CN109816714A - A kind of point cloud object type recognition methods based on Three dimensional convolution neural network - Google Patents
A kind of point cloud object type recognition methods based on Three dimensional convolution neural network Download PDFInfo
- Publication number
- CN109816714A CN109816714A CN201910034966.XA CN201910034966A CN109816714A CN 109816714 A CN109816714 A CN 109816714A CN 201910034966 A CN201910034966 A CN 201910034966A CN 109816714 A CN109816714 A CN 109816714A
- Authority
- CN
- China
- Prior art keywords
- layer
- neural network
- value
- dimensional
- convolution neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The point cloud object type recognition methods based on Three dimensional convolution neural network that the present invention provides a kind of, specifically includes: acquiring and handle the point cloud data of the object of known object type, obtain the three-dimensional matrice being made of CP；It will be input in Three dimensional convolution neural network, successively every layer of Three dimensional convolution neural network handled, and Three dimensional convolution neural network is updated by the three-dimensional matrice that CP is formed in training set data, obtain new Three dimensional convolution neural network；It will be input in new Three dimensional convolution neural network in verifying collection data by the three-dimensional matrice that CP is formed, judgement obtains trained Three dimensional convolution neural network；Object to be identified is identified according to trained Three dimensional convolution neural network, finally obtains the type of object to be identified；Operation of the present invention is convenient, overcomes the shortcomings that being needed in conventional point cloud object identification according to different type object using distinct methods；And overcome the disadvantage for needing that the expert of different majors knowledge participates in and automatic identification degree is low.
Description
Technical field
The invention belongs to area of pattern recognition, and in particular to a kind of point cloud object type based on Three dimensional convolution neural network
Recognition methods.
Background technique
Point cloud data is distributed in three dimensions using structured light scanner or three-dimensional laser scanner are collected
Discrete point set, it has special advantage to the expression of the shape of complex scene and object.With Artificial Intelligence Development and
The promotion of industry the degree of automation, knowing method for distinguishing based on two dimensional image can no longer meet requirement, people start from two dimension to
Three-dimensional research field transition.Relative to two dimensional image, three dimensional point cloud includes that the information of object is more, is related to each side of object
Face factor is more abundant, expresses more comprehensive.
Existing recognition methods is broadly divided into four classes: the point cloud object identification based on global characteristics, based on local feature
Point cloud object identification, point cloud object identification based on machine learning and knows method based on matched cloud object of figure.Above four
Class recognition methods has certain application in cloud object identification, but all haves the defects that certain.Firstly, since point
Cloud data volume is big, and leading to the target signature extracted is a high dimension vector, computationally intensive, inefficiency；Secondly, needing a large amount of people
Work intervention, the i.e. participation of the priori knowledge of expert, algorithm the degree of automation are lower；In addition, conventional point cloud object identification method one
As for certain objects there is specific identification method, the frame of a unified identification different objects can not be formed.
Summary of the invention
The computationally intensive, low efficiency for existing in the prior art cloud object identification method, and the degree of automation is lower
The shortcomings that, the purpose of the application is, provides a kind of point cloud object type recognition methods based on Three dimensional convolution neural network.
To achieve the goals above, the present invention takes following technical scheme to be achieved:
A kind of building method of the Three dimensional convolution neural network of cloud object type identification, comprising the following steps:
The point cloud data of the object of step 1, acquisition known object type, and peel off to the point cloud data of the object
Point deletion obtains muting object point cloud data；The point cloud data of the object includes training set data, verifying collection data；
Step 2 carries out voxelization operation to obtained muting object point cloud data, obtains three-dimensional matrice, and to institute
It states three-dimensional matrice and operation is normalized, obtain the three-dimensional matrice being made of CP, the three-dimensional matrice being made of CP includes instruction
The three-dimensional matrice practicing the three-dimensional matrice being made of in collection data CP and being made of in verifying collection data CP；
Step 3, using the three-dimensional matrice being made of in training set data CP as input data, input data is input to three
It ties up in convolutional neural networks, obtains the Three dimensional convolution neural network with input data, the Three dimensional convolution of input data will be had
The layer of neural network is successively denoted as the 1st layer to the 32nd layer in accordance with the order from top to bottom；
Step 4 successively calculates the 1st layer to 19 layers of the Three dimensional convolution neural network with input data, obtains
19th layer of feature graphic sequence；By the 19th layer of characteristic pattern sequence inputting to the Three dimensional convolution neural network with input data
20th layer, successively the 20th layer to 30 layers are calculated, obtains the 30th layer of feature graphic sequence；By the 30th layer of feature graphic sequence
It is input to the 31st layer of the Three dimensional convolution neural network with input data, successively the 31st layer to the 32nd layer is calculated, is obtained
To final feature vector；According to final feature vector, the probability and network losses value of every kind of identification types is calculated；
The probability of every kind of identification types and network losses value are input to three with input data by step 5
It ties up in convolutional neural networks, obtains the new Three dimensional convolution neural network with input data；Obtained new having is inputted
The Three dimensional convolution neural network of data carries out calculating update, obtains new Three dimensional convolution neural network.
Further, the building method of the Three dimensional convolution neural network of described cloud object type identification, further includes following
Step:
Step 6 will be input to the new Three dimensional convolution neural network by the three-dimensional matrice that CP is formed in verifying collection data
In, successively obtain first network losses value, second network losses value and third network losses value；
When obtained first network losses value, second network losses value and third network losses value are equal, or according to
When secondary increase, using obtained new Three dimensional convolution neural network as trained Three dimensional convolution neural network；
When obtained first network losses value, second network losses value and third network losses value successively increase it is small
When, execute operation identical with step 4-6.
A kind of point cloud object type recognition methods based on Three dimensional convolution neural network, specifically includes the following steps:
The point cloud data of step 1, acquisition object to be identified, and outlier is carried out to the point cloud data of the object to be identified
It deletes, obtains muting object point cloud data to be identified；
Step 2 carries out voxelization operation to obtained muting object point cloud data to be identified, obtains three-dimensional matrice,
And operation is normalized to the three-dimensional matrice, obtain the three-dimensional matrice being made of CP；
It is further comprising the steps of:
The three-dimensional matrice being made of CP value is input to new Three dimensional convolution neural network or trained three-dimensional by step 3
In convolutional neural networks, the Three dimensional convolution neural network with input data is obtained；Successively to the three-dimensional volume with input data
The 1st layer to the 32nd layer of product neural network is calculated, and final feature vector is obtained；According to final feature vector, calculate
Obtain the probability of every kind of identification types；According to every kind of obtained identification types probability, the type of object to be identified is finally obtained.
Further, to the 1st layer to 19 layers progress of the Three dimensional convolution neural network with input data described in step 4
It calculates, concrete operations are as follows:
Step 41 is used as current layer for the 1st layer of the Three dimensional convolution neural network with input data；
Step 42 carries out convolution operation to current layer using formula (3), the output valve after obtaining convolution operation；
Wherein, x indicates output valve after convolution operation, x ∈ R；L indicates to be rolled up in the three-dimensional matrice being made of in training set CP
Long-pending channel position, 1≤l≤M；M represents the number for the characteristic pattern that current layer has, M >=3；XlIt indicates in training set by CP group
At three-dimensional matrice in the channel l value, Xl∈ R, WlIndicate the channel l convolution in the three-dimensional matrice being made of in training set CP
Core weight, Wl∈R；
To output valve after obtained convolution operation, carry out nonlinear transformation using formula (4), obtain it is non-linear after value；
Wherein, x indicates the output valve after convolution operation；Value after Relu (x) expression is non-linear, Relu (x) are more than or equal to
0；
To obtain it is non-linear after value carry out pondization operation, the pondization operates specifically: uses step-length for 2, size
For 2*2*2 window to obtain it is non-linear after value divide, obtain multiple windows；Choose 8 numerical value in each window
Typical value of the maximum value as corresponding window, obtain the typical value of multiple windows；Made by the typical value of obtained multiple windows
For the feature graphic sequence of current layer, and will be in next layer of the characteristic pattern sequence inputting of current layer to current layer；By current layer
Next layer is used as current layer, executes operation identical with step 42, until the 19th layer has executed, the three-dimensional volume with input data
The operation of 1st layer to 19 layers progress of product neural network terminates.
Further, the 20th layer to the 30th layer is calculated described in step 4, concrete operations are as follows:
Step 43 is used as current layer for the 20th layer；
Step 44 calculates current layer to learning method using residual error, obtains the feature graphic sequence of current layer, and will
The characteristic pattern sequence inputting of obtained current layer is into next layer of current layer；It is used as current layer by next layer of current layer, is held
Row operation identical with step 44, until the 30th layer has executed, to the 20th of the Three dimensional convolution neural network with input data the
The operation that layer is carried out to the 30th layer terminates.
Further, the 31st layer to 32 layers are calculated described in step 4, obtains final feature vector；According to most
Whole feature vector, probability and network losses value, the concrete operations that every kind of identification types are calculated are as follows:
Step 45 uses full attended operation to the 31st layer, obtains the 31st layer of feature vector；By the 31st layer of feature vector
It is input to the 32nd layer, full attended operation is used to the 32nd layer, obtains final feature vector；
Step 46, according to demand and training set number of types, is arranged m kind identification types；It is calculated often using formula (5)
The probability of kind identification types,
Wherein, i is i-th kind of identification types, and m is the species number of identification types, 1≤i≤j≤m, m >=1；wiKnow for i-th kind
The weighted value of other type, wi∈ R,Indicate the transposition of the weighted value of i-th kind of identification types,The final spy that V is
Levy vector, V ∈ R；PiIndicate the probability of i-th kind of identification types, 0≤Pi≤1；biIndicate the bias term of identification types in i-th, bi
∈R；bjRepresent the bias term of jth kind identification types, bj∈R；
Step 47 calculates network losses value:
Wherein, m is the species number of identification types, and i is i-th kind of identification types, 1≤i≤m, m >=1；PiIndicate i-th kind of knowledge
The probability of other type, when known object type is identical as i-th kind of identification types, yiIt is 1；When known object type and i-th kind
When identifying classification difference, yiIt is 0.
Further, step 5 concrete operations are as follows:
The probability of every kind of identification types and network losses value are input to the Three dimensional convolution for having input data
In neural network, the new Three dimensional convolution neural network with input data is obtained；
Successively according to the 32nd layer to the 1st layer of sequence of Three dimensional convolution neural network, using gradient descent algorithm to described
The new Three dimensional convolution neural network with input data is calculated, and each layer of weighted value is obtained；
Using each layer of obtained weighted value weighted value new as this layer, new Three dimensional convolution neural network is obtained.
Further, step 6 concrete operations are as follows:
Step 61 will be input to the new Three dimensional convolution neural network by the three-dimensional matrice that CP is formed in verifying collection data
In；
Step 62 executes operation identical with step 4, obtains first network losses value；The new three-dimensional volume that will be obtained
Product neural network executes behaviour identical with step 4-5, step 61-62 as the Three dimensional convolution neural network for having input data
Make, the Three dimensional convolution network updated and second network losses value；Using the Three dimensional convolution neural network of update as having
The Three dimensional convolution neural network of input data executes operation identical with step 4-5, step 61-62, obtains third network damage
Mistake value；
When obtained first network losses value, second network losses value and third network losses value are equal, or according to
When secondary increase, using obtained new Three dimensional convolution neural network as trained Three dimensional convolution neural network；
When obtained first network losses value, second network losses value and third network losses value are sequentially reduced
When, execute operation identical with step 4-6.
Further, step 3 concrete operations are as follows:
Step 31 is used as current layer for the 1st layer of the Three dimensional convolution neural network with input data；
Step 32 carries out convolution operation to current layer using formula (6), the output valve after obtaining convolution operation；
Wherein, x indicates output valve after convolution operation, x ∈ R；L indicates to be rolled up in the three-dimensional matrice being made of in training set CP
Long-pending channel position, 1≤l≤M；M represents the number for the characteristic pattern that current layer has, M >=3；XlIt indicates in training set by CP group
At three-dimensional matrice in the channel l value, Xl∈ R, WlIndicate the channel l convolution in the three-dimensional matrice being made of in training set CP
Core weight, Wl∈R；
To output valve after obtained convolution operation, carry out nonlinear transformation using formula (7), obtain it is non-linear after value；
Wherein, x indicates the output valve after convolution operation；Value after Relu (x) expression is non-linear, Relu (x) are more than or equal to
0；
To obtain it is non-linear after value carry out pondization operation, the pondization operates specifically: uses step-length for 2, size
For 2*2*2 window to obtain it is non-linear after value divide, obtain multiple windows；Choose 8 numerical value in each window
Typical value of the maximum value as corresponding window, obtain the typical value of multiple windows；Made by the typical value of obtained multiple windows
For the feature graphic sequence of current layer, and will be in next layer of the characteristic pattern sequence inputting of current layer to current layer；By current layer
Next layer is used as current layer, executes operation identical with step 32, until the 19th layer has executed, executes step 33；
Step 33 calculates current layer to learning method using residual error, obtains the feature graphic sequence of current layer, and will
The characteristic pattern sequence inputting of obtained current layer is into next layer of current layer；It is used as current layer by next layer of current layer, is held
Row operation identical with step 33 executes step 34 until the 30th layer has executed；
Step 34, current layer is calculated using full attended operation, obtains the feature vector of current layer, and will obtained
The characteristic pattern vector of current layer is input in next layer of current layer；It is used as current layer by next layer of current layer, executes and walks
Rapid 34 identical operation, until the 30th layer of feature vector is obtained, using the 30th layer of feature vector as final feature vector；
Step 35, according to demand and training set number of types, is arranged m kind identification types；It is calculated often using formula (8)
The probability of kind identification types；
Wherein, i is i-th kind of identification types, and m is the species number of identification types, 1≤i≤j≤m, m >=1；wiKnow for i-th kind
The weighted value of other type, wi∈ R,Indicate the transposition of the weighted value of i-th kind of identification types,V is final
Feature vector, V ∈ R；PiIndicate the probability of i-th kind of identification types, 0≤Pi≤1；biIndicate the bias term of identification types in i-th,
bi∈R；bjRepresent the bias term of jth kind identification types, bj∈R。
Further, the class of object to be identified is finally obtained according to every kind of obtained identification types probability described in step 3
Type, concrete operations are as follows:
Maximum value is chosen from the probability of every kind of obtained identification types, using the corresponding type of maximum value as object to be identified
The type of body obtains the type of final object to be identified.
Compared with prior art, beneficial effects of the present invention are as follows:
1, the present invention in order to solve conventional point cloud the degree of automation it is low and can not accomplish using a kind of frame identify inhomogeneity
Three dimensional convolution neural network is added, by Three dimensional convolution neural network in the shortcomings that type object during cloud object identification
Learning training, obtain new Three dimensional convolution neural network, can be improved effectively in conjunction with object to be identified wait know
The accuracy rate of other object type identification.
2, new Three dimensional convolution neural network is added during cloud object identification and carries out sentencing for object type by the present invention
It is disconnected, the accuracy rate of object type identification is improved, the point cloud for needing expert in the prior art according to different type object is overcome
Design the priori knowledge disadvantage of different operators；And it is low and can not accomplish using a kind of frame to overcome prior art the degree of automation
The shortcomings that identifying different type object has high degree of automation, easy to operate, the high feature of accuracy rate；
3, it tends towards stability in the present invention with the variation of the intensifications target value of the number of plies, that is, learns to degenerate, by the
20 layers to 30 layers are changed to learn the variation of residual error item by the variation of learning objective value, the feature for the extraction point cloud data that can be automated
Vector, and difference cloud object type can be judged using a frame.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is result schematic diagram of the object point cloud data under XYZ three dimensional space coordinate and after voxelization operation；Wherein:
Figure a is schematic diagram of the object point cloud data under XYZ three dimensional space coordinate；
Scheming b is the result schematic diagram after voxelization operation.
Specific embodiment
A kind of building method of the Three dimensional convolution neural network of cloud object type identification, comprising the following steps:
The point cloud data of the object of step 1, acquisition known object type, and peel off to the point cloud data of the object
Point deletion obtains muting object point cloud data；The point cloud data of the object includes training set data, verifying collection data；
Step 2 carries out voxelization operation to obtained muting object point cloud data, obtains three-dimensional matrice, and to institute
It states three-dimensional matrice and operation is normalized, obtain the three-dimensional matrice being made of CP；The three-dimensional matrice being made of CP includes instruction
The three-dimensional matrice practicing the three-dimensional matrice being made of in collection data CP and being made of in verifying collection data CP；
Further include following steps:
Step 3, using the three-dimensional matrice being made of in training set data CP as input data, input data is input to three
It ties up in convolutional neural networks, obtains the Three dimensional convolution neural network with input data；The Three dimensional convolution neural network uses
The layer of Three dimensional convolution neural network with input data is successively denoted as the 1st layer by 32 layers of structure in accordance with the order from top to bottom
To the 32nd layer；
Step 4, successively according to the direction of forward conduction, to the 1st layer of the Three dimensional convolution neural network with input data
It is calculated to 19 layers, obtains the 19th layer of feature graphic sequence；
By the 19th layer of characteristic pattern sequence inputting to the 20th layer of the Three dimensional convolution neural network with input data, successively
20th layer to 30 layers are calculated, the 30th layer of feature graphic sequence is obtained；
By the 30th layer of characteristic pattern sequence inputting to the 31st layer of the Three dimensional convolution neural network with input data, successively
31st layer to the 32nd layer is calculated, final feature vector is obtained；According to final feature vector, every kind of knowledge is calculated
The probability and network losses value of other type；
The probability of every kind of identification types and network losses value are input to three with input data by step 5
It ties up in convolutional neural networks, obtains the new Three dimensional convolution neural network with input data；Obtained new having is inputted
The Three dimensional convolution neural network of data carries out calculating update, obtains new Three dimensional convolution neural network；
The present invention in order to solve conventional point cloud the degree of automation it is low and can not accomplish using a kind of frame identify different type
Three dimensional convolution neural network is added, by Three dimensional convolution neural network in the shortcomings that object during cloud object identification
Learning training obtains new Three dimensional convolution neural network, can be improved to be identified effectively in conjunction with object to be identified
The accuracy rate of object type identification.
Specifically, further comprising the steps of:
Step 6 will be input to the new Three dimensional convolution neural network by the three-dimensional matrice that CP is formed in verifying collection data
In, successively obtain first network losses value, second network losses value and third network losses value；
When obtained first network losses value, second network losses value and third network losses value are equal, or according to
When secondary increase, using obtained new Three dimensional convolution neural network as trained Three dimensional convolution neural network；
When obtained first network losses value, second network losses value and third network losses value are sequentially reduced
When, execute operation identical with step 4-6.
Whether which is qualified by the new Three dimensional convolution neural network that verifying collection data judge, works as Three dimensional convolution
After convolutional neural networks train, the accuracy rate of object type identification to be identified can be improved.
Specifically, the concrete operations of the step 1 are as follows:
Object is scanned by spatial digitizer, obtains the point cloud data of the object, the point cloud data of object includes training
Collect data, verifying collection data；Outlier delete operation is carried out using point cloud data of the k nearest neighbor method to obtained object, is obtained
Muting object point cloud data.
The data mode of the point cloud data of object is point in object point cloud data respectively under XYZ three-dimensional coordinate system
X-axis, Y-axis, the set of RGB triple channel value corresponding to projection value and the point on three directions of Z axis.RGB indicates three primary colors, packet
Include red, green, blue.
In order to solve the problems, such as manually to acquire there are a large amount of outliers in data, using k nearest neighbor method to obtained object
Point cloud data carry out outlier delete operation, this method can be good at complete outlier rejecting, ensure that point cloud data compared with
It is accurate.
Specifically, step 2 concrete operations are as follows:
Step 21 finds out noiseless object point cloud data X-axis, Y-axis, Z axis three under XYZ three-dimensional coordinate system respectively
Maximum projection value and the smallest projection value on direction, obtain Xmax,Xmin,Ymax,Ymin,Zmax,ZminSix values, with (Xmax,
0,0)、(Xmin, 0,0), (0, Ymax, 0), (0, Ymin, 0), (0,0, Zmax), (0,0, Zmin) it is that coordinate constitutes cube；
Cube cutting along three X-axis, Y-axis, Z axis directions is respectively M unit by step 22, is obtained M*M*M and sub is stood
Cube；The quantity at noiseless object point cloud data midpoint in each sub-cube is counted, and nothing in each sub-cube is made an uproar
Typical value of the quantity at sound object point cloud data midpoint as corresponding sub-cube, obtains the typical value of M*M*M sub-cube；
One three-dimensional matrice is formed by the typical value of M*M*M sub-cube, obtains three-dimensional matrice, of numerical value in the three-dimensional matrice
Number is M*M*M, wherein 20≤M≤30；
Step 23, using normalized function, i.e. operation is normalized to obtained three-dimensional matrice in formula (1), obtain by
The three-dimensional matrice of CP value composition, the three-dimensional matrice being made of CP value includes the three-dimensional square being made of in training set data CP value
The three-dimensional matrice being made of in battle array, verifying collection data CP value；
Wherein R represents the number at the point cloud midpoint after normalization, and R >=1000, N represent at current object point cloud midpoint
Number；CP indicates the value obtained after normalization, CP >=0；X, y, z respectively represent the x of three-dimensional matrice, the coordinate on tri- directions y, z
Value, nx,y,zThe value of (x, y, z) is designated as under representing in three-dimensional matrice.
Operation, the precision of identification model are normalized to obtained three-dimensional matrice using the normalized function of formula (1)
Higher, speed is slower；
Formula (2) also can be used, operation is normalized to obtained three-dimensional matrice, obtains the three-dimensional square being made of CP value
Battle array, including the three-dimensional matrice being made of in training set data CP value, the three-dimensional matrice being made of CP value in test set data, verifying
The three-dimensional matrice being made of in collection data CP value；
Wherein, P indicates the numerical value in three-dimensional matrice, P > 0；μ is the mean value of three-dimensional matrice, μ > 0；σ is the mark of three-dimensional matrice
Quasi- poor, σ > 0.
Operation is normalized to obtained three-dimensional matrice using formula (2), the precision of identification model is slightly lower, but speed compared with
Fastly.
In order to reduce point cloud data amount and the input requirements of Three dimensional convolution neural network can be met, need to muting
Object point cloud data carries out voxelization.Which can reduce the number of points cloud processing under the premise of remaining valid data
According to measure and realize meet Three dimensional convolution Neural Network Data requirement format；It is different due to the influence of human factor in acquisition
The point cloud data amount of object acquisition is different, will affect the final result of prediction, therefore obtained three-dimensional matrice is normalized
Operation, influence of the point cloud data amount to final object type prediction is finally eliminated by method for normalizing.
Specifically, being counted described in step 4 to the 1st layer to 19 layers of the Three dimensional convolution neural network with input data
It calculates, concrete operations are as follows:
Step 41 regard the 1st layer of Three dimensional convolution neural network with input data as current layer；
Step 42 carries out convolution operation to current layer using formula (3), the output valve after obtaining convolution operation；
Wherein, x indicates output valve after convolution operation, x ∈ R；L indicates to be rolled up in the three-dimensional matrice being made of in training set CP
Long-pending channel position, 1≤l≤M；M represents the number for the characteristic pattern that current layer has, M >=3；XlIt indicates in training set by CP group
At three-dimensional matrice in the channel l value, Xl∈ R, WlIndicate the channel l convolution in the three-dimensional matrice being made of in training set CP
Core weight, Wl∈R；
To output valve after obtained convolution operation, using Relu activation primitive, i.e. formula (4) carries out nonlinear transformation, obtains
To the value after non-linear；
Wherein, x indicates the output valve after convolution operation；Value after Relu (x) expression is non-linear, Relu (x) are more than or equal to
0；
To obtain it is non-linear after value carry out pondization operation, the pondization operates specifically: uses step-length for 2, size
For 2*2*2 window to obtain it is non-linear after value divide, obtain multiple windows；Choose 8 numerical value in each window
Typical value of the maximum value as corresponding window, obtain the typical value of multiple windows；Made by the typical value of obtained multiple windows
For the feature graphic sequence of current layer, and will be in next layer of the characteristic pattern sequence inputting of current layer to current layer；By current layer
Next layer is used as current layer, executes operation identical with step 42, until the 19th layer has executed, to the three-dimensional for having input data
The operation that the 1st layer of convolutional neural networks is carried out to the 19th layer terminates.
Convolution operation is carried out by the 1st layer to the 19th layer, nonlinear transformation and pondization operation are realized to input number
According to the 1st layer to the 19th layer of Three dimensional convolution neural network of learning objective.
Specifically, described in step 4 the 20th layer to the 30th layer calculated, the specific steps are as follows:
Step 43 is used as current layer for the 20th layer；
Step 44 calculates current layer to learning method using residual error, obtains the feature graphic sequence of current layer, and will
The characteristic pattern sequence inputting of obtained current layer is into next layer of current layer；It is used as current layer by next layer of current layer, is held
Row operation identical with step 44, until the 30th layer has executed, to the 20th of the Three dimensional convolution neural network with input data the
The operation that layer is carried out to the 30th layer terminates；
Since the Three dimensional convolution neural network with input data is in learning objective, with the intensification target value of the number of plies
Variation tends towards stability, i.e. study is degenerated, then change learning objective is needed, by being become to 30 layers by learning objective value to the 20th layer
Change the variation for being changed to study residual error item, the feature vector for the extraction point cloud data that can be automated, and a frame can be used
Difference cloud object type is judged.
Specifically, calculating described in step 4 the 31st layer to 32 layers, final feature vector is obtained；According to final
Feature vector, probability and network losses value, the concrete operations that every kind of identification types are calculated are as follows:
Step 45 uses full attended operation to the 31st layer, obtains the 31st layer of feature vector；By the 31st layer of feature vector
The 32nd layer is inputted, full attended operation is used to the 32nd layer, obtains final feature vector；
Step 46, according to demand and training set number of types, is arranged m kind identification types；Using softmax function, that is, formula
(5), it calculates separately to obtain the probability of every kind of identification types,
Wherein, i is i-th kind of identification types, and m is the species number of identification types, 1≤i≤j≤m, m >=1；wiKnow for i-th kind
The weighted value of other type, wi∈ R,Indicate the transposition of the weighted value of i-th kind of identification types,The final spy that V is
Levy vector, V ∈ R；PiIndicate the probability of i-th kind of identification types, 0≤Pi≤1；, biIndicate the bias term of identification types in i-th, bi
∈R；bjRepresent the bias term of jth kind identification types, bj∈R；
Step 47 calculates network losses value:
Wherein, m is the species number of identification types, and i is i-th kind of identification types, 1≤i≤m, m >=1；PiIndicate i-th kind of knowledge
The probability of other type；When known object type is identical as i-th kind of identification types, yiIt is 1；When known object type and i-th kind
When identifying classification difference, yiIt is 0.
Which is capable of the accuracy rate of judgment object identification types by the probability of every kind of identification types of calculating；Calculate net
Network penalty values can effectively judge whether Three dimensional convolution neural network trains qualification.
In the embodiment of the present invention, according to the actual situation, step 4 and step are repeated to new Three dimensional convolution neural network
Rapid 5, iteration 5 times according to this obtain new Three dimensional convolution neural network, then execute step 6.
Specifically, step 5 concrete operations are as follows:
The probability of every kind of identification types and network losses value are input to the Three dimensional convolution for having input data
In neural network, the new Three dimensional convolution neural network with input data is obtained；Successively according to Three dimensional convolution neural network
32nd layer to the 1st layer of sequence, the i.e. direction of reverse conduction, using gradient descent algorithm to the new three-dimensional with input data
Convolutional neural networks are calculated, and each layer of weighted value is obtained；Using each layer of obtained weighted value weight new as this layer
Value, obtains new Three dimensional convolution neural network.
Which is realized by the direction of reverse conduction to the new Three dimensional convolution neural network with input data
It updates, obtains new Three dimensional convolution neural network.The iteration of network is completed eventually by forward conduction and reverse conduction, it is final real
Show the autonomous learning of Three dimensional convolution neural network parameter, and reached and participated in without expert's priori knowledge, independently judges object
The target of body type.
Specifically, the step 6 specifically includes the following steps:
Step 61, the new Three dimensional convolution nerve net that will be input in verifying collection data by the three-dimensional matrice that CP is formed
In network；
Step 62 executes operation identical with step 4, obtains first network losses value；The new three-dimensional volume that will be obtained
Product neural network executes behaviour identical with step 4-5, step 61-62 as the Three dimensional convolution neural network for having input data
Make, the Three dimensional convolution network updated and second network losses value；Using the Three dimensional convolution neural network of update as having
The Three dimensional convolution neural network of input data executes operation identical with step 4-5, step 61-62, obtains third network damage
Mistake value；
When obtained first network losses value, second network losses value and third network losses value are equal, or according to
When secondary increase, using obtained new Three dimensional convolution neural network as trained Three dimensional convolution neural network；
When obtained first network losses value, second network losses value and third network losses value successively increase it is small
When, execute operation identical with step 4-6.
Whether which is qualified by the new Three dimensional convolution neural network that verifying collection data judge, and to unqualified
New Three dimensional convolution neural network optimize.
A kind of point cloud object type recognition methods based on Three dimensional convolution neural network, specifically includes the following steps:
The point cloud data of step 1, acquisition object to be identified, and outlier is carried out to the point cloud data of the object to be identified
It deletes, obtains muting object point cloud data to be identified；
Step 2 carries out voxelization operation to obtained muting object point cloud data to be identified, obtains three-dimensional matrice,
And operation is normalized to the three-dimensional matrice, obtain the three-dimensional matrice being made of CP；
It is further comprising the steps of:
The three-dimensional matrice being made of CP value is input to new Three dimensional convolution neural network or trained three-dimensional by step 3
In convolutional neural networks, the Three dimensional convolution neural network with input data is obtained；Successively to the three-dimensional volume with input data
The 1st layer to the 32nd layer of product neural network is calculated, and final feature vector is obtained；According to final feature vector, calculate
Obtain the probability of every kind of identification types；According to every kind of obtained identification types probability, the type of object to be identified is finally obtained.
Step 1, step 2 difference in a kind of point cloud object type recognition methods based on Three dimensional convolution neural network
With step 1, the operating process phase in step 2 in the building method of the Three dimensional convolution neural network of a kind of cloud object type identification
Together.
The present invention in order to solve conventional point cloud the degree of automation it is low and can not accomplish using a kind of frame identify different type
New Three dimensional convolution neural network or trained Three dimensional convolution nerve is added in the shortcomings that object during cloud object identification
Network carries out the judgement of object type to be identified, by the training to Three dimensional convolution neural network, improves object type to be identified
The accuracy rate of type identification overcomes the elder generation for needing expert to design different operators according to the point cloud of different type object in the prior art
Test knowledge disadvantage；And overcome prior art the degree of automation it is low and can not accomplish using a kind of frame identify different type object
The shortcomings that, there is high degree of automation, easy to operate, the high feature of accuracy rate.
Specifically, successively to the 1st layer of Three dimensional convolution neural network with input data to the described in the step 3
32 layers are calculated, and final feature vector is obtained, and concrete operations are as follows:
Step 31 is used as current layer for the 1st layer of the Three dimensional convolution neural network with input data；
Step 32 carries out convolution operation to current layer using formula (6), the output valve after obtaining convolution operation；
Wherein, x indicates output valve after convolution operation, x ∈ R；L indicates to be rolled up in the three-dimensional matrice being made of in training set CP
Long-pending channel position, 1≤l≤M；M represents the number for the characteristic pattern that current layer has, M >=3；XlIt indicates in training set by CP group
At three-dimensional matrice in the channel l value, Xl∈ R, WlIndicate the channel l convolution in the three-dimensional matrice being made of in training set CP
Core weight, Wl∈R；
To output valve after obtained convolution operation, carry out nonlinear transformation using formula (7), obtain it is non-linear after value；
Wherein, x indicates the output valve after convolution operation；Value after Relu (x) expression is non-linear, Relu (x) are more than or equal to
0；
To obtain it is non-linear after value carry out pondization operation, the pondization operates specifically: uses step-length for 2, size
For 2*2*2 window to obtain it is non-linear after value divide, obtain multiple windows；Choose 8 numerical value in each window
Typical value of the maximum value as corresponding window, obtain the typical value of multiple windows；Made by the typical value of obtained multiple windows
For the feature graphic sequence of current layer, and will be in next layer of the characteristic pattern sequence inputting of current layer to current layer；By current layer
Next layer is used as current layer, executes operation identical with step 32, until the 19th layer has executed, executes step 33；
Step 33 calculates current layer to learning method using residual error, obtains the feature graphic sequence of current layer, and will
The characteristic pattern sequence inputting of obtained current layer is into next layer of current layer；It is used as current layer by next layer of current layer, is held
Row operation identical with step 33 executes step 34 until the 30th layer has executed；
Step 34, current layer is calculated using full attended operation, obtains the feature vector of current layer, and will obtained
The characteristic pattern vector of current layer is input in next layer of current layer；It is used as current layer by next layer of current layer, executes and walks
Rapid 34 identical operation, until the 30th layer of feature vector is obtained, using the 30th layer of feature vector as final feature vector；
Step 35, according to demand and training set number of types, is arranged m kind identification types；It is calculated often using formula (8)
The probability of kind identification types；
Wherein, i is i-th kind of identification types, and m is the species number of identification types, 1≤i≤j≤m, m >=1；wiKnow for i-th kind
The weighted value of other type, wi∈ R,Indicate the transposition of the weighted value of i-th kind of identification types,V is final
Feature vector, V ∈ R；PiIndicate the probability of i-th kind of identification types, 0≤Pi≤1；biIndicate the bias term of identification types in i-th,
bi∈R；bjRepresent the bias term of jth kind identification types, bj∈R。
Specifically, finally obtaining the class of object to be identified according to every kind of obtained identification types probability described in step 3
Type, concrete operations are as follows:
Maximum value is chosen from the probability of every kind of obtained identification types, using the corresponding type of maximum value as object to be identified
The type of body obtains the type of final object to be identified.
According to trained Three dimensional convolution neural network, the identification that the type of object to be identified can be automated,
To obtain the type of object to be identified, the final identification realized to object type to be identified.
Embodiment:
The point cloud data that sample is chosen in the embodiment is RGBD common data sets, i.e. depth three primary colors common data sets,
The RGBD common data concentration that embodiment is taken includes the point cloud number of 6 apple, ball, banana, bell pepper, eyeshade and bowl objects
According to；Specific steps are as follows:
Step 1, the point cloud data for acquiring object；Since the point cloud data of the sample of selection is RGBD common data sets, and
Pretreatment is had been subjected to, which, which directly obtains, schemes muting object point cloud under XYZ three-dimensional coordinate system shown in a in Fig. 2
Data.Wherein the point cloud data of sample chooses the point cloud data that RGBD common data concentrates 300 objects, the point of 200 objects
Cloud data are as training set data, and the point cloud data of 50 objects is as verifying collection data, the point cloud data conduct of 50 objects
Test set data；
Step 2 carries out voxelization to the muting object point cloud data that step 1 obtains, and voxelization number is 20*20*
20 totally 8000 cells, color are RGB triple channel, obtain the three-dimensional matrice as shown in the figure b in Fig. 2 in three sides of x, y, z
Upward schematic diagram finally obtains the three-dimensional matrice of 3 20*20*20；The normalized side of standard is used to obtained three-dimensional matrice
Operation is normalized in method, obtains the three-dimensional matrice being made of CP；
Step 3, using the three-dimensional matrice being made of in training set data CP as input data, be input to Three dimensional convolution nerve
In network, the Three dimensional convolution neural network with input data is obtained, by the Three dimensional convolution neural network with input data
Layer is successively denoted as the 1st layer to the 32nd layer according to the sequence of top to bottm；
Step 4 successively calculates the 1st layer to 19 layers of the Three dimensional convolution neural network with input data, obtains
19th layer of feature graphic sequence；
By the 19th layer of characteristic pattern sequence inputting to the 20th layer of the Three dimensional convolution neural network with input data, successively
20th layer to 30 layers are calculated, the 30th layer of feature graphic sequence is obtained；
By the 30th layer of characteristic pattern sequence inputting to the 31st layer of the Three dimensional convolution neural network with input data, successively
31st layer to the 32nd layer is calculated, final feature vector is obtained；According to demand with training set number of types, apple is set
Every kind of identification types are calculated according to final feature vector in 6 kinds of fruit, ball, banana, bell pepper, eyeshade and bowl identification types
Probability and network losses value；
Step 5: the probability of every kind of identification types and network losses value are input to three with input data
It ties up in convolutional neural networks, obtains the new Three dimensional convolution neural network with input data；Successively according to Three dimensional convolution nerve
32nd layer to the 1st layer of sequence of network, using gradient descent algorithm to the new Three dimensional convolution nerve net with input data
Network is calculated, and each layer of weighted value is obtained；Using each layer of obtained weighted value weighted value new as this layer, obtain new
Three dimensional convolution neural network.
The initialization network parameter in Three dimensional convolution neural metwork training, is arranged learning rate lr=0.01, and the number of iterations is
50 hyper parameters confirm optimal parameter value using the strategy of grid search.The every iteration of training dataset 5 times, by Three dimensional convolution mind
Apply to verifying collection data through network model and finds out loss value；
Verifying collection data are input to the new Three dimensional convolution neural network by the three-dimensional matrice that CP is formed by step 6
In, operation identical with step 4 is executed, first network losses value is obtained；Obtained new Three dimensional convolution neural network is made
For the Three dimensional convolution neural network with input data, operation identical with step 4-5, step 6, the three-dimensional updated are executed
Convolutional network and second network losses value；Using the Three dimensional convolution neural network of update as the Three dimensional convolution for having input data
Neural network executes operation identical with step 4-5, step 6, obtains third network losses value；
When obtained first network losses value, second network losses value and third network losses value are equal, or according to
When secondary increase, using obtained new Three dimensional convolution neural network as trained Three dimensional convolution neural network, step 7 is executed；
When obtained first network losses value, second network losses value and third network losses value are sequentially reduced
When, execute operation identical with step 4-6；
Step 7: in order to verify accuracy of the invention, the three-dimensional matrice being made of in obtained test set data CP is defeated
Enter into trained Three dimensional convolution neural network, the probability of every kind of object identification type is calculated；According to every kind obtained
Object identification type obtains the type of final object.
When the actual type of object is respectively apple, ball, banana, bell pepper, eyeshade and bowl, pass through survey in statistical sample
The type of examination collection data prediction is respectively the ratio of apple, ball, banana, bell pepper, eyeshade and bowl, and statistical result is as shown in table 1.
The meaning that each column indicates in table 1: first row represents the actual type of object, and the first row represents the object type of prediction, numerical tabular
Show ratio of the sum for predicting a certain type in total sample number.With the second row data instance, the second row first row represents object
The actual type of body is apple, and when the actual type of object is apple, the type for using method of the invention to predict is apple
Ratio of the sum in total sample number be 84%, predict that ratio of the sum of ball in total sample number is 7%, predict bell pepper
Ratio of the sum in total sample number be 7%, predict that ratio of the sum of eyeshade in total sample number is 1%, predict banana
Sum and ratio of the sum in total sample number of bowl be 0；By test statistics, identified using method provided by the invention
The accuracy rate of the type of object reaches 91.7% or so, has high degree of automation, easy to operate, the high feature of accuracy rate.
Table 1
| Apple | Ball | Banana | Bell pepper | Eyeshade | Bowl | |
| Apple | 0.84 | 0.07 | 0.00 | 0.07 | 0.01 | 0.00 |
| Ball | 0.06 | 0.90 | 0.00 | 0.02 | 0.00 | 0.02 |
| Banana | 0.00 | 0.00 | 0.97 | 0.02 | 0.06 | 0.00 |
| Bell pepper | 0.10 | 0.03 | 0.00 | 0.89 | 0.01 | 0.00 |
| Eyeshade | 0.00 | 0.00 | 0.03 | 0.00 | 0.92 | 0.00 |
| Bowl | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.98 |
In short, the embodiment of the present invention announcement is its preferable embodiment, but it is not limited to this.The technology of this field
Personnel can understand core of the invention thought, without departing from technical solution of the present invention easily according to above-described embodiment
Basis deformation or replacement, all within protection scope of the present invention.
Claims (10)
1. the building method of the Three dimensional convolution neural network of a kind of cloud object type identification, comprising the following steps:
The point cloud data of the object of step 1, acquisition known object type, and outlier is carried out to the point cloud data of the object and is deleted
It removes, obtains muting object point cloud data；The point cloud data of the object includes training set data, verifying collection data；
Step 2 carries out voxelization operation to obtained muting object point cloud data, obtains three-dimensional matrice, and to described three
Operation is normalized in dimension matrix, obtains the three-dimensional matrice being made of CP, and described by the three-dimensional matrice that CP is formed includes training set
The three-dimensional matrice being made of in the three-dimensional matrice and verifying collection data being made of in data CP CP；
It is characterized in that, further including following steps:
Step 3, using the three-dimensional matrice being made of in training set data CP as input data, input data is input to three-dimensional volume
In product neural network, the Three dimensional convolution neural network with input data is obtained, by the Three dimensional convolution nerve with input data
The layer of network is successively denoted as the 1st layer to the 32nd layer in accordance with the order from top to bottom；
Step 4 successively calculates the 1st layer to 19 layers of the Three dimensional convolution neural network with input data, obtains the 19th
The feature graphic sequence of layer；
By the 19th layer of characteristic pattern sequence inputting to the 20th layer of Three dimensional convolution neural network with input data, successively to the
20 layers to 30 layers are calculated, and the 30th layer of feature graphic sequence is obtained；
By the 30th layer of characteristic pattern sequence inputting to the 31st layer of Three dimensional convolution neural network with input data, successively to the
31 layers to the 32nd layer are calculated, and final feature vector is obtained；According to final feature vector, every kind of identification class is calculated
The probability and network losses value of type；
The probability of every kind of identification types and network losses value are input to the three-dimensional volume with input data by step 5
In product neural network, the new Three dimensional convolution neural network with input data is obtained；It is new with input data to what is obtained
Three dimensional convolution neural network carry out calculating update, obtain new Three dimensional convolution neural network.
2. the building method of the Three dimensional convolution neural network of point cloud object type identification, feature exist as described in claim 1
In further comprising the steps of:
Step 6 will be input in the new Three dimensional convolution neural network in verifying collection data by the three-dimensional matrice that CP is formed, according to
It is secondary to obtain first network losses value, second network losses value and third network losses value；
When obtained first network losses value, second network losses value and third network losses value are equal, or successively increase
When big, using obtained new Three dimensional convolution neural network as trained Three dimensional convolution neural network；
When obtained first network losses value, second network losses value and third network losses value are sequentially reduced, hold
Row operation identical with step 4-6.
3. a kind of point cloud object type recognition methods based on Three dimensional convolution neural network, specifically includes the following steps:
The point cloud data of step 1, acquisition object to be identified, and outlier is carried out to the point cloud data of the object to be identified and is deleted
It removes, obtains muting object point cloud data to be identified；
Step 2 carries out voxelization operation to obtained muting object point cloud data to be identified, obtains three-dimensional matrice, and right
Operation is normalized in the three-dimensional matrice, obtains the three-dimensional matrice being made of CP；
It is characterized in that, further comprising the steps of:
The three-dimensional matrice being made of CP value is input to new Three dimensional convolution neural network or trained Three dimensional convolution by step 3
In neural network, the Three dimensional convolution neural network with input data is obtained；Successively to the Three dimensional convolution mind with input data
The 1st layer to the 32nd layer through network is calculated, and final feature vector is obtained；According to final feature vector, it is calculated
The probability of every kind of identification types；According to every kind of obtained identification types probability, the type of object to be identified is finally obtained.
4. the point cloud object type recognition methods based on Three dimensional convolution neural network as described in claim 1, which is characterized in that
The 1st layer to 19 layers of the Three dimensional convolution neural network with input data are calculated described in step 4, concrete operations are such as
Under:
Step 41 is used as current layer for the 1st layer of the Three dimensional convolution neural network with input data；
Step 42 carries out convolution operation to current layer using formula (3), the output valve after obtaining convolution operation；
Wherein, x indicates output valve after convolution operation, x ∈ R；It is convolved in the three-dimensional matrice being made of in l expression training set CP
Channel position, 1≤l≤M；M represents the number for the characteristic pattern that current layer has, M >=3；XlIt is made of in expression training set CP
The value in the channel l in three-dimensional matrice, Xl∈ R, WlIndicate that the channel l convolution kernel is weighed in the three-dimensional matrice being made of in training set CP
Weight, Wl∈R；
To output valve after obtained convolution operation, carry out nonlinear transformation using formula (4), obtain it is non-linear after value；
Wherein, x indicates the output valve after convolution operation；Value after Relu (x) expression is non-linear, Relu (x) are more than or equal to 0；
To obtain it is non-linear after value carry out pondization operation, the pondization operates specifically: uses step-length for 2, size 2*
The window of 2*2 to obtain it is non-linear after value divide, obtain multiple windows；Choose in each window 8 numerical value most
Typical value of the big value as corresponding window, obtains the typical value of multiple windows；It is used as and is worked as by the typical value of obtained multiple windows
The feature graphic sequence of front layer, and will be in next layer of the characteristic pattern sequence inputting of current layer to current layer；By the next of current layer
Layer is used as current layer, executes operation identical with step 42, until the 19th layer has executed, to the Three dimensional convolution for having input data
The operation of 1st layer to 19 layers progress of neural network terminates.
5. the point cloud object type recognition methods based on Three dimensional convolution neural network as described in claim 1, which is characterized in that
The 20th layer to the 30th layer is calculated described in step 4, concrete operations are as follows:
Step 43 is used as current layer for the 20th layer；
Step 44 calculates current layer to learning method using residual error, obtains the feature graphic sequence of current layer, and will obtain
Current layer characteristic pattern sequence inputting into next layer of current layer；By next layer of current layer be used as current layer, execute with
The identical operation of step 44, until the 30th layer has executed, extremely to the 20th layer of the Three dimensional convolution neural network for having input data
The operation of 30th layer of progress terminates.
6. the point cloud object type recognition methods based on Three dimensional convolution neural network as described in claim 1, which is characterized in that
The 31st layer to 32 layers are calculated described in step 4, obtains final feature vector；According to final feature vector, calculate
Probability and network losses value, the concrete operations for obtaining every kind of identification types are as follows:
Step 45 uses full attended operation to the 31st layer, obtains the 31st layer of feature vector；31st layer of feature vector is inputted
To the 32nd layer, full attended operation is used to the 32nd layer, obtains final feature vector；
Step 46, according to demand and training set number of types, is arranged m kind identification types；Every kind of knowledge is calculated using formula (5)
The probability of other type,
Wherein, i is i-th kind of identification types, and m is the species number of identification types, 1≤i≤j≤m, m >=1；wiFor i-th kind of identification class
The weighted value of type, wi∈ R,Indicate the transposition of the weighted value of i-th kind of identification types,The final feature that V is
Vector, V ∈ R；PiIndicate the probability of i-th kind of identification types, 0≤Pi≤1；biIndicate the bias term of identification types in i-th, bi∈
R；bjRepresent the bias term of jth kind identification types, bj∈R；
Step 47 calculates network losses value:
Wherein, m is the species number of identification types, and i is i-th kind of identification types, 1≤i≤m, m >=1；PiIndicate i-th kind of identification types
Probability；When known object type is identical as i-th kind of identification types, yiIt is 1；When known object type and i-th kind of identification class
When not different, yiIt is 0.
7. the point cloud object type recognition methods based on Three dimensional convolution neural network as described in claim 1, which is characterized in that
Step 5 concrete operations are as follows:
The probability of every kind of identification types and network losses value are input to the Three dimensional convolution nerve with input data
In network, the new Three dimensional convolution neural network with input data is obtained；
Successively according to the 32nd layer to the 1st layer of sequence of Three dimensional convolution neural network, using gradient descent algorithm to described new
Three dimensional convolution neural network with input data is calculated, and each layer of weighted value is obtained；
Using each layer of obtained weighted value weighted value new as this layer, new Three dimensional convolution neural network is obtained.
8. the point cloud object type recognition methods based on Three dimensional convolution neural network as claimed in claim 2, which is characterized in that
Step 6 concrete operations are as follows:
Step 61 will be input in the new Three dimensional convolution neural network in verifying collection data by the three-dimensional matrice that CP is formed；
Step 62 executes operation identical with step 4, obtains first network losses value；The new Three dimensional convolution mind that will be obtained
Through network as the Three dimensional convolution neural network for having input data, operation identical with step 4-5, step 61-62 is executed, is obtained
To the Three dimensional convolution network of update and second network losses value；Using the Three dimensional convolution neural network of update as with input number
According to Three dimensional convolution neural network, execute identical with step 4-5, step 61-62 operation, obtain third network losses value；
When obtained first network losses value, second network losses value and third network losses value are equal, or successively increase
When big, using obtained new Three dimensional convolution neural network as trained Three dimensional convolution neural network；
When obtained first network losses value, second network losses value and third network losses value are sequentially reduced, hold
Row operation identical with step 4-6.
9. the point cloud object type recognition methods based on Three dimensional convolution neural network as claimed in claim 3, which is characterized in that
Successively the 1st layer to the 32nd layer of the Three dimensional convolution neural network with input data is calculated described in step 3, is obtained most
Whole feature vector, concrete operations are as follows:
Step 31 is used as current layer for the 1st layer of the Three dimensional convolution neural network with input data；
Step 32 carries out convolution operation to current layer using formula (6), the output valve after obtaining convolution operation；
Wherein, x indicates output valve after convolution operation, x ∈ R；It is convolved in the three-dimensional matrice being made of in l expression training set CP
Channel position, 1≤l≤M；M represents the number for the characteristic pattern that current layer has, M >=3；XlIt is made of in expression training set CP
The value in the channel l in three-dimensional matrice, Xl∈ R, WlIndicate that the channel l convolution kernel is weighed in the three-dimensional matrice being made of in training set CP
Weight, Wl∈R；
To output valve after obtained convolution operation, carry out nonlinear transformation using formula (7), obtain it is non-linear after value；
Wherein, x indicates the output valve after convolution operation；Value after Relu (x) expression is non-linear, Relu (x) are more than or equal to 0；
To obtain it is non-linear after value carry out pondization operation, the pondization operates specifically: uses step-length for 2, size 2*
The window of 2*2 to obtain it is non-linear after value divide, obtain multiple windows；Choose in each window 8 numerical value most
Typical value of the big value as corresponding window, obtains the typical value of multiple windows；It is used as and is worked as by the typical value of obtained multiple windows
The feature graphic sequence of front layer, and will be in next layer of the characteristic pattern sequence inputting of current layer to current layer；By the next of current layer
Layer is used as current layer, executes operation identical with step 32, until the 19th layer has executed, executes step 33；
Step 33 calculates current layer to learning method using residual error, obtains the feature graphic sequence of current layer, and will obtain
Current layer characteristic pattern sequence inputting into next layer of current layer；By next layer of current layer be used as current layer, execute with
The identical operation of step 33 executes step 34 until the 30th layer has executed；
Step 34, current layer is calculated using full attended operation, obtains the feature vector of current layer, and is current by what is obtained
The characteristic pattern vector of layer is input in next layer of current layer；It is used as current layer by next layer of current layer, is executed and step 34
Identical operation, until the 30th layer of feature vector is obtained, using the 30th layer of feature vector as final feature vector；
Step 35, according to demand and training set number of types, is arranged m kind identification types；Every kind of knowledge is calculated using formula (8)
The probability of other type；
Wherein, i is i-th kind of identification types, and m is the species number of identification types, 1≤i≤j≤m, m >=1；wiFor i-th kind of identification class
The weighted value of type, wi∈ R,Indicate the transposition of the weighted value of i-th kind of identification types,The final feature that V is
Vector, V ∈ R；PiIndicate the probability of i-th kind of identification types, 0≤Pi≤1；biIndicate the bias term of identification types in i-th, bi∈
R；bjRepresent the bias term of jth kind identification types, bj∈R。
10. the point cloud object type recognition methods based on Three dimensional convolution neural network, feature exist as claimed in claim 3
In every kind of identification types probability that basis obtains described in step 3 finally obtains the type of object to be identified, concrete operations are as follows:
Maximum value is chosen from the probability of every kind of obtained identification types, using the corresponding type of maximum value as object to be identified
Type obtains the type of final object to be identified.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910034966.XA CN109816714B (en) | 2019-01-15 | 2019-01-15 | Point cloud object type identification method based on three-dimensional convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910034966.XA CN109816714B (en) | 2019-01-15 | 2019-01-15 | Point cloud object type identification method based on three-dimensional convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109816714A true CN109816714A (en) | 2019-05-28 |
| CN109816714B CN109816714B (en) | 2023-03-21 |
Family
ID=66604434
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910034966.XA Active CN109816714B (en) | 2019-01-15 | 2019-01-15 | Point cloud object type identification method based on three-dimensional convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109816714B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490915A (en) * | 2019-08-19 | 2019-11-22 | 重庆大学 | A kind of point cloud registration method being limited Boltzmann machine based on convolution |
| CN110738194A (en) * | 2019-11-05 | 2020-01-31 | 电子科技大学中山学院 | three-dimensional object identification method based on point cloud ordered coding |
| CN112154448A (en) * | 2019-09-29 | 2020-12-29 | 深圳市大疆创新科技有限公司 | Target detection method and device and movable platform |
| CN112651986A (en) * | 2020-12-25 | 2021-04-13 | 北方工业大学 | Environment recognition method, recognition device, recognition system, electronic device, and medium |
| CN112949380A (en) * | 2021-01-08 | 2021-06-11 | 浙江大学 | Intelligent underwater target identification system based on laser radar point cloud data |
| WO2023096908A1 (en) * | 2021-11-23 | 2023-06-01 | Trustees Of Tufts College | Detection and identification of defects using artificial intelligence analysis of multi-dimensional information data |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104103093A (en) * | 2014-07-10 | 2014-10-15 | 北京航空航天大学 | Three-dimensional grid semantic marking method based on deep convolution neural network |
| CN107122396A (en) * | 2017-03-13 | 2017-09-01 | 西北大学 | Three-dimensional model searching algorithm based on depth convolutional neural networks |
| WO2017215284A1 (en) * | 2016-06-14 | 2017-12-21 | 山东大学 | Gastrointestinal tumor microscopic hyper-spectral image processing method based on convolutional neural network |
| CN107590432A (en) * | 2017-07-27 | 2018-01-16 | 北京联合大学 | A kind of gesture identification method based on circulating three-dimensional convolutional neural networks |
| CN107657285A (en) * | 2017-10-13 | 2018-02-02 | 哈尔滨工业大学 | Hyperspectral image classification method based on Three dimensional convolution neutral net |
| CN107862275A (en) * | 2017-11-01 | 2018-03-30 | 电子科技大学 | Human bodys' response model and its construction method and Human bodys' response method |
| CN108171217A (en) * | 2018-01-29 | 2018-06-15 | 深圳市唯特视科技有限公司 | A kind of three-dimension object detection method based on converged network |
| CN108717568A (en) * | 2018-05-16 | 2018-10-30 | 陕西师范大学 | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network |
| CN108921029A (en) * | 2018-06-04 | 2018-11-30 | 浙江大学 | A kind of SAR automatic target recognition method merging residual error convolutional neural networks and PCA dimensionality reduction |
| CN109118564A (en) * | 2018-08-01 | 2019-01-01 | 湖南拓视觉信息技术有限公司 | A kind of three-dimensional point cloud labeling method and device based on fusion voxel |
| WO2019001209A1 (en) * | 2017-06-28 | 2019-01-03 | 苏州比格威医疗科技有限公司 | Classification algorithm for retinal oct image based on three-dimensional convolutional neural network |
-
2019
- 2019-01-15 CN CN201910034966.XA patent/CN109816714B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104103093A (en) * | 2014-07-10 | 2014-10-15 | 北京航空航天大学 | Three-dimensional grid semantic marking method based on deep convolution neural network |
| WO2017215284A1 (en) * | 2016-06-14 | 2017-12-21 | 山东大学 | Gastrointestinal tumor microscopic hyper-spectral image processing method based on convolutional neural network |
| CN107122396A (en) * | 2017-03-13 | 2017-09-01 | 西北大学 | Three-dimensional model searching algorithm based on depth convolutional neural networks |
| WO2019001209A1 (en) * | 2017-06-28 | 2019-01-03 | 苏州比格威医疗科技有限公司 | Classification algorithm for retinal oct image based on three-dimensional convolutional neural network |
| CN107590432A (en) * | 2017-07-27 | 2018-01-16 | 北京联合大学 | A kind of gesture identification method based on circulating three-dimensional convolutional neural networks |
| CN107657285A (en) * | 2017-10-13 | 2018-02-02 | 哈尔滨工业大学 | Hyperspectral image classification method based on Three dimensional convolution neutral net |
| CN107862275A (en) * | 2017-11-01 | 2018-03-30 | 电子科技大学 | Human bodys' response model and its construction method and Human bodys' response method |
| CN108171217A (en) * | 2018-01-29 | 2018-06-15 | 深圳市唯特视科技有限公司 | A kind of three-dimension object detection method based on converged network |
| CN108717568A (en) * | 2018-05-16 | 2018-10-30 | 陕西师范大学 | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network |
| CN108921029A (en) * | 2018-06-04 | 2018-11-30 | 浙江大学 | A kind of SAR automatic target recognition method merging residual error convolutional neural networks and PCA dimensionality reduction |
| CN109118564A (en) * | 2018-08-01 | 2019-01-01 | 湖南拓视觉信息技术有限公司 | A kind of three-dimensional point cloud labeling method and device based on fusion voxel |
Non-Patent Citations (6)
| Title |
|---|
| DANIL PROKHOROV: "A Convolutional Learning System for Object Classification in 3-D Lidar Data", 《IEEE TRANSACTIONS ON NEURAL NETWORKS ( VOLUME: 21, ISSUE: 5, MAY 2010)》 * |
| LUÍS A. ALEXANDRE: "3D Object Recognition Using Convolutional Neural Networks with Transfer Learning Between Input Channels", 《INTELLIGENT AUTONOMOUS SYSTEMS》 * |
| 孙江冬: "基于激光雷达点云的车辆检测方法研究", 《中国优秀硕士学位论文全文数据库 (工程科技Ⅱ辑)》 * |
| 李羿辰: "基于区域扩张的三维点云模型配准算法", 《计算机工程》 * |
| 李鲜: "无人机三维立体重构与物体识别系统研究与实现", 《中国优秀硕士学位论文全文数据库 (工程科技Ⅱ辑)》 * |
| 杨军等: "基于深度体素卷积神经网络的三维模型识别分类", 《光学学报》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490915A (en) * | 2019-08-19 | 2019-11-22 | 重庆大学 | A kind of point cloud registration method being limited Boltzmann machine based on convolution |
| CN110490915B (en) * | 2019-08-19 | 2023-11-24 | 重庆大学 | Point cloud registration method based on convolution-limited Boltzmann machine |
| CN112154448A (en) * | 2019-09-29 | 2020-12-29 | 深圳市大疆创新科技有限公司 | Target detection method and device and movable platform |
| CN110738194A (en) * | 2019-11-05 | 2020-01-31 | 电子科技大学中山学院 | three-dimensional object identification method based on point cloud ordered coding |
| CN110738194B (en) * | 2019-11-05 | 2023-04-07 | 电子科技大学中山学院 | Three-dimensional object identification method based on point cloud ordered coding |
| CN112651986A (en) * | 2020-12-25 | 2021-04-13 | 北方工业大学 | Environment recognition method, recognition device, recognition system, electronic device, and medium |
| CN112651986B (en) * | 2020-12-25 | 2024-05-24 | 北方工业大学 | Environment recognition method, recognition device, recognition system, electronic equipment and medium |
| CN112949380A (en) * | 2021-01-08 | 2021-06-11 | 浙江大学 | Intelligent underwater target identification system based on laser radar point cloud data |
| CN112949380B (en) * | 2021-01-08 | 2022-06-10 | 浙江大学 | Intelligent underwater target identification system based on laser radar point cloud data |
| WO2023096908A1 (en) * | 2021-11-23 | 2023-06-01 | Trustees Of Tufts College | Detection and identification of defects using artificial intelligence analysis of multi-dimensional information data |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109816714B (en) | 2023-03-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109816714A (en) | A kind of point cloud object type recognition methods based on Three dimensional convolution neural network | |
| CN109559320A (en) | Realize that vision SLAM semanteme builds the method and system of figure function based on empty convolution deep neural network | |
| CN106778604B (en) | Pedestrian re-identification method based on matching convolutional neural network | |
| CN111489358A (en) | Three-dimensional point cloud semantic segmentation method based on deep learning | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN109859171A (en) | A kind of flooring defect automatic testing method based on computer vision and deep learning | |
| CN110276264B (en) | Crowd density estimation method based on foreground segmentation graph | |
| CN108710919A (en) | A kind of crack automation delineation method based on multi-scale feature fusion deep learning | |
| CN108648233A (en) | A kind of target identification based on deep learning and crawl localization method | |
| CN111783772A (en) | Grabbing detection method based on RP-ResNet network | |
| CN108564565A (en) | A kind of power equipment infrared image multi-target orientation method based on deep learning | |
| CN108647655A (en) | Low latitude aerial images power line foreign matter detecting method based on light-duty convolutional neural networks | |
| CN106446930A (en) | Deep convolutional neural network-based robot working scene identification method | |
| CN106127702A (en) | A kind of image mist elimination algorithm based on degree of depth study | |
| CN113420643B (en) | Lightweight underwater target detection method based on depth separable cavity convolution | |
| CN104992223A (en) | Intensive population estimation method based on deep learning | |
| CN108074243A (en) | A kind of cellular localization method and cell segmentation method | |
| CN110334584B (en) | Gesture recognition method based on regional full convolution network | |
| CN105955708A (en) | Sports video lens classification method based on deep convolutional neural networks | |
| CN111127423B (en) | Rice pest and disease identification method based on CNN-BP neural network algorithm | |
| CN108875819B (en) | Object and component joint detection method based on long-term and short-term memory network | |
| CN113705641A (en) | Hyperspectral image classification method based on rich context network | |
| CN111368637B (en) | Transfer robot target identification method based on multi-mask convolutional neural network | |
| CN108154158B (en) | Building image segmentation method for augmented reality application | |
| CN109919246A (en) | Pedestrian's recognition methods again based on self-adaptive features cluster and multiple risks fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |