CN115661767A - Image front vehicle target identification method based on convolutional neural network - Google Patents

Abstract

The application relates to the technical field of vehicle target identification, in particular to a method for identifying a vehicle target in front of an image based on a convolutional neural network, wherein the method comprises the following steps: acquiring an environment image around a vehicle; the method comprises the steps of inputting an environmental image into a pre-trained target detection model, and outputting a target recognition result of the environmental image, wherein the target detection model comprises a residual error network, a feature extraction network and a prediction network, the residual error network comprises a plurality of feature layers and a dense connection network, the dense connection network is used for splicing the output features of the current feature layer with the output features of all feature layers before the current feature layer to serve as the input of the next feature layer, extracting a feature map of the environmental image by using the residual error network, inputting the feature map into the feature extraction network, outputting a fusion feature, inputting the fusion feature into the prediction network, and outputting the target recognition result of the environmental image. Therefore, the problems of difficulty in detecting small targets in a road scene, low detection rate and the like in the related art are solved.

Description

A method for vehicle target recognition in front of images based on convolutional neural network

technical field

The present application relates to the technical field of vehicle target recognition, in particular to a method for recognizing a vehicle target in front of an image based on a convolutional neural network.

Background technique

YOLO (You Only Look Once) is a network for target detection. From the initial YOLOV1 to the current YOLOV4, the detection accuracy and speed of its algorithm have been significantly improved. The network structure of YOLOV4 can be divided into three parts: the backbone feature extraction network CSPDarknet53, the enhanced feature extraction network and the prediction network Yolo Head.

The structure of the backbone feature extraction network CSPDarknet53 is from YOLOV3. The ResNet (Residual Neural Network, residual network) structure uses the CSPNet (Cross Stage Partial Network, cross-stage local network) structure, but the feature extraction ability still needs to be strengthened. In addition, the number and size of the pooling cores used by the SPP (Spatial Pyramid Pooling) module in YOLOV4 cannot fully integrate the information of multi-scale perception on the large feature map, and the performance improvement of its target detection is limited.

Contents of the invention

The present application provides a vehicle target recognition method, device, vehicle and storage medium to solve the problems in related technologies such as difficulty in detecting small targets in road scenes and low detection rate.

The embodiment of the first aspect of the present application provides a vehicle target recognition method, including the following steps: acquiring an environmental image around the vehicle; inputting the environmental image into a pre-trained target detection model, and outputting the target recognition result of the environmental image , wherein the target detection model includes a residual network, a feature extraction network and a prediction network, the residual network includes a plurality of feature layers and a densely connected network, and the densely connected network is used to combine the output features of the current feature layer with splicing the output features of all feature layers before the current feature layer, as the input of the next feature layer, using the residual network to extract the feature map of the environment image, and inputting the feature map into the feature extraction network, Outputting the fusion feature, and inputting the fusion feature into the prediction network, and outputting the target recognition result of the environment image.

Optionally, the feature extraction network includes a feature pyramid network and a pooling network, wherein the output features in the feature layer and the pooling network are respectively provided with an attention mechanism network between the feature pyramid network, Use the pooling network to convert the feature map into a feature vector, and use the attention mechanism network to input the feature map and the feature vector into the feature pyramid network, and output the image feature and the feature A vector of fused features.

Optionally, the attention mechanism of the attention mechanism network includes: performing a global average pooling operation on the feature images of each channel output by the feature layer to obtain a new feature map of each channel; The convolution operation of k, and the weight of each channel is obtained through the preset activation function, and the cross-channel interaction information is obtained based on the new feature map and weight of each channel.

Optionally, the pooling network includes multiple convolution kernels, where the multiple convolution kernels may be 1×1, 4×4, 7×7, 10×10, and 13×13 respectively.

The embodiment of the second aspect of the present application provides a vehicle target recognition device, including: an acquisition module, used to acquire the environment image around the vehicle; a recognition module, used to input the environment image into the pre-trained target detection model, and output The target recognition result of the environment image, wherein the target detection model includes a residual network, a feature extraction network and a prediction network, and the residual network includes a plurality of feature layers and a densely connected network, and the densely connected network is used for Splicing the output features of the current feature layer with the output features of all feature layers before the current feature layer, as the input of the next feature layer, using the residual network to extract the feature map of the environment image, and combining the feature The graph is input into the feature extraction network, and the fusion feature is output, and the fusion feature is input into the prediction network, and the target recognition result of the environment image is output.

Optionally, the feature extraction network includes a feature pyramid network and a pooling network, wherein the output features in the feature layer and the pooling network are respectively provided with an attention mechanism network between the feature pyramid network, Use the pooling network to convert the feature map into a feature vector, and use the attention mechanism network to input the feature map and the feature vector into the feature pyramid network, and output the image feature and the feature A vector of fused features.

Optionally, the attention mechanism of the attention mechanism network includes: performing a global average pooling operation on the feature images of each channel output by the feature layer to obtain a new feature map of each channel; The convolution operation of k, and the weight of each channel is obtained through the preset activation function, and the cross-channel interaction information is obtained based on the new feature map and weight of each channel.

Optionally, the pooling network includes multiple convolution kernels, where the multiple convolution kernels may be 1×1, 4×4, 7×7, 10×10, and 13×13 respectively.

The embodiment of the third aspect of the present application provides a vehicle, including: a memory, a processor, and a computer program stored on the memory and operable on the processor, and the processor executes the program to implement the following: The vehicle object recognition method described in the above embodiments.

The embodiment of the fourth aspect of the present application provides a computer-readable storage medium, on which a computer program is stored, and the program is executed by a processor, so as to realize the object recognition method for a vehicle as described in the above-mentioned embodiment.

Thus, the application at least has the following beneficial effects:

Dense connections are added to the residual network structure, which improves the feature extraction ability of the network, especially for small targets, and reduces the amount of parameters; adding the attention mechanism to the output feature layer of the backbone feature network, without improving At the same time as the depth of the network, a large amount of information is used to selectively add features, which improves the accuracy of the algorithm; the size of the convolution kernel of the pooling layer is changed, and the fusion of multi-scale sensory field information of the feature map is improved. After the pooling layer module, the neural network The contained image feature information is richer, the information contained is more complex, and the image shows more detailed features. As a result, technical problems such as difficulty in detecting small targets in a road scene and low detection rate in related technologies are solved.

Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the application.

Description of drawings

The above and/or additional aspects and advantages of the present application will become apparent and easy to understand from the following description of the embodiments in conjunction with the accompanying drawings, wherein:

FIG. 1 is a flow chart of a vehicle target recognition method provided according to an embodiment of the present application;

FIG. 2 is a diagram of a dense connection structure added to a residual structure provided according to an embodiment of the present application;

FIG. 3 is a diagram of an ECA (Efficient Channel Attention) attention mechanism provided according to an embodiment of the present application;

Fig. 4 is according to the YOLOV4 target detection model figure in the related art;

FIG. 5 is a diagram of an improved YOLOV4 target detection model provided according to an embodiment of the present application;

FIG. 6 is a comparison diagram of P-R curves for vehicle detection provided according to an embodiment of the present application;

FIG. 7 is a flow chart of a vehicle target recognition method provided according to an embodiment of the present application;

FIG. 8 is an example diagram of a target recognition device for a vehicle provided according to an embodiment of the present application;

Fig. 9 is a schematic structural diagram of a vehicle provided according to an embodiment of the present application.

Detailed ways

Embodiments of the present application are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary, and are intended to explain the present application, and should not be construed as limiting the present application.

Related technologies use YOLO to achieve target detection. YOLO is a network for target detection. From the initial YOLOV1 to the current YOLOV4, the detection accuracy and speed of its algorithm have been significantly improved. The network structure of YOLOV4 can be divided into three parts: the backbone feature extraction network CSPDarknet53; the enhanced feature extraction network, including SPP and PANet (PathAggregation Network, path aggregation network); the prediction network Yolo Head, which uses the extracted features for result prediction. But there are still following disadvantages:

1. The backbone feature extraction network CSPDarknet53 is modified from the ResNet structure in YOLOV3 to the CSPResNet structure. The CSPResNet structure divides the input layer into two parts and uses convolution operations respectively. The right part uses the Resblock residual block structure for feature extraction, and the left part is directly spliced with the output feature map of the right part after convolution, regularization and activation function processing, but its feature extraction ability still needs to be strengthened.

2. When using a deep CNN (Convolution Neural Network, convolutional neural network) model to identify images, the local information of the image is generally extracted through the convolution kernel. However, the influence of each local information on whether the image can be correctly recognized How to let the model know the importance of different local information in the image is a more critical issue. The backbone network does not add an attention mechanism, which will cause the weight index to shift during the training process. Adding an attention mechanism can enhance the combination of feature fusion and channel and spatial attention, and improve the ability to extract features.

3. With the deepening of the network level of the convolutional neural network, the image information of the deep feature map is highly abstract, the semantic information of the image increases, and the direct feature information of the image is lost, resulting in the detection of small targets using the deep feature map of the neural network. The accuracy of the model need to be improved. The SPP module structure can realize the fusion of multi-scale local features and global features, and enrich the expressive ability of feature maps. The number and size of pooling kernels used by the SPP module in YOLOV4 cannot fully integrate multi-scale receptive field information on large feature maps, and its target detection performance is limited.

The vehicle object recognition method, device, vehicle and storage medium according to the embodiments of the present application will be described below with reference to the accompanying drawings. Aiming at the problem that the current network YOLOV4 used for target detection mentioned in the above background technology has poor target feature extraction ability and low target detection performance, this application provides a vehicle target recognition method, in which , improve the YOLOV4 network structure, integrate the dense connection structure into the CSPResNet structure, and fully reuse the shallow feature information; in the channel attention mechanism, a local cross-channel interaction strategy without dimensionality reduction is adopted to avoid dimensionality reduction for channel attention The influence of force learning effect; the size of the pooling kernel in YOLOV4 is refined to make more detailed features on the image. Thus, problems such as difficulty in detecting small target vehicles, low detection rate, and poor detection performance in the related art are solved when vehicle target detection is performed.

Specifically, FIG. 1 is a schematic flowchart of a vehicle object recognition method provided by an embodiment of the present application.

As shown in Figure 1, the vehicle target recognition method includes the following steps:

In step S101, an environment image around the vehicle is acquired.

Wherein, in the embodiment of the present application, the environment image around the vehicle may be obtained through the vehicle-mounted camera, or may be obtained through other methods, which are not specifically limited.

In step S102, the environment image is input into the pre-trained target detection model, and the target recognition result of the environment image is output, wherein the target detection model includes a residual network, a feature extraction network and a prediction network, and the residual network includes multiple feature layers And densely connected network, the densely connected network is used to splice the output features of the current feature layer with the output features of all feature layers before the current feature layer, as the input of the next feature layer, and use the residual network to extract the feature map of the environment image, Input the feature map into the feature extraction network, output the fusion feature, and input the fusion feature into the prediction network, and output the target recognition result of the environment image.

It can be understood that, in the embodiment of the present application, the environment image may be input into a pre-trained object detection model to obtain the object recognition result of the output environment image.

Among them, the residual network includes multiple feature layers and a densely connected network; the prediction network is used to predict the result by using the extracted features, and output the recognition result of the environment image.

It can be understood that in the embodiment of the present application, a densely connected network is added to the residual network. As shown in Figure 2, all the previous feature layers are densely connected to the current feature layer. Starting from the input feature layer, each layer is used as the following The input of each layer enables the shallow feature information to be fully reused and improves the ability to detect small target objects. In a densely connected network, the input features of the n+1th layer, as shown in the formula:

x _n = _n ([x ₀ , x ₁ , . . . , x _n-2 , x _n-1 ]),

Among them, [x ₀ , x ₁ , ..., x _n-2 , x _n-1 ] is the new input x _n-1 and all previous input feature layers [x ₀ , x ₁ , ..., x _n-3 , x _n-2 | Splicing in the channel direction; H _n is the conversion function of the nth layer; x _n is the output feature of the nth layer.

In the densely connected network, the real input feature layer of the nth layer is the new input x _n-1 and all previous input feature layers x ₀ , x ₁ ,..., x _n-3 , x _n-2 in the channel direction Splicing, namely: [x ₀ , x ₁ , ..., x _n-3 , x _n-2 ]. The input is subjected to regularization, activation function and convolution operation to obtain the output feature layer x _n of the nth layer, as shown in the formula:

x _n = H _n ([x ₀ , x ₁ , . . . , x _n-2 , x _n-1 ]),

Among them, x _n and all previous input feature layers are spliced into [x ₀ , x ₁ ,..., x _n-2 , x _n-1 ] as the input of the n+1th layer. The number k of new channels in each layer is usually not too large, and the number of new channels k is called the growth rate. Since the splicing method is used instead of the linear addition used in the residual network, and the number of channels added each time k is small, it seems that the connection density is more complicated, but the actual number of parameters and calculations are more complex than the residual network. Poor network is less.

It should be noted that the embodiment of this application aims at the problem of insufficient utilization of some shallow feature information in the residual block structure used by the YOLOV4 backbone feature extraction network CSPDarknet53, and uses densely connected blocks with the same number of layers as the original residual stacking structure Dense block. The connection block meets the requirement that the output size of the stacked residual block in the skip connection of the CSPNet structure is consistent with the size of the input feature map, and will not change the size of the input feature map. This structure can be realized by dense connection block Denseblock, and the downsampling operation of the transition layer (Transition Layer) is not required.

In the embodiment of the present application, the feature extraction network includes a feature pyramid network and a pooling network, wherein an attention mechanism network is set between the output features in the feature layer and the pooling network and the feature pyramid network, and the pooling network is used to The feature map is converted into a feature vector, and the feature map and feature vector are input into the feature pyramid network by using the attention mechanism network, and the fusion feature of the image feature and feature vector is output.

It should be noted that since the channel attention mechanism has been proven to have great potential in improving the performance of deep convolutional neural networks, the output features in the feature layer and the pooling network in the embodiment of the present application are respectively equipped with an attention mechanism.

It can be understood that the embodiment of the present application can use the pooling network to convert the feature map into a feature vector, and use the attention mechanism network to input the feature map and feature vector into the feature pyramid network, so as to output the fusion feature of image features and feature vectors, In order to facilitate the subsequent use of fusion features to output the target recognition results of the environmental image.

In the embodiment of the present application, the attention mechanism of the attention mechanism network includes: performing a global average pooling operation on the feature images of each channel output by the feature layer to obtain a new feature map of each channel; The convolution operation is k, and the weight of each channel is obtained through the preset activation function, and the cross-channel interaction information is obtained based on the new feature map and weight of each channel.

Wherein, the preset activation function refers to a function that can calculate the weight of each channel, for example, it can be a Sigmod activation function.

Specifically, the embodiment of the present application adds an attention mechanism to the output feature layer of the backbone network, as shown in Figure 3, ECA-Net (Efficient Channel Attention for Deep Convolutional Neural Networks, lightweight module attention mechanism) according to the importance of the channel Different weights are assigned to it, and the cross-channel interaction information is obtained in a lightweight way by adaptively determining the local cross-channel interaction range for modeling. Specific steps are as follows:

1. Perform a global average pooling operation on the input feature layer;

2. Perform a 1-dimensional convolution operation with a convolution kernel size of k, and obtain the weight ω of each channel through the Sigmod activation function;

3. Multiply the weight with the corresponding element of the original input feature map.

It can be understood that the ECA module of the efficient channel attention mechanism in the embodiment of the present application adopts a local cross-channel interaction strategy without dimensionality reduction, which can effectively avoid the influence of dimensionality reduction on the learning effect of channel attention. Channel interaction can significantly reduce model complexity while maintaining performance.

In the embodiment of the present application, the pooling network includes multiple convolution kernels, wherein the multiple convolution kernels are 1×1, 4×4, 7×7, 10×10, and 13×13, respectively.

It should be noted that the original YOLOV4 pooling layer convolution kernels are 1×1, 5×5, 9×9, 13×13 respectively, and the embodiment of this application refines the pooling kernel size to 1×1, 4 ×4, 7×7, 10×10, 13×13, effectively improving the multi-scale sensory field of view of the feature map, after the refinement of the pooling layer module, the image feature information contained in the neural network is more abundant, and the included The information is more complex, and there are more detailed features on the image.

Specifically, this application improves the original YOLOV4 target model (as shown in Figure 4), and the improved YOLOV4 target detection model is shown in Figure 5, and the improvement method is specifically divided into the following three aspects:

1. In view of the difficulties in detecting small target vehicles and low detection rate in machine vision detection vehicles, the network structure of YOLOV4 is improved. The improved YOLOV4 network structure draws on the advantages of densely connected network feature extraction, incorporates densely connected structure into the CSPResNet structure, and fully reuses the shallow feature information; the densely connected network is to establish a short-circuit connection between the front feature layer and the back feature layer.

2. The channel attention mechanism has been shown to have great potential in improving the performance of deep convolutional neural. The Efficient Channel Attention (ECA) module adopts a local cross-channel interaction strategy without dimensionality reduction, which effectively avoids the influence of dimensionality reduction on channel attention learning effects. Proper cross-channel interaction can significantly reduce model complexity while maintaining performance.

3. The original YOLOV4 pooling layer convolution kernels are 1×1, 5×5, 9×9, 13×13, and the pooling kernel size is refined to 1×1, 4×4, 7×7, 10 ×10, 13×13, enhanced receptive field range. After refining the pooling layer module, the image feature information contained in the neural network is more abundant, the information contained is more complex, and the image shows more detailed features.

As shown in Figure 6, the P-R curve comparison between the original YOLOV4 and the improved YOLOV4, the improved YOLOV4 algorithm of this application has a better detection effect on small targets in road scenes due to the original YOLOV4 algorithm, which has a lower missed detection rate and is better at detecting targets. In terms of confidence, there are also differences in the results before and after the improvement. For the detection target of YOLOV4, the confidence value of the improved YOLOV4 output is generally higher, which shows that the improved network has stronger detection ability and can pay more attention to the key information of the target. , The detection speed is also slightly improved. Compared with the original YOLOV4, the improved YOLOV4 algorithm has a slightly higher detection speed, and the average accuracy has increased by 2.61%, reaching 92.63%. The improved YOLOV4 algorithm has better performance in small target detection, and the missed detection rate lower.

The target recognition method of the vehicle will be described below through a specific embodiment, as shown in Figure 7, the steps are as follows:

1. Collect images of the surrounding environment of the vehicle and make a data set;

2. Divide the data set into validation set, training set and test set;

3. Construct a vehicle detection model based on improved YOLOV4, which adds dense connection, ECA attention mechanism and different pooling convolution kernels;

4. Train and tune the model;

5. Use the verification set to evaluate the performance of the trained front vehicle model;

6. Build a development platform to read the monocular camera and perform video prediction in the model.

According to the vehicle target recognition method proposed in the embodiment of the present application, dense connections are added to the residual network structure, which improves the feature extraction capability of the network, especially for small targets, and reduces the amount of parameters; it will be noted that The force mechanism is added to the output feature layer of the backbone feature network. While not increasing the depth of the network, a large amount of information is used to selectively increase the feature, which improves the accuracy of the algorithm; the size of the convolution kernel of the pooling layer is changed, and the multi-scale visual field information of the feature map is improved. After the fusion of the pooling layer module, the image feature information contained in the neural network is more abundant, the information contained is more complex, and there are more detailed features on the image.

Next, the object recognition device for a vehicle proposed according to the embodiment of the present application will be described with reference to the accompanying drawings.

FIG. 8 is a schematic block diagram of an object recognition device for a vehicle according to an embodiment of the present application.

As shown in FIG. 8 , the object recognition device 10 of the vehicle includes: an acquisition module 100 and a detection module 200 .

Wherein, the acquisition module 100 is used to acquire the environment image around the vehicle; the detection module 200 is used to input the environment image into the pre-trained target detection model, and output the target recognition result of the environment image, wherein the target detection model includes a residual network, a feature Extraction network and prediction network, the residual network includes multiple feature layers and dense connection network, the dense connection network is used to splice the output features of the current feature layer with the output features of all feature layers before the current feature layer, as the next feature layer input, use the residual network to extract the feature map of the environment image, input the feature map into the feature extraction network, output the fusion feature, and input the fusion feature into the prediction network, and output the target recognition result of the environment image.

In the embodiment of the present application, the feature extraction network includes a feature pyramid network and a pooling network, wherein an attention mechanism network is set between the output features in the feature layer and the pooling network and the feature pyramid network, and the pooling network is used to The feature map is converted into a feature vector, and the feature map and feature vector are input into the feature pyramid network by using the attention mechanism network, and the fusion feature of the image feature and feature vector is output.

In the embodiment of the present application, the attention mechanism of the attention mechanism network includes: performing a global average pooling operation on the feature images of each channel output by the feature layer to obtain a new feature map of each channel; The convolution operation is k, and the weight of each channel is obtained through the preset activation function, and the cross-channel interaction information is obtained based on the new feature map and weight of each channel.

In the embodiment of the present application, the pooling network includes multiple convolution kernels, wherein the multiple convolution kernels are 1×1, 4×4, 7×7, 10×10, and 13×13, respectively.

It should be noted that the foregoing explanations of the embodiment of the vehicle object recognition method are also applicable to the vehicle object recognition device of this embodiment, and will not be repeated here.

According to the vehicle target recognition device proposed in the embodiment of the present application, dense connections are added to the residual network structure, which improves the feature extraction capability of the network, especially for small targets, and reduces the amount of parameters; it will be noted The force mechanism is added to the output feature layer of the backbone feature network. While not increasing the depth of the network, a large amount of information is used to selectively increase the feature, which improves the accuracy of the algorithm; the size of the convolution kernel of the pooling layer is changed, and the multi-scale visual field information of the feature map is improved. After the fusion of the pooling layer module, the image feature information contained in the neural network is more abundant, the information contained is more complex, and there are more detailed features on the image.

FIG. 9 is a schematic structural diagram of a vehicle provided by an embodiment of the present application. The vehicle can include:

A memory 901 , a processor 902 , and a computer program stored in the memory 901 and executable on the processor 902 .

When the processor 902 executes the program, the object recognition method for the vehicle provided in the above-mentioned embodiments is implemented.

Further, the vehicle also includes:

The communication interface 903 is used for communication between the memory 901 and the processor 902 .

The memory 901 is used to store computer programs that can run on the processor 902 .

The memory 901 may include a high-speed RAM (Random Access Memory, random access memory) memory, and may also include a non-volatile memory, such as at least one disk memory.

If the memory 901, the processor 902, and the communication interface 903 are independently implemented, the communication interface 903, the memory 901, and the processor 902 may be connected to each other through a bus to complete mutual communication. The bus may be an ISA (Industry Standard Architecture, industry standard architecture) bus, a PCI (Peripheral Component, external device interconnection) bus, or an EISA (Extended Industry Standard Architecture, extended industry standard architecture) bus, etc. The bus can be divided into address bus, data bus, control bus and so on. For ease of representation, only one thick line is used in FIG. 9 , but it does not mean that there is only one bus or one type of bus.

Optionally, in terms of specific implementation, if the memory 901, processor 902, and communication interface 903 are integrated on one chip, then the memory 901, processor 902, and communication interface 903 can communicate with each other through the internal interface.

The processor 902 may be a CPU (Central Processing Unit, central processing unit), or an ASIC (Application Specific Integrated Circuit, specific integrated circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

The embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the above object recognition method for a vehicle is realized.

In the description of this specification, descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific examples", or "some examples" mean that specific features described in connection with the embodiment or example , structure, material or characteristic is included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics may be combined in any one or N embodiments or examples in an appropriate manner. In addition, those skilled in the art can combine and combine different embodiments or examples and features of different embodiments or examples described in this specification without conflicting with each other.

In addition, the terms "first" and "second" are used for descriptive purposes only, and cannot be interpreted as indicating or implying relative importance or implicitly specifying the quantity of indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include at least one of these features. In the description of the present application, "N" means at least two, such as two, three, etc., unless otherwise specifically defined.

Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing a custom logical function or step of a process , and the scope of preferred embodiments of the present application includes additional implementations in which functions may be performed out of the order shown or discussed, including in substantially simultaneous fashion or in reverse order depending on the functions involved, which shall It should be understood by those skilled in the art to which the embodiments of the present application belong.

It should be understood that each part of the present application may be realized by hardware, software, firmware or a combination thereof. In the above embodiments, the N steps or methods may be implemented by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented by any one or a combination of the following techniques known in the art: a discrete Logic circuits, ASICs with suitable combinational logic gates, programmable gate arrays, field programmable gate arrays, etc.

Those of ordinary skill in the art can understand that all or part of the steps carried by the methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. During execution, one or a combination of the steps of the method embodiments is included.

Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limitations on the present application, and those skilled in the art can make the above-mentioned The embodiments are subject to changes, modifications, substitutions and variations.

Claims (10)

1. A target recognition method for a vehicle, comprising the following steps: Obtain an image of the environment around the vehicle; Inputting the environmental image into a pre-trained target detection model, outputting the target recognition result of the environmental image, wherein the target detection model includes a residual network, a feature extraction network and a prediction network, and the residual network includes multiple A feature layer and a densely connected network, the densely connected network is used to splice the output features of the current feature layer with the output features of all feature layers before the current feature layer, as the input of the next feature layer, using the residual The difference network extracts the feature map of the environment image, inputs the feature map into the feature extraction network, outputs fusion features, and inputs the fusion features into the prediction network, and outputs the target recognition result of the environment image. 2. The method according to claim 1, wherein the feature extraction network comprises a feature pyramid network and a pooling network, wherein the output features in the feature layer and the pooling network are respectively related to the feature An attention mechanism network is set between the pyramid networks, and the feature map is converted into a feature vector by using the pooling network, and the feature map and the feature vector are input into the feature pyramid by using the attention mechanism network A network that outputs a fusion feature of the image feature and the feature vector. 3. The method according to claim 2, wherein the attention mechanism of the attention mechanism network comprises: Perform a global average pooling operation on the feature images of each channel output by the feature layer to obtain a new feature map of each channel; A convolution operation with a convolution kernel size of k is performed on each channel, and the weight of each channel is obtained through a preset activation function, and cross-channel interaction information is obtained based on the new feature map and weight of each channel. 4. The method according to claim 2, wherein the pooling network comprises a plurality of convolution kernels, wherein the plurality of convolution kernels are 1×1, 4×4, 7×7, 10×10 and 13×13. 5. A target recognition device for a vehicle, comprising: An acquisition module, configured to acquire environmental images around the vehicle; A detection module, configured to input the environmental image into a pre-trained target detection model, and output the target recognition result of the environmental image, wherein the target detection model includes a residual network, a feature extraction network and a prediction network, the The residual network includes multiple feature layers and a densely connected network, and the densely connected network is used to splice the output features of the current feature layer with the output features of all feature layers before the current feature layer, as the input of the next feature layer , using the residual network to extract the feature map of the environment image, input the feature map into the feature extraction network, output the fusion feature, and input the fusion feature into the prediction network, and output the feature map of the environment image Target recognition results. 6. The device according to claim 5, wherein the feature extraction network comprises a feature pyramid network and a pooling network, wherein the output features in the feature layer and the pooling network are respectively related to the feature An attention mechanism network is set between the pyramid networks, and the feature map is converted into a feature vector by using the pooling network, and the feature map and the feature vector are input into the feature pyramid by using the attention mechanism network A network that outputs a fusion feature of the image feature and the feature vector. 7. The device according to claim 6, wherein the attention mechanism of the attention mechanism network comprises: performing a global average pooling operation on the feature images of each channel output by the feature layer to obtain new features of each channel Figure: Perform a convolution operation with a convolution kernel size of k on each channel, and obtain the weight of each channel through a preset activation function, and obtain cross-channel interaction information based on the new feature map and weight of each channel. 8. The device according to claim 6, wherein the pooling network comprises a plurality of convolution kernels, wherein the plurality of convolution kernels are 1×1, 4×4, 7×7, 10×10 and 13×13. 9. A vehicle, characterized in that it comprises: a memory, a processor, and a computer program stored on the memory and operable on the processor, the processor executes the program, so as to realize the The target recognition method for the vehicle described in any one of 1-4. 10. A computer-readable storage medium, on which a computer program is stored, wherein the program is executed by a processor, so as to implement the vehicle object recognition method according to any one of claims 1-4.

Images