CN103218824A - Motion key frame extracting method based on distance curve amplitudes - Google Patents

Abstract

This paper invents a motion key frame extraction method based on the distance curve amplitude, and extracts the key pose of the motion to describe the original motion sequence. The first step is to select a new set of joint distances as the distance feature, and the second step is to use the principal component analysis method (PCA method for short) to reduce the dimensionality of this distance feature, extract the first-dimensional principal component, and use smoothing filter to remove noise to obtain the feature curve, which can better reflect the essential characteristics of the original motion. The third step obtains the initial key frame by extracting the local extremum points on the characteristic curve. The fourth step is between the initial adjacent key frames, according to the characteristic curve The amplitude difference is evenly inserted into the corresponding number of frames, and then the dense key frames are merged to obtain the final key frame set. A large amount of experimental data proves that the present invention not only satisfies the good effect on the visual generalization of motion, but also has a lower compression rate and error ratio, and it can be completed automatically without manual intervention in the processing process.

Description

Motion Keyframe Extraction Based on Distance Curve Amplitude

technical field

The present invention relates to human body motion capture technology, more specifically, relates to key frame extraction of human body motion.

Background technique

In recent decades, human motion capture technology has developed rapidly, and the importance of motion capture has also increased. Motion capture systems are already widely used in movies and games. Accompanied by a problem, due to the huge amount of captured data, the size of the motion capture database is also very large. How to deal with huge motion capture data has become a hot research topic at home and abroad.

Key frame technology is an effective solution. Select the most important and critical frame in the motion as the key frame, representing the entire motion sequence, while other non-key frames are not as important as these key frames, which can be calculated by the key frame through an interpolation algorithm get. Because the key frame extraction technology plays an irreplaceable role in the representation of the entire motion sequence, it can not only speed up the data processing speed, but also has obvious advantages in the storage, compression, browsing and reuse of motion capture data.

That is to say, "key frame extraction of human motion" is to automatically extract a certain number of key motion postures for a motion sequence, which has a better visual generalization for this motion, and at the same time can perform motion reconstruction and restore the original frame. Exercise, keep a low error rate.

Key frame extraction should meet the following requirements. On the one hand, key frames under a certain compression rate can effectively summarize the original motion sequence. On the other hand, keyframes can be used to reconstruct the original motion sequence as accurately as possible. In addition to subjective visual judgment, there are two evaluation criteria for the quality of key frame extraction: error rate and compression ratio.

So far, key frame extraction methods include uniform sampling, curve simplification, clustering-based methods, etc.

Human body motion capture data is in BVH format, and its motion data is composed of frame-by-frame motion data, and each frame of data contains motion posture information, and each posture is composed of all relevant nodes of the human body. Import the BVH data into MATLAB to display the human skeleton model shown in Figure 2. The model contains 31 joint points (the joint points to be used are marked), each joint point adopts a tree structure, and the root node (root ) is the root node of the tree-shaped human skeleton, and extends layer by layer from the root joint point to each end joint of the human skeleton to form each subtree of the root node. Among them, the root joint is represented by 3 translations and 3 rotations, and the other non-root joints are represented by 3 rotations. There are 96 degrees of freedom in total. The translation of the root determines the current position of the human body, and the rotation of the root determines the orientation of the human body; the rotation of each other joint point indicates the direction of the joint point in the local coordinate system where its parent joint point is located, and they together determine the human body posture.

j=2,…,31表示非root关节点的方向。The data of human body motion capture is a sequence of human body postures obtained by sampling discrete time points, each sampling point is a frame, and the posture of each frame is jointly determined by 31 joint points. In this way, at any time i, the human body pose is expressed as $f_i=(p_i^{\left(1\right)},r_i^{\left(1\right)},r_i^{\left(2\right)},....,r_i^{\left(31\right)}),$ in $p_i^{\left(1\right)}\&Element;{\mathrm{<figure-callout\; id="3"\; label="R"\; filenames="HDA00002636509800042.png"\; state="\{\{state\}\}">R</figure-callout>}}^3$ and $r_i^{\left(1\right)}\&Element;{\mathrm{<figure-callout\; id="3"\; label="R"\; filenames="HDA00002636509800042.png"\; state="\{\{state\}\}">R</figure-callout>}}^3$ Respectively represent the position and direction of the root joint point, that is, the amount of translation and the amount of rotation,

j=2,...,31 represent the direction of non-root joint points.

Contents of the invention

The purpose of the present invention is to: propose a new distance characteristic curve based on the essential characteristics of the reaction movement, perform two key frame extractions according to the magnitude of the characteristic curve, and focus on solving the problem of automatically extracting a certain number of key frames for a section of motion sequence Motion posture has a better visual generalization for this segment of motion, and at the same time reconstructs the original motion sequence, maintaining a low error rate.

Based on the method of curve simplification, the present invention provides a motion key frame extraction based on distance curve amplitude, comprising the following steps:

S1. Select a group of joint distances as distance features;

S2, using the PCA method to extract the first dimension principal component, using smoothing filter to remove the noise, and obtaining the characteristic curve;

S3. Obtain an initial key frame by extracting local extremum points on the characteristic curve;

S4. Between the adjacent initial key frames, calculate the characteristic curve amplitude difference, and then insert additional key frames evenly according to uniform sampling;

S5. Merge the over-dense additional keyframes and the initial keyframes to leave a final set of keyframes.

Wherein, in step S1, the selection of the joint distance is made in the logical semantic manner of the following table:

d1: degree of bending of the left leg d4: left arm bending degree d7: bow/raise/shake degree d2: degree of bending of the right leg d5: degree of bending of the right arm d8: swing degree of left arm d3: the distance between the two feet d6: degree of bending d9: swing degree of right arm

Then a motion frame is represented as: θ=(d1,d2,d3,d4,d5,d6,d7,d8,d9), where d1,...,d9 are the joint distances with logical semantics selected in step S1.

The PCA method described in step S2 is divided into the following five steps:

S21: Construct the average of the 9 distance feature samples shown in S1

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; T</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ji</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span>$

Where i=1...9 represents the 9 distance feature samples, and T is the total number of frames of this movement;

然后构建距离特征的差值矩阵D=[Δθ₁，....Δθ₉]；S22: Calculate the difference between the original value and the average value of these 9 distance features

Then construct the difference matrix D=[Δθ ₁ ,....Δθ ₉ ] of the distance feature;

S23: Calculate the covariance matrix C=DD ^T , where D ^T is the transpose of D;

S24: Calculate the eigenvalue λ of the covariance matrix, and the corresponding eigenvector L;

这样就重新构建了一组新的9维主成分

其值按贡献率从大到小的顺序排列；然后提取第一维主成分，也就是的最大贡献率

S25: Extract feature vector

In this way, a new set of 9-dimensional principal components is reconstructed

Its values are arranged in order of contribution rate from large to small; then extract the first dimension principal component, that is, the maximum contribution rate of

其中N_frame为运动的总帧数;This motion M can in turn be expressed as:

Among them, N _frame is the total number of frames of the motion;

In addition, Lowess smoothing filter is selected in step S2.

In a preferred manner, the implementation process of step S4 is:

S41: Calculate the magnitude difference of the characteristic curve between the adjacent initial key frames;

S42: Set a threshold; if the amplitude difference ≥ the threshold, insert one or more frames of the additional key frame.

In a preferred manner, the implementation method of step S5: In view of the situation that the additional key frames and the initial key frames are too dense, combine by using a method of limiting the number of frames between adjacent key frames.

Compared with the prior art, the present invention has the following advantages:

1. A new set of distance features is selected, and the extracted feature curve can better reflect the essential features of human motion.

2. Using two key frame extraction methods, the first time extracts the boundary key frame as the initial key frame according to the local extreme point of the characteristic curve, and the second time uses the uniform sampling algorithm to automatically interpolate and merge the frame according to the amplitude difference of the characteristic curve Over-dense frames have better automation without artificially specifying the number of sampling frames, and can better adapt to motion. Fewer key frames are extracted at gentle motion, and more key frames are extracted at severe motion. . It can not only generalize the original motion well, but also has a low error rate and compression rate.

Description of drawings

Figure 1 is a flow chart of the present invention.

Figure 2 shows the human skeleton model diagram in MATLAB.

Figure 3 Selected nine joint distance feature representations, θ=(d1,d2,d3,d4,d5,d6,d7,d8,d9).

Fig. 4 The characteristic curve and key frame distribution diagram of the "kicking ball" movement; here the circle represents the initial key frame, and the asterisk represents the final key frame.

Figure 5 shows the comparison results of different key frame extraction methods for the "kicking ball" movement; among them, (a) the method of the present invention; (b) the uniform sampling method, the ellipse represents oversampling and undersampling; (c) the simplified curve, the ellipse represents Oversampling and undersampling; (d) only quaternion distance method.

Figure 6 Comparison of key frame extraction for similar motion types (taking walking as an example); among them, (a) walking with big arm swing; (b) walking with small arm swing; (c) walking cheerfully; (d) crazy Go wild.

Fig. 7 is a comparison chart of compression rates of six different motion types (kicking ball, jumping, running-stop, walking, dancing, walking-jumping-walking) using the method of the present invention.

Fig. 8 Four different methods (the method of the present invention, the quaternion distance method, the curve simplification, and the uniform sampling method) compare the reconstruction errors of six sampling movements; (a) kicking error (extract 33 key frames) ; (b) jumping error (24 keyframes extracted); (c) run-stop error (11 keyframes extracted); (d) walking error (16 keyframes extracted); (e) dancing Error (extract 37 keyframes); (f) go-jump-walk error (extract 50 keyframes).

Detailed ways

The technical solution of the present invention is: the motion key frame extraction method based on the range of the distance curve is invented in this paper, which includes selecting a group of new joint distances as the distance feature, using the PCA method to extract the first dimension principal component, and using smoothing filter to remove noise to obtain the feature Curve, the initial key frame is obtained by extracting the local extremum point on the characteristic curve, and the amplitude difference of the characteristic curve is calculated between the initial adjacent key frames, and then the corresponding number of frames is evenly inserted according to the uniform sampling, and then merged and over-dense keyframes to get the final set of keyframes. Accompanying drawing 1 shows the algorithm flowchart of the present invention, and it specifically comprises the following technical links:

1. Select a new set of joint distances as distance features.

The present invention selects nine human body joint distances with logical semantics as distance features to reflect the essential features of this movement, θ=(d1, d2, d3, d4, d5, d6, d7, d8, d9), see Figure 3. Its logical semantics are shown in Table 1. This reduces the original 96-dimensional data to 9-dimensional.

Table 1. Logical semantics of the extracted nine distance features

d1: degree of bending of the left leg d4: left arm bending degree d7: bow/raise/shake degree d2: degree of bending of the right leg d5: degree of bending of the right arm d8: swing degree of left arm d3: the distance between the two feet d6: degree of bending d9: the degree of swing of the right arm

Among them, the distance between human body joints can be determined according to the needs of specific human body motion capture. The nine features selected by the present invention are the optimal choice for this embodiment, and are also the optimal choice for most overall human body motion capture. For the steps of local movements, such as arm The movement of the arm should be realized by selecting the joint points of the arm.

2. Use PCA method to analyze the nine distance features, further reduce the dimension, extract the first dimension principal component, use Lowess smoothing filter, set the corresponding parameters to denoise the characteristic curve, and obtain the distance characteristic curve.

PCA analysis can be divided into the following five steps:

Step 1: Construct the average of the 9 distance feature samples shown in S1

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; T</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ji</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span>$

Where i=1...9 represents the 9 distance feature samples, and T is the total number of frames of this movement;

然后构建距离特征的差值矩阵D=[Δθ₁，....Δfθ₉]；Step 2: Calculate the difference between the original value and the mean value of these 9 distance features

Then construct the difference matrix D=[Δθ ₁ ,....Δfθ ₉ ] of the distance feature;

Step 3: Calculate the covariance matrix C=DD ^T , where D ^T is the transpose of D;

Step 4: Calculate the eigenvalue λ of the covariance matrix, and the corresponding eigenvector L;

这样就重新构建了一组新的9维主成分

其值按贡献率从大到小的顺序排列；然后提取第一维主成分，也就是的最大贡献率

此一个运动M进而可以表示成为：

其中N_frame为运动的总帧数;Step 5: Extract feature vectors

In this way, a new set of 9-dimensional principal components is reconstructed

Its values are arranged in order of contribution rate from large to small; then extract the first dimension principal component, that is, the maximum contribution rate of

This motion M can in turn be expressed as:

Among them, N _frame is the total number of frames of the motion;

The Lowess smoothing filter is used, and the corresponding parameters are set to denoise the curve, and a characteristic curve is obtained, which can better reflect the original motion.

3. For a motion clip, key stores the initial keyframe. The local extremum point loc on the characteristic curve is extracted, which is regarded as the boundary pose of the motion, so as to obtain the initial key frame key. The specific steps are described as follows:

Step 1: We use the "findpeak" function in MATLAB to get the local maximum;

Step 2: Invert the data, continue to use the "findpeak" function to get the local minimum, and then invert the data back for further use in the future;

Step 3: Merge the local maximum and local minimum to obtain the local extremum point loc (according to the order of frame numbers from small to large). The obtained local extremum point is used as the initial key frame key.

These initial keyframes are just boundary poses, we don't take into account the intensity of the motion. For lower error requirements and better generalization of human motion, it is necessary to insert one or more frames between adjacent initial key frames. We believe that if there is a large difference in the magnitude of the characteristic curve between adjacent initial keyframes, the motion between the adjacent keyframes has a large change in posture, and if the difference is small, the corresponding motion is considered to be gentle. Changes are minor. The following splitting and merging algorithm of the magnitude of the distance characteristic curve is used to further extract the key frames.

4. For a motion segment, key stores the initial key frame, and Lkey stores the final key frame. Then, between the initial keyframe keys, based on the magnitude of the distance characteristic curve, the keyframes are split or merged to obtain the final keyframe Lkey. The specific description is as follows:

Step 1: According to the characteristic curve, calculate the magnitude difference vary of the characteristic curve between the adjacent initial keyframes key1 and key2;

Step 2: Set a threshold threshold. If this vary<threshold, it means that the pose of the initial adjacent keyframes does not change drastically, and frame interpolation is not required; otherwise, the changes between adjacent initial keyframes are drastic, and one or more frames need to be inserted. For example, if the change is vary=0.4, we don’t need to insert more frames; if the change is vary=10, we need to insert 10 frames between key1 and key2 in a uniform sampling method, adding to the initial key frame key. In this way, a set of keyframes can be obtained;

Step 3: But the adjacent keyframes in the obtained keyframe set may be too dense, so it is necessary to merge the too dense keyframes to obtain the final keyframe. We use a method of merging that limits the number of frames between adjacent keyframes. If the number of frames between adjacent keyframes is greater than a threshold, we save this keyframe to the final keyframe Lkey.

During the implementation of the present invention, the values of the specific parameters are determined according to the needs, that is, there are no specific limitations on the value ranges, thresholds, etc., and the values can be selected according to the situation.

In the following, a specific embodiment is specifically realized by the present invention. The embodiments of the present invention are implemented on the premise of the technical solutions of the present invention, and detailed implementation methods and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments. The examples are selected from the CMU database (that is, the motion capture database created by the Graphics Laboratory of Carnegie Mellon University in the United States).

The specific implementation steps are:

Step 1: Select a representative movement from the CMU database, the 'kicking' movement (the total frame length of this movement is 802 frames). Extract nine joint distances as distance features θ=(d1,d2,d3,d4,d5,d6,d7,d8,d9).

Step 2: Use the PCA method to analyze the nine-dimensional distance features for further dimensionality reduction. Extract its first-dimensional principal component, use Lowess smoothing filter, set corresponding parameters to denoise the characteristic curve, obtain the distance characteristic curve we need, and then extract key frames.

Step 3: Use the 'findpeak' function in MATLAB to find the local extremum point loc on the characteristic curve, which is considered to be the boundary pose of the motion, and get the initial key frame key. See attached drawing 4.

Step 4: Split and merge keyframes based on the magnitude of the distance characteristic curve, and Lkey stores the final keyframe. See attached drawing 4. We extracted the local extremum point as the initial keyframe key represented by a circle, and we obtained the final keyframe Lkey represented by an asterisk. Through the study of the "kicking" movement, we found that the movement is an irregular movement, the first half of the movement is relatively slow and the posture changes are small, and the middle and rear parts of the movement are violent and the posture changes are large. When the movement is slow, the corresponding characteristic curve has a small amplitude, and no additional keyframes need to be inserted. However, when the movement is severe, it is necessary to split new keyframes between adjacent initial keyframes based on the difference in characteristic curve amplitude. , because the initial keyframe does not reflect this change very well, and then merge the over-dense keyframes. Through this method, we can effectively obtain the final set of keyframes.

Step 5: Comparison of different keyframe extraction algorithms. We used different methods: the method of the present invention, uniform sampling, curve simplification, and only quaternion distance method, extracting the same number of keyframes from a 'kicking' motion (same compression ratio). The comparison results of key frame extraction by different methods are shown in Figure 5. We extract 33 keyframes, which generalize motion well and avoid oversampling and undersampling issues.

Step 6: Using the method of the present invention to compare key frames of similar motions. We choose four types of walking motions to extract key frames within a certain frame range (130 frames). See attached drawing 6. From Figure 6(a)-(d), the key frames extracted by the present invention can clearly distinguish similar motions.

Step 7: Using the method of the present invention to test six different types of motion sequences, including kicking ball, jumping, running-stop, walking, dancing, walking-jumping-walking. See attached drawing 7. From Table 2, it is found that the compression ratio obtained by the method of the present invention is within 10%.

Table 2. Comparison of compression ratios for the six motion types

type play football Jump run-stop Walk Dance go-jump-walk total 802 439 239 343 1033 1200 Keyframe 33 twenty four 11 16 37 50 Compression ratio (%) 4.1 5.5 4.6 4.6 3.5 4.3

Then, we calculate the absolute mean error with the following formula:

E=[∑(F(n)-F'(n)) ² ]/N

Here, F(n) is the original motion data, F'(n) is the corresponding reconstructed motion data, and N is the total number of frames of this motion multiplied by 96 degrees of freedom.

As can be seen from Table 3, we use six sampled motions to calculate the absolute average error after reconstruction, where the data in brackets represent the number of key frames extracted. Figure 8, shows the error comparison of the six sampled motions.

Table 3. Comparison of error rates of the four methods

We use four methods to get the same number of keyframes. Afterwards, we reconstruct the motion sequence by linear interpolation to get the post-reconstruction error. We can find that the error rate obtained by using the method of the present invention is lower than other methods under the same compression ratio. This is because it extracts important keyframes and thus achieves a lower error. From the table, it can be found that the method of the present invention has an obvious advantage in both irregular motion and regular motion.

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone familiar with the technical field within the technical scope disclosed in the present invention, according to the technical solution of the present invention Any equivalent replacement or change of the inventive concepts thereof shall fall within the protection scope of the present invention.

Claims (5)

1. the motion key-frame extraction based on the distance Curve amplitude is characterized in that, comprises the steps:

S1, one group of joint distance of selection are as distance feature;

S2, employing PCA method are extracted the first dimension major component, adopt smothing filtering to remove noise, obtain characteristic curve;

S3, the Local Extremum of passing through to extract on the characteristic curve obtain initial key frame;

S4, between adjacent described initial key frame, the characteristic curve amplitude difference of calculating, and then insert uniformly additional key frame according to uniform sampling;

S5, overstocked described additional key frame and the described initial key frame of merging stay final key frame set.

2. according to the described motion key-frame extraction of claim 1, it is characterized in that among the step S1 based on the distance Curve amplitude, the selection of described joint distance, the logical semantics mode of following table is made:

D1: the degree of crook of left leg D4: left arm degree of crook D7: the degree of bowing/come back/shake the head D2: the degree of crook of right leg D5: right arm degree of crook D8: the degree of waving of left arm D3: the distance between the bipod D6: the degree of bending over D9: the degree of waving of right arm

Then a motion frame is represented to become: θ=(d1, d2, d3, d4, d5, d6, d7, d8, d9), d1 wherein .., d9 are the joint distance of choosing in the S1 step with logical semantics.

3. according to claim 1 or 2 described motion key-frame extraction, it is characterized in that the method for PCA described in the step S2 is divided into following five steps based on the distance Curve amplitude:

S21: the mean value that makes up 9 distance feature samples shown in the S1

${\stackrel{\&OverBar;}{\mathrm{\&theta;}}}_i=\frac{1}{T}\underset{j=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&theta;}}_{\mathrm{ji}}(i=1..9,j=1...T)$

I=1 wherein ... 9 represent this 9 distance feature samples, and T is the total frame number of this motion;

S22: the difference of calculating the original value and the mean value of these 9 distance feature

Make up the matrix of differences D=[Δ θ of distance feature then ₁.... and Δ θ ₉];

S23: calculate covariance matrix C=DD ^T, D wherein ^TTransposition for D;

S24: calculate the eigenvalue of covariance matrix, and corresponding proper vector L;

S25: extract proper vector So just rebuild one group of 9 new dimension major component

Its value is by contribution rate series arrangement from big to small; Extract the first dimension major component then, maximum contribution rate just

This motion M and then can represent to become:

N wherein _FrameTotalframes for motion;

In addition, select the Lowess smothing filtering for use among the step S2.

4. according to the described motion key-frame extraction of claim 3, it is characterized in that the implementation procedure of step S4 is based on the distance Curve amplitude:

S41: the amplitude difference of calculating the characteristic curve between the adjacent described initial key frame;

S42 a: threshold value is set; If described amplitude difference 〉=described threshold value is then inserted the described additional key frame of a frame or multiframe.

5. according to the described motion key-frame extraction of claim 4, it is characterized in that the implementation method of step S5 based on the distance Curve amplitude:

At described additional key frame and the overstocked situation of described initial key frame, adopt the method for the frame number between the adjacent key frame of restriction to merge.

Images