CN106529441A - Fuzzy boundary fragmentation-based depth motion map human body action recognition method - Google Patents

Abstract

The invention discloses a human behavior recognition method based on fuzzy boundary slicing in depth motion graphs. The model training method includes the following steps: slice the video depth map sequence and determine the blurred boundary of the slice according to the blur parameter α; calculate the depth motion map DMM of the main view, left view and top view for each sub-sequence after fragmentation ; use interpolation to convert these depth motion maps to a fixed size and normalize them; concatenate the normalized depth motion maps DMM of each video sequence subsequence to obtain the feature vector of the video sequence; use probabilistic collaborative representation The classifier R‑ProCRC classifies the features and finally realizes human action recognition. The human body behavior recognition method disclosed by the invention effectively captures the change rule of the time domain feature, enhances the anti-interference ability of the behavior feature to the time domain difference, and can realize the robust recognition of the human body behavior.

Description

Human action recognition method based on fuzzy boundary slices in deep motion map

Technical field:

The invention belongs to the field of machine vision, and in particular relates to a human behavior recognition method based on a fuzzy boundary slice in a deep motion map.

Background technique:

Human behavior recognition technology is a technology that extracts behavioral features from video sequences of human behaviors and recognizes actions through these features.

In the field of machine vision and pattern recognition, human behavior recognition has now become a very active branch. This technology has many potential applications in the field of human-computer interaction, including video analysis, surveillance systems, intelligent robots, etc. In the past few years, research on human behavior recognition was mainly based on image frame sequences collected by color cameras[1][2]. This kind of data has inherent defects, and they are very sensitive to illumination, occlusion, and complex backgrounds, which affect recognition accuracy. Nowadays, depth cameras have received widespread attention. This sensor can provide 3D structural information of objects, including the shape and motion information of the human body. At present, there are two main branches of the feature extraction method of depth: the method based on the estimated 3D skeleton point [3] and the method based on the original depth image. However, this kind of joint point estimation can sometimes be wrong, especially for cases where the human body is partially occluded or the background is complex. C. Chen et al. [4] obtained relatively good results by calculating the depth motion map (DMM, DepthMotion Maps) of the original depth image, and using pixels as features after PCA dimensionality reduction. However, this method accumulates all video frames on one DMM and loses the time information of the action.

In view of the above-mentioned problems of human behavior recognition, the present invention proposes a human behavior recognition method based on the depth action map of the fuzzy boundary slice, which not only absorbs the advantages of the original depth image information relative to the skeleton information, but also uses the fuzzy boundary slice The technology slices the DMM, extracts the time information of the action, and has high robustness to factors such as illumination, occlusion, and complex background.

Invention content:

The main purpose of the present invention is to propose a human behavior recognition method based on fuzzy boundary slices in depth action graphs, capture the time information of actions and use the robust probabilistic collaborative representation classifier R-ProCRC (Robust Probabilistic Collaborative Representation based Classifier) [5 ] to classify and improve the recognition accuracy.

In order to achieve the above object, the present invention provides the following technical solutions, including a training phase and a testing phase.

The technical scheme of the training phase of the human action recognition method based on the depth action map of the fuzzy boundary fragmentation is as follows:

Step 1: Given a training set of depth map samples of human behavior video sequences where X ^(k) represents the depth map sequence of the kth training sample, Be the original depth image of the i-th frame in the k-th sample, N _k is the total number of frames of the k-th sample; Y ^(k) represents the behavior category to which the k-th training sample belongs; M represents the number of samples in the training set;

Step 2: Divide each video sequence training sample X ^(k) in the training set directly into DIV equal time slices according to the time axis, and the length of each time slice is The divided time slice is expressed as in

Step 3: Select the appropriate fuzzy parameter α, and perform slice fuzzy processing on each time slice; the fuzzy time slice is expressed as To avoid subscript overflow, forward blurring is not performed for the first time slice, and backward blurring is not performed for the last time slice;

Step 4: Calculate each fuzzy time slice Depth motion maps in three different projection directions where v ∈ {f, s, t} respectively represent the three views (three directions) of the video sequence projection, the main view, the left view and the top view; so far, all training samples are calculated Corresponding set of deep motion maps

Step 5: Use the bicubic difference method to solve the depth action map obtained in step 4 Adjust to the same size, and normalize these depth action maps between 0 and 1, the normalized recorded as

Step 6: Correspond any training sample X ^(k) to the set of depth motion maps Carry out vectorization, and concatenate all vectorized sequence action graphs to complete the feature construction corresponding to the training sample X ^(k) , which is recorded as H ^(k) , and the feature set of all samples is expressed as {H ^(k) } _k∈[1M] ;

Step 7: All the samples obtained in Step 6 The output features of {H ^(k) } _k∈[1M] are dimensionally reduced by PCA, and all features after dimensionality reduction are saved

The technical scheme of the testing phase of the human action recognition method based on the depth action map of the fuzzy boundary fragmentation is as follows:

Step 1: Given a test sample (TestX, TestY) of a depth map sample of a human behavior video sequence, where TestX represents the depth map sequence of the test sample, X _i is the original depth image of the i-th frame in the test sample, _NT is the total number of frames of the test sample; TestY indicates the behavior category to which the test sample belongs;

Step 2: Divide the test sample TestX directly into DIV equal time slices according to the time axis (the segmentation method is the same as the training stage), and the length of each time slice is The divided time slice is expressed as in

Step 3: According to the fuzzy parameter α used in the training stage, perform slice fuzzy processing on each time slice; the fuzzy time slice is expressed as To avoid subscript overflow, forward blurring is not performed for the first time slice, and backward blurring is not performed for the last time slice;

Step 4: Calculate each fuzzy time slice Depth motion map DMM _j,v in three different projection directions, where v∈{f,s,t} respectively represent three views (three directions) of video sequence projection, front view, left view and top view; so far, calculate Obtain the depth motion map set {DMM _j,v } _{j∈[1DIV],v∈[f,s,t]} corresponding to the test sample TestX;

Step 5: Use the bicubic difference method to adjust the test sample depth action map {DMM _j,v } _{j∈[1DIV],v∈[f,s,t]} obtained in step 4 to the same size as the training stage, and according to The normalization method in the training phase normalizes these depth action maps to be between 0 and 1, and the normalized DMM _j,v is recorded as

Step 6: Collect the depth action images corresponding to the test sample TestX Carry out vectorization and concatenate all vectorized sequence action graphs to complete the feature construction corresponding to the test sample TestX, which is denoted as H _T ;

Step 7: Reduce the dimensionality of the output feature H _T of the test sample TestX obtained in Step 6 through PCA to obtain the dimensionality-reduced feature The PCA dimensionality reduction method is consistent with the training phase;

Step 8: Then the output features after dimensionality reduction Send it to the R-ProCRC[5] classifier to obtain the classification output PridY;

Step 9: Compare PridY and TestY, if PridY=TestY, the recognition is correct, otherwise the recognition is wrong.

Compared with the prior art, the present invention has the following beneficial effects:

1. This method uses depth data for human body behavior recognition. Compared with traditional color video data, depth data can achieve efficient segmentation of human body, while preserving the shape and structural characteristics of human body, which helps to improve classification accuracy;

2. The traditional feature extraction method using the depth motion map DMM projects the entire video frame onto a DMM map, which loses time information. The human behavior recognition method of depth action map based on fuzzy boundary segmentation proposed by this method, slices the depth image sequence in the time dimension, and effectively captures the change law of time domain features;

3. Aiming at the time-domain differences of human behaviors, the human behavior recognition method of the depth motion map of the fuzzy boundary slices proposed by this method uses the fuzzy parameter α to control the boundaries between slices, so that the information of adjacent slices can be shared. The robustness of the feature to the time domain difference of human behavior is further improved, and the recognition accuracy of the R-ProCRC classifier is significantly improved.

Description of drawings:

Fig. 1 is a flow chart of a method for constructing human behavior recognition features based on a depth action map of fuzzy boundary slices;

Figure 2 Schematic diagram of fuzzy boundary fragmentation strategy;

Figure 3 is an example of the projection of the depth action image under three viewing angles;

detailed description

In order to better illustrate the purpose, specific steps and characteristics of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings, taking the MSRAction3D data set as an example:

The human behavior recognition method of the depth motion map of the fuzzy boundary segmentation proposed by the present invention, wherein the flow chart of feature extraction is shown in FIG. 1 . Firstly, the sample is divided into equal segments to determine the boundary, and then the blurring degree of the boundary is determined according to the parameter α. For each segmented video sub-sequence, its depth motion map DMM is calculated, and all The DMM of the sample is fixed to the same size and normalized, and the features of the subsequence are obtained after serial vectorization, and the construction of the output feature of the training sample is completed.

A human behavior recognition method based on a deep motion graph of fuzzy boundary slices proposed by the present invention includes a training phase and a testing phase.

The technical scheme of the training phase of the human action recognition method based on the depth action map of the fuzzy boundary fragmentation is as follows:

Step 1: Given a training set of depth map samples of human behavior video sequences where X ^(k) represents the depth map sequence of the kth training sample, Be the original depth image of the i-th frame in the k-th sample, N _k is the total number of frames of the k-th sample; Y ^(k) represents the behavior category to which the k-th training sample belongs; M represents the number of samples in the training set;

Step 2: Divide each video sequence training sample X ^(k) in the training set directly into DIV equal time slices according to the time axis, and the length of each time slice is The divided time slice is expressed as in

Step 3: Select the appropriate fuzzy parameter α, and perform slice fuzzy processing on each time slice; the fuzzy time slice is expressed as To avoid subscript overflow, forward blurring is not performed for the first time slice, and backward blurring is not performed for the last time slice;

Step 4: Calculate each fuzzy time slice Depth motion maps in three different projection directions where v ∈ {f, s, t} respectively represent the three views (three directions) of the video sequence projection, the main view, the left view and the top view; so far, all training samples are calculated Corresponding set of deep motion maps

Step 5: Use the bicubic difference method to solve the depth action map obtained in step 4 Adjust to the same size, and normalize these depth action maps between 0 and 1, the normalized recorded as

Step 6: Correspond any training sample X ^(k) to the set of depth motion maps Carry out vectorization, and concatenate all vectorized sequence action graphs to complete the feature construction corresponding to the training sample X ^(k) , which is recorded as H ^(k) , and the feature set of all samples is expressed as {H ^(k) } _k∈[1M] ;

Step 7: All the samples obtained in Step 6 The output features of {H ^(k) } _k∈[1M] are dimensionally reduced by PCA, and all features after dimensionality reduction are saved

The technical scheme of the testing phase of the human action recognition method based on the depth action map of the fuzzy boundary fragmentation is as follows:

Step 1: Given a test sample (TestX, TestY) of a depth map sample of a human behavior video sequence, where TestX represents the depth map sequence of the test sample, X _i is the original depth image of the i-th frame in the test sample, _NT is the total number of frames of the test sample; TestY indicates the behavior category to which the test sample belongs;

Step 2: Divide the test sample TestX directly into DIV equal time slices according to the time axis (the segmentation method is the same as the training stage), and the length of each time slice is The divided time slice is expressed as in

Step 3: According to the fuzzy parameter α used in the training stage, perform slice fuzzy processing on each time slice; the fuzzy time slice is expressed as To avoid subscript overflow, forward blurring is not performed for the first time slice, and backward blurring is not performed for the last time slice;

Step 4: Calculate each fuzzy time slice Depth motion map DMM _j,v in three different projection directions, where v∈{f,s,t} respectively represent three views (three directions) of video sequence projection, front view, left view and top view; so far, calculate Obtain the depth motion map set {DMM _j,v } _{j∈[1DIV],v∈[f,s,t]} corresponding to the test sample TestX;

Step 5: Use the bicubic difference method to adjust the test sample depth action map {DMM _j,v } _{j∈[1DIV],v∈[f,s,t]} obtained in step 4 to the same size as the training stage, and according to The normalization method in the training phase normalizes these depth action maps to be between 0 and 1, and the normalized DMM _j,v is recorded as

Step 6: Collect the depth action images corresponding to the test sample TestX Carry out vectorization and concatenate all vectorized sequence action graphs to complete the feature construction corresponding to the test sample TestX, which is denoted as H _T ;

Step 7: Reduce the dimensionality of the output feature H _T of the test sample TestX obtained in Step 6 through PCA to obtain the dimensionality-reduced feature The PCA dimensionality reduction method is consistent with the training phase;

Step 8: Then the output features after dimensionality reduction Send it to the R-ProCRC[5] classifier to obtain the classification output PridY;

Step 9: Compare PridY and TestY, if PridY=TestY, the recognition is correct, otherwise the recognition is wrong.

In the above-mentioned technical scheme, in the method of dividing the video sequence into equal length slices in step 2 of the training phase, the selection of the number of slices DIV is based on the specific human behavior sample data set to determine the optimal number of slices, taking the MSR Action3D data set as For example, DIV=3.

In the above-mentioned technical scheme, in the method for dividing the video sequence into equal time slices in step 2 of the training phase, the length of each time slice is The length of the time slice is rounded down; if the length of the last time slice is insufficient Then choose according to the actual length;

In the above technical solution, in step 3 of the training stage, in the segmental fuzzy processing of time slices, the selection of the fuzzy parameter α is based on the specific human behavior sample data set to determine the optimal parameter. Taking the MSR Action3D data set as an example, α=0.8.

In the above technical solution, in step three of the training phase, in the segmental fuzzy processing of time slices, no forward fuzzy processing is performed on the first time slice of each sample; no backward fuzzy processing is performed on the last time slice of each sample deal with.

In the above technical solution, step 2 and step 3 of the training stage jointly complete the fuzzy slice processing of the video sequence, as shown in Figure 2, the fuzzy parameter α controls the boundary between slices, so that adjacent slice information is shared, The robustness of features to temporal differences in human behavior is further improved.

In the above technical scheme, step four of the training phase is for each fuzzy time slice Depth motion maps in three different projection directions The calculation of uses the video frame absolute difference superposition method, specifically:

The video frames in the same time slice are projected in three visual directions to obtain the main view, left view and top view of each video, and then the projections of the same viewing angle of adjacent frames are subtracted, and the absolute values are superimposed after subtraction. The trajectory of is saved, the specific formula is as follows:

in with respectively represent the first time slice and the last time slice in the k-th sample, and v∈{f,s,t} respectively represent the three views (three directions) of the video sequence projection, that is, the front view, the left view and the top view ; Indicates the v-direction projection of the depth map of the i-th frame in the k-th sample; j∈(2,...,DIV-1), α∈(0,1); as shown in Figure 3, the DMM map of each projection direction is effectively saved Trajectory information of a single fuzzy slice of human behavior sequences in three projection directions is obtained.

In the above technical scheme, step 5 of the training stage adopts the bicubic difference method to convert the depth action map to Adjusted to the same size, the dimensions of the front view, left view, and top view used in this patent are defined as: 50×25, 50×40, and 40×20; in actual implementation, different interpolation methods can be selected to achieve the size of the depth action map The scaling of , the selection premise is to minimize the loss of image information.

In the above technical solution, all samples calculated in step 7 of the training phase against step 6 The output features of {H ^(k) } _k∈[1M] are dimensionally reduced by PCA, and the specific feature dimension after dimensionality reduction can be determined according to the number of training samples. In the implementation example of this patent, if the number of training samples is M, the final feature dimension is (M-20)×1.

In the above technical solution, the feature construction methods and parameters used in steps 2 to 6 of the test phase are the same as those used in the training phase.

In the above-mentioned technical scheme, step 7 of the test stage uses PCA to reduce the dimensionality of the output feature H _T of the test sample TestX calculated in step 6, and the dimension after dimensionality reduction is (M-20)×1, and M is the number of samples in the training stage number.

In the above technical solution, step 8 of the test phase uses R-ProCRC classification to classify the dimensionality-reduced behavioral features obtained in step 7 The specific methods for classification are:

(1) Calculate the optimal parameters

in is the training feature after dimensionality reduction, H _T is the test sample feature, Is the set of all input feature vectors belonging to category c (c∈C), ‖C‖ is the total number of categories, λ and γ are parameters between 0 and 1; in The construction method is as follows: firstly, the initialized as a dictionary with 0 matrix of the same size, then the Assigned to , the position is exist The relative position in can be obtained E.g The value is: is a diagonalized weight matrix:

here, Represents all elements of the i-th row, Indicates the i-th value of the test sample feature vector;

(2) Estimate the output characteristics of the test sample Probability of belonging to class c

Test sample output features The probability of belonging to category c is estimated as follows:

for all samples are the same, so the above formula can be simplified as:

From this we can get the characteristics category to which it belongs.

In order to verify the validity of the present invention, the present invention has carried out experiments successively on the famous human body behavior depth information database MSRAction3D and Action Pair. Table 1 shows the characteristics of the two human behavior depth information databases.

Table 1: Depth Database Feature Description

As shown in Table 2, we divide the MSR Action3D database into three fixed subsets in the experiments. Each subset uses 3 experimental methods, Test One uses the first demonstration behavior of each subject as the training set, and the rest as the test set. Test Two uses the first two demonstration behaviors of each subject as the training set, and the rest as the test set. Cross Test uses all video sequences of 1, 3, 5, 7, and 9 subjects as the training set, and the rest are used for testing. The experimental results are shown in Table 3. It can be seen that the behavior recognition accuracy of the present invention is better than the traditional DMM method in most cases.

Table 2: MSR Action3D database subset

Table 3: Comparison of MSR Action3D database subsets

Table 4 shows the recognition rate of the present invention on the Action Pair database and the comparison with DMM. Since the actions in the Action Pair database have a large number of actions with opposite timing, such as "pick up" and "put down", "stand up" and "sit down", they are very sensitive to time information. The recognition rate of traditional DMM is only 50.6%, but the recognition rate of the present invention reaches 97.2%.

Table 4: Action Pair recognition rate of different algorithms

Since the present invention uses depth data for human body behavior recognition, compared with traditional color video data, depth data can quickly and accurately complete human body separation, preserve the shape and structural features of the human body, and help improve accuracy. The processing method is more robust than the bone points estimated based on bone tracking technology; the traditional feature extraction method using the depth motion map DMM projects the entire video frame onto a DMM map, which loses time information; The human behavior recognition of the depth action map of the fuzzy boundary segmentation proposed by this method divides the existing DMM into slices, and uses the fuzzy parameter α to control the boundaries between slices, so that the information of adjacent slices can be shared, and the It enables DMM to better capture time information, and the R-ProCRC [5] classifier can be used together to obtain excellent recognition accuracy.

The specific implementation of the present invention has been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned implementation, and within the knowledge of those of ordinary skill in the art, it can also be done without departing from the gist of the present invention. Various changes.

references

[1]. Bian W, Tao D, Rui Y. Cross-domain human action recognition. [J]. IEEE Transactions on Systems Man&Cybernetics Part B Cybernetics A Publication of the IEEE Systems Man&Cybernetics Society, 2012,42(2):298-307.

[2].Niebles J C,Wang H,Li F F.Unsupervised Learning of Human Action Categories Using Spatial-Temporal Words[J].International Journal of Computer Vision,2008,79(3):299-318.

[3].Wang J, Liu Z, Wu Y, et al.Mining actionlet ensemble for actionrecognition with depth cameras[C]//IEEE Conference on Computer Vision and Pattern Recognition.IEEE Computer Society,2012:1290-1297.

[4]. Chen C, Liu K, Kehtarnavaz N. Real-time human action recognition based on depth motion maps [J]. Journal of Real-Time Image Processing, 2013: 1-9.

[5]. Sijia C, Lei Z, et al. A Probabilistic Collaborative Representation based Approach for Pattern Classification. IEEE Trans. on Pattern Analysis and Machine Intelligence, 2016.

[6]. Xu H, Chen E, Liang C, et al.Spatio-Temporal Pyramid Model based ondepth maps for action recognition[C]//Mmsp2015 IEEE, International Workshop on Multimedia Signal Processing.2015.

Claims (7)

1. The human action recognition based on the depth action graph of fuzzy boundary slice, is characterized in that, comprises training stage and testing stage. 2. the human body behavior recognition based on the depth action map of fuzzy boundary slice according to claim 1, is characterized in that, training stage comprises the following steps: Step 1. Given the training set of depth map samples of human behavior video sequences where X ^(k) represents the depth map sequence of the kth training sample, Be the original depth image of the i-th frame in the k-th sample, N _k is the total number of frames of the k-th sample; Y ^(k) represents the behavior category to which the k-th training sample belongs; M represents the number of samples in the training set; Step 2, each video sequence training sample X ^(k) in the training set is directly divided into DIV equal time slices according to the time axis, and the length of each time slice is The divided time slice is expressed as in Step 3: Select the appropriate fuzzy parameter α, and perform slice fuzzy processing on each time slice; the fuzzy time slice is expressed as To avoid subscript overflow, forward blurring is not performed for the first time slice, and backward blurring is not performed for the last time slice; Step 4. Calculate each fuzzy time slice Depth motion maps in three different projection directions where v ∈ {f, s, t} respectively represent the three views (three directions) of the video sequence projection, the main view, the left view and the top view; so far, all training samples are calculated Corresponding set of deep motion maps Step 5. Use the bicubic difference method to solve the depth action map obtained in step 4 Adjust to the same size, and normalize these depth action maps between 0 and 1, the normalized recorded as Step 6. Correspond any training sample X ^(k) to the set of depth motion maps Carry out vectorization, and concatenate all vectorized sequence action graphs to complete the feature construction corresponding to the training sample X ^(k) , which is recorded as H ^(k) , and the feature set of all samples is expressed as {H ^(k) } _{k∈[ 1M]} ; Step 7, all the samples obtained in step 6 The output features of {H ^(k) } _{k∈[1 M]} are dimensionally reduced by PCA, and all features after dimensionality reduction are saved 3. the human body behavior recognition based on the depth action map of fuzzy boundary slice according to claim 1, is characterized in that, testing stage comprises the following steps: Step 1, given a test sample (TestX, TestY) of the depth map sample of a human behavior video sequence, wherein TestX represents the depth map sequence of the test sample, X _i is the original depth image of the i-th frame in the test sample, _NT is the total number of frames of the test sample; TestY indicates the behavior category to which the test sample belongs; Step 2. Divide the test sample TestX directly into DIV equal time slices according to the time axis (the slice method is the same as the training stage), and the length of each time slice is The divided time slice is expressed as in Step 3. According to the fuzzy parameter α used in the training stage, perform fragmentation fuzzy processing on each time slice; the fuzzy time slice is denoted as To avoid subscript overflow, no forward blurring is done for the first time slice, and no backward blurring is done for the last time slice; Step 4. Calculate each fuzzy time slice Depth motion map DMM _j,v in three different projection directions, where v∈{f,s,t} respectively represent three views (three directions) of video sequence projection, front view, left view and top view; so far, calculate Obtain the depth motion map set {DMM _{j, v} } _{j∈[1 DIV], v∈[f, s, t]} corresponding to the test sample TestX; Step 5. Use the bicubic difference method to adjust the test sample depth motion map {DMM _j,v } _{j∈[1 DIV],v∈[f,s,t]} obtained in step 4 to the same size as the training stage, and According to the normalization method of the training stage, these depth action maps are normalized between 0 and 1, and the normalized DMM _j,v is recorded as Step 6: Collect the depth action images corresponding to the test sample TestX Carry out vectorization and concatenate all vectorized sequence action graphs to complete the feature construction corresponding to the test sample TestX, which is denoted as H _T ; Step 7. The output feature H _T of the test sample TestX obtained in step 6 is reduced by PCA to obtain the feature after dimension reduction The PCA dimensionality reduction method is consistent with the training phase; Step 8, and then the output features after dimensionality reduction Send it to the R-ProCRC classifier to obtain the classification output PridY; Step 9: Comparing PridY and TestY, if PridY=TestY, the recognition is correct, otherwise the recognition is wrong. 4. in the method that step 2 of training stage according to claim 2 is carried out to video sequence in the method for equal time slice division, each time slice length is The length of the time slice is rounded down; if the length of the last time slice is insufficient Then choose according to the actual length. 5. in the fragmentation fuzzy processing of time slice according to training stage step 3 according to claim 2, the first time slice of each sample is not done forward fuzzy processing; The last time slice of each sample is not Do backward blurring. 6. training phase step four according to claim 2 to each fuzzy time slice Depth motion maps in three different projection directions The method of superimposing the absolute difference of video frames is used for the calculation of , specifically: projection in three visual directions on the video frames in the same time slice to obtain the main view, left view and top view of each video, and then the same viewing angle of adjacent frames The projection is subtracted, and the absolute value is superimposed after the subtraction, so that the trajectory of the action is saved. The specific formula is as follows: in with respectively represent the first time slice and the last time slice in the k-th sample, and v∈{f,s,t} respectively represent the three views (three directions) of the video sequence projection, that is, the front view, the left view and the top view ; Represents the v-direction projection of the i-th frame depth map in the k-th sample; j∈(2,...,DIV-1), α∈(0,1); the DMM map of each projection direction effectively preserves a single human behavior sequence Trajectory information of fuzzy slices in three projection directions. 7. according to claim 3, test phase step eight uses R-ProCRC classification to the behavioral characteristics after the dimensionality reduction that test step seven obtains The specific methods for classification are: (1) Calculate the optimal parameters in is the training feature after dimensionality reduction, H _T is the test sample feature, Is the set of all input feature vectors belonging to category c (c∈C), ‖C‖ is the total number of categories, λ and γ are parameters between 0 and 1; in The construction method is as follows: firstly, the initialized as a dictionary with 0 matrix of the same size, then the Assigned to , the position is exist The relative position in can be obtained E.g The value is: is a diagonalized weight matrix: here, Represents all elements of the i-th row, Indicates the i-th value of the test sample feature vector; (2) Estimate the output characteristics of the test sample Probability of belonging to class c Test sample output features The probability of belonging to category c is estimated as follows: for all samples are the same, so the above formula can be simplified as: From this we can get the characteristics category to which it belongs.