Multi-person behavior recognition method and device, computer device, and storage medium

Abstract

Embodiments of the present application disclose a multi-person behavior recognition method and device, computer equipment and a storage medium, wherein the method comprises: acquiring image data in a detection area; performing frame cutting processing on the image data to obtain a plurality of static pictures; inputting the plurality of static pictures in the form of a picture sequence into a human behavior prediction model for processing to obtain a human behavior recognition result. The present application combines multi-target tracking, key point detection and optical flow estimation method, can effectively perform real-time behavior recognition on a moving human body of a multi-target, has good compatibility for behavior detection of a human body moving greatly in a short time period, improves the robustness and accuracy of the model in different scene recognition, and makes the model have good understanding ability for complex content images.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically to a method, apparatus, computer device, and storage medium for multi-person behavior recognition. Background Technology

[0002] With rapid societal development, human behavior prediction is currently a research hotspot and challenge in both industry and academia, and it has significant practical applications. Several methods are currently available for human behavior prediction:

[0003] The first method involves processing the acquired visible light and infrared images to obtain the target area for tracking. Next, it detects whether the target area includes pedestrians. If pedestrians are included, they are tracked, and during tracking, pedestrian edges are detected to obtain the pedestrian identification area from the target area. This pedestrian identification area is then input into a recognition model to obtain the pedestrian behavior recognition result. This patent uses infrared tracking, which requires additional equipment costs. Furthermore, it relies on traditional histogram-based feature extraction to determine pedestrian targets, which is severely affected by lighting conditions and has poor performance. Additionally, this method only recognizes the outline of the person, ignoring spatial-temporal fusion information, resulting in poor recognition performance.

[0004] The second approach involves training a 3D convolutional neural network using multi-task deep learning. This method uses a fixed number of consecutive frames from a background video as input to various human behavioral attributes, and the network is trained to perform the recognition task. However, this method only recognizes human behavior at fixed locations and is limited to a single individual, which has significant limitations. Furthermore, behavior recognition based on 3D convolution is inefficient, and the trained model is heavily influenced by background factors, resulting in poor recognition performance.

[0005] The third method involves acquiring video clips, extracting features and reducing dimensionality from the image frame sequences within those clips, encoding the reduced feature vectors, and then expanding these expanded feature vectors to obtain a predetermined number of expanded feature vectors. These expanded feature vectors and the encoded feature vectors are then input into a three-layer single-layer decoder for decoding. The final decoded feature vector is then input into a single-layer fully connected feedforward network for calculation, yielding multiple predicted values. These predicted values ​​are then input into a logistic regression network to obtain the corresponding predicted probabilities. The category corresponding to the highest probability value is selected as the human action category of the bounding box corresponding to the final decoded feature vector. This method first extracts spatial information and then extracts temporal features using spatial features from different time periods. However, this process easily loses spatial information during temporal feature extraction, making it difficult to achieve a fusion of spatial and temporal features, thus reducing the accuracy of action recognition. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device, computer equipment and storage medium for multi-person behavior recognition.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The first aspect is multi-person behavior recognition methods, including:

[0009] Acquire image data within the detection area;

[0010] The image data is sliced ​​into frames to obtain multiple static images;

[0011] Multiple static images are input as an image sequence into the human behavior prediction model for processing to obtain human behavior recognition results.

[0012] The further technical solution is as follows: The step of inputting multiple static images as an image sequence into a human behavior prediction model for processing to obtain human behavior recognition results includes:

[0013] Multiple static images are input into the tracking model as an image sequence for processing to obtain human images with ID indexes;

[0014] Different images of people with ID indexes are input into a keypoint detection model for processing to obtain a keypoint heatmap;

[0015] The image of a person with an ID index is input into the optical flow model for processing to obtain an optical flow feature map;

[0016] The key point heatmap and optical flow feature map are input into the video classification model for processing to obtain the human behavior recognition results.

[0017] The further technical solution is as follows: The step of inputting different person images with ID indexes into a keypoint detection model for processing to obtain a keypoint heatmap includes:

[0018] Different images of people with ID indices are input into multiple concatenated ExtrackBlocks for processing to obtain the features of the people;

[0019] The character's features are input into RefineBlock for processing to obtain a key point heatmap.

[0020] The further technical solution is as follows: The process of inputting different person images with ID indexes into multiple concatenated ExtrackBlocks for processing to obtain person features includes:

[0021] Different images of people with ID indices are processed by depthwise separable convolution and spatial separable convolution respectively to obtain depthwise convolution features and spatial convolution features.

[0022] The depthwise convolutional features and the spatial convolutional features are added together to obtain the summed features.

[0023] The summed features are processed by the first convolution, the second convolution, the third convolution, and the fourth convolution to obtain the first convolution feature, the second convolution feature, the third convolution feature, and the fourth convolution feature.

[0024] The first, second, third, and fourth convolutional features are respectively input into four attention mechanism branches for processing to obtain the first attention branch features, the second attention branch features, the third attention branch features, and the fourth attention branch features, respectively.

[0025] The features from the first attention branch, the second attention branch, the third attention branch, and the fourth attention branch are concatted to obtain combined features.

[0026] The combined features and spatial convolution features are merged to obtain the features of the person.

[0027] The further technical solution is as follows: The input of character features into RefineBlock for processing to obtain a key point heatmap includes:

[0028] The features of the person are processed by deformable convolution to obtain deformable convolution features;

[0029] Deformable convolutional features are input into SCSE for processing to obtain attention weight features;

[0030] Deformable convolutional features are input into a hybrid deep convolutional network and connected to a sigmoid layer for processing to obtain attention feature maps;

[0031] The result of multiplying the attention weight features with the attention feature map is added to the deformable convolution features to obtain the output features.

[0032] The output features are enlarged to their original size using bilinear interpolation to obtain a keypoint heatmap.

[0033] The further technical solution is as follows: The step of inputting the image of a person with an ID index into the optical flow model for processing to obtain an optical flow feature map includes:

[0034] The two frames of character images with ID indices are input into residual convolution for processing to obtain the first residual convolution feature and the second residual convolution feature.

[0035] The first residual convolutional feature and the second residual convolutional feature are respectively input into the depthwise separable convolutional network and the spatially separable convolutional network for processing to obtain the first depthwise separable convolutional feature, the first spatially separable convolutional feature, the second depthwise separable convolutional feature, and the second spatially separable convolutional feature.

[0036] The first depthwise separable convolutional feature and the first spatially separable convolutional feature are concatted to obtain the first combined feature;

[0037] The second depthwise separable convolutional feature and the second spatially separable convolutional feature are concatted to obtain the second combined feature;

[0038] The result of multiplying the first and second combined features is then processed by the sigmoid activation function to obtain the activation function features.

[0039] The first residual convolutional feature is convolved to obtain the convolutional processed feature;

[0040] The convolutional features and the activation function features are concatenated to obtain merged features;

[0041] The merged features are input into the convolutional block for processing to obtain the convolutional block features;

[0042] The activation function features are subjected to dilated convolution upsampling to obtain dilated convolution upsampling features;

[0043] The merged features are subjected to deconvolution upsampling to obtain deconvolution upsampled features;

[0044] The convolutional block features are subjected to unpooling upsampling to obtain unpooling upsampling features;

[0045] The values ​​of the dilated convolution upsampling feature, the deconvolution upsampling feature, and the unpooling upsampling feature are added together to obtain the first summed feature.

[0046] The two frames of character images with ID indices are concatted and then input into transformerEncoder and deformable convolution respectively to obtain the first and second processed features.

[0047] The result of resizing the first processed feature is concatted with the second processed feature to obtain the combined feature;

[0048] The combined features are input into two convolutional blocks for processing to obtain the third and fourth processed features.

[0049] The third and fourth processing features are upsampled separately, and the results of the upsampling are added together to obtain the second summed feature.

[0050] The features of the first addition result and the features of the second addition result are concatted to obtain the intermediate result features;

[0051] The intermediate feature data are input into two convolutional layers for processing to obtain optical flow feature maps.

[0052] The further technical solution is as follows: The key point heatmap and optical flow feature map are input into a video classification model for processing to obtain human behavior recognition results, including:

[0053] Map the key points of the Gaussian heatmap on each channel of the key point heatmap to a single-channel heatmap to obtain a single-channel Gaussian heatmap.

[0054] The optical flow feature map was scaled proportionally to the single-channel Gaussian thermal image. Figure 1 The same size is obtained and concatted with it to obtain key point features containing optical flow information;

[0055] Keypoint features containing optical flow information are input into TIN for behavior classification to obtain human behavior recognition results.

[0056] Secondly, the multi-person behavior recognition device includes:

[0057] The acquisition unit is used to acquire image data within the detection area;

[0058] The first processing unit is used to perform frame-slicing processing on the image data to obtain multiple frames of static images;

[0059] The second processing unit is used to input multiple frames of static images into the human behavior prediction model in the form of an image sequence for processing, so as to obtain human behavior recognition results.

[0060] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the multi-person behavior recognition method as described above.

[0061] Fourthly, a computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the steps of the multi-person behavior recognition method described above.

[0062] The advantages of this invention compared to existing technologies are: this invention combines multi-target tracking, key point detection, and optical flow estimation methods, which can effectively perform real-time behavior recognition of multi-target moving human bodies and has good compatibility with the behavior detection of human bodies that move significantly within a short period of time.

[0063] By integrating optical flow features and key point heatmap features, more accurate action recognition and analysis can be performed, solving the problem of poor training results caused by excessive reliance on background for behavior, and improving the robustness and accuracy of the model in different scenarios.

[0064] In addition, spatial and temporal information were considered. For spatial information, supervision information of the original pixel and key point positions was added. For temporal information, long-period video sequence information and short-period optical flow information were considered at the same time, which enabled the model to have a better understanding of images with complex content.

[0065] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of the present invention more obvious and understandable, preferred embodiments are described in detail below. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 A schematic diagram illustrating an application scenario of the multi-person behavior recognition method provided in a specific embodiment of the present invention;

[0068] Figure 2 A flowchart of a multi-person behavior recognition method provided in a specific embodiment of the present invention;

[0069] Figure 3 A schematic block diagram of a multi-person behavior recognition device provided in a specific embodiment of the present invention;

[0070] Figure 4 This is a schematic block diagram of a computer device provided for a specific embodiment of the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0073] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0074] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0075] Please see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram illustrating an application scenario of the multi-person behavior recognition method provided in this embodiment of the invention. Figure 2 The flowchart illustrates a multi-person behavior recognition method provided in a specific embodiment of the present invention. This multi-person behavior recognition method is applied to a server and is executed by application software installed on the server.

[0076] like Figure 2 As shown, the multi-person behavior recognition method includes the following steps: S10-S30.

[0077] S10. Acquire image data within the detection area.

[0078] Use surveillance equipment, such as surveillance cameras, to monitor and capture video data (image data) of the area to be detected. Any commercially available surveillance equipment is acceptable; this application does not impose any restrictions on its use.

[0079] S20. Perform frame segmentation on the image data to obtain multiple static images.

[0080] The detected image data is sliced ​​into individual RGB static images, which are then used as input data for the human behavior prediction model.

[0081] S30. Input multiple static images as an image sequence into the human behavior prediction model for processing to obtain human behavior recognition results.

[0082] In one embodiment, step S30 specifically includes the following steps: S301-S304.

[0083] S301. Input multiple static images as an image sequence into the tracking model for processing to obtain a person image with an ID index.

[0084] In this embodiment, the tracking model used is BoT-SORT. BoT-SORT uses the YOLOv7 model as a human detector. In order to reduce the ID switching caused by human occlusion, it not only uses human features as the matching benchmark, but also adds optical flow features of the previous and next frames. The human features and optical flow features obtained from the detector are fused together to determine whether the index of the pedestrian in the previous and next frames matches, adding auxiliary information.

[0085] In this embodiment, step S301 specifically includes the following steps:

[0086] S3011. Input multiple static images as an image sequence into the tracking model to detect different human target boxes.

[0087] In this embodiment, the obtained multi-frame static images (frame video stream) are input into the tracking model as an image sequence, which can detect different human target boxes. These target boxes are identified by their image position coordinates. The coordinates are used to represent the position of the target bounding box relative to the top left corner of the image and the bottom right corner of the image.

[0088] S3012. Assign an index id to the different human target boxes detected to obtain human images with id indexes.

[0089] S302. Input the images of different people with id indexes into the key point detection model for processing to obtain the key point heatmap.

[0090] A Multi-Residual Network (MRN) was designed based on an attention mechanism and a convolutional neural network for keypoint detection. The keypoint detection model consists of two sub-network structures: ExtractBlock, a feature extraction model structure, and RefineBlock, which generates keypoint heatmaps.

[0091] In one embodiment, step S302 specifically includes the following steps: S3021-S3022.

[0092] S3021. Input different person images with id indexes into multiple concatenated ExtrackBlocks for processing to obtain person features.

[0093] In one embodiment, step S3021 specifically includes the following steps: S30211-S30216.

[0094] S30211. Perform depthwise separable convolution and spatial separable convolution on different images of people with id indices to obtain depthwise convolution features and spatial convolution features.

[0095] In this embodiment, for each ExtrackBlock structure, the input images of different people with id indices are processed by 3*3 depthwise separable convolution and spatially separable convolution to obtain features. as well as .

[0096] S30212. Add the depthwise convolutional features and the spatial convolutional features together to obtain the summed features.

[0097] In this embodiment, the features and The feature is obtained by performing addition calculation. .

[0098] S30213. The summed features are processed by the first convolution, the second convolution, the third convolution, and the fourth convolution to obtain the first convolution feature, the second convolution feature, the third convolution feature, and the fourth convolution feature.

[0099] In this embodiment, the features are obtained by performing 3*3 convolution, 5*5 convolution, 7*7 convolution, and 9*9 convolution, respectively. Features ,feature and characteristics .

[0100] S30214. Input the first convolutional feature, the second convolutional feature, the third convolutional feature and the fourth convolutional feature into the four attention mechanism branches for processing, so as to obtain the first attention branch feature, the second attention branch feature, the third attention branch feature and the fourth attention branch feature respectively.

[0101] In this embodiment, the features ,feature ,feature and characteristics The inputs are processed in four attention mechanism branches. For the first branch, the inputs are first processed... The input is fed into a ConvBlock, which consists of 1x1 convolutions, 3x3 convolutions, and 3x3 dilated convolutions using BacthNormalization, ultimately yielding the features. At the same time, features Features are obtained through CBAM (Convolution Block Attention Module) and 1*1 convolution processing. Next and Perform concat merging to obtain features .

[0102] The second branch uses the same structure as the first branch, and the resulting features The features are obtained by inputting them into the CBAM branch and the 1*1 convolution branch respectively. At the same time, Parallel input to ConvBlock yields features , will feature With features Perform concat merging to obtain features The first branch yielded Then use 3x3 convolution and... Merging yields features Next, Features are obtained by inputting into the ConvBlock module. .

[0103] For the third branch, the features obtained from the input... The features are input into CBAM and a 1*1 convolution, respectively, and also into ConvBlock. The features are then concatenated to obtain the final feature. ,Will Features connected to 3x3 convolution and Perform concat to merge the inputs into the ConvBlock model to obtain And obtained from the second branch Features are obtained by concatenating the results of 3x3 convolutions. Then input it into ConvBlock and combine it with the features in the second branch. concat yields features .

[0104] The fourth branch can be obtained by following the same processing method as the three branches above. .

[0105] S30215. Perform concat processing on the first attention branch features, the second attention branch features, the third attention branch features, and the fourth attention branch features to obtain combined features.

[0106] In this embodiment, the features will be obtained as well as They are concatted to obtain features. .

[0107] S30216. Combine the combined features with the spatial convolution features to obtain the character features.

[0108] In this embodiment, the features Features obtained now Merging to obtain features After obtaining the ExtrackBlock through a series of concatenations, the final feature can be obtained. , and features Among its features These are the characteristics of a person.

[0109] S3022. Input the character features into RefineBlock for processing to obtain a key point heatmap.

[0110] In one embodiment, step S3022 specifically includes the following steps: S30221-S30225.

[0111] S30221. Process the human features through deformable convolution to obtain deformable convolution features.

[0112] In this embodiment, the following will be used: Features are obtained through 3x3 deformable convolution. .

[0113] S30222. Input the deformable convolutional features into SCSE for processing to obtain attention weight features.

[0114] In this embodiment, the features The input is fed into the SCSE (Spatial and Channel SE Block) module to obtain attention weight features. .

[0115] S30223. Input deformable convolutional features into a hybrid deep convolutional network and connect it to a sigmoid layer for processing to obtain attention feature maps.

[0116] In this embodiment, The input is fed into a Mixed Depthwise Convolution network and connected to a sigmoid layer to obtain an attention feature map. .

[0117] S30224. The result obtained by multiplying the attention weight feature with the attention feature map is added to the deformable convolution feature to obtain the output feature.

[0118] In this embodiment, the features of the obtained attention weights are... With the obtained attention map Perform multiplication and combine with features Adding them together yields the output features. .

[0119] S30225. The output features are enlarged to their original size using bilinear interpolation to obtain a key point heatmap.

[0120] In this embodiment, the obtained features The image is then enlarged to the size of the original input image using bilinear interpolation to obtain a heatmap of the key points. The number of output channels for keypoints here is the number of keypoints. Here, we have a heatmap with 18 keypoints, which contains 18 channels.

[0121] S303. Input the image of the person with the id index into the optical flow model for processing to obtain the optical flow feature map.

[0122] In this embodiment, the optical flow model uses OFNet, which consists of two main parts: OFsBlock and OFpBlock. The optical flow model is primarily used to extract human optical flow feature maps; that is, for different person indices, the corresponding person optical flow feature maps are obtained by inputting them into the optical flow model.

[0123] In one embodiment, step S303 specifically includes the following steps:

[0124] S3031. Input the two frames of character images with id indexes into residual convolution for processing to obtain the first residual convolution feature and the second residual convolution feature.

[0125] In this embodiment, for OFsBlock, the person images with id indices from the two consecutive frames are input into residual convolution to obtain the corresponding features. and features .

[0126] S3032. Input the first residual convolution feature and the second residual convolution feature into the depthwise separable convolutional network and the spatially separable convolutional network respectively for processing to obtain the first depthwise separable convolutional feature, the first spatially separable convolutional feature, the second depthwise separable convolutional feature, and the second spatially separable convolutional feature.

[0127] S3033. Perform concat processing on the first depthwise separable convolutional feature and the first spatially separable convolutional feature to obtain the first combined feature.

[0128] S3034. Perform concat processing on the second depthwise separable convolutional feature and the second spatially separable convolutional feature to obtain the second combined feature.

[0129] For S3032-S3034, in this embodiment, the features are... and characteristics The inputs are fed into depthwise separable convolutional networks and spatially separable convolutional networks, respectively. The features obtained through the above operations and features The features are obtained by concatenating the two. For In other words, the features obtained after the above operations and features Concatenate the two to obtain the features. .

[0130] S3035. The result of multiplying the first and second combined features is processed by the sigmoid activation function to obtain the activation function features.

[0131] In this embodiment, the features With features After multiplication and passing through the sigmoid activation function, new features are obtained. This feature can better reflect pixel motion information.

[0132] S3036. Perform convolution processing on the first residual convolution feature to obtain the convolution-processed feature.

[0133] In this embodiment, features Features are obtained through 5x5 convolution. .

[0134] S3037. Concat the convolutional features with the activation function features to obtain merged features.

[0135] In this embodiment, the features will be With features Perform concat merging to obtain features .

[0136] S3038. Input the merged features into the convolutional block for processing to obtain the convolutional block features.

[0137] In this embodiment, the features Features are obtained by inputting into convolutional blocks The convolutional block consists of 3x3 convolution, BatchNormalization, and SiLu activation function.

[0138] S3039. Perform dilated convolution upsampling on the activation function features to obtain dilated convolution upsampling features.

[0139] S3040. Perform deconvolution upsampling on the merged features to obtain deconvolution upsampled features.

[0140] S30401. Perform unpooling upsampling on the convolutional block features to obtain unpooling upsampling features.

[0141] S30402. Add the values ​​of the dilated convolution upsampling feature, the deconvolution upsampling feature, and the unpooling upsampling feature to obtain the first added feature.

[0142] For S3039-S30402, in this embodiment, the features... Perform dilated convolution upsampling to extract features. Perform deconvolution upsampling for features Perform unpooling upsampling to ultimately obtain features of consistent size. , and features Finally, the values ​​of the three features are added together to obtain the feature. .

[0143] S30403. After concatting the two frames of character images with id indices, input them into transformerEncoder and deformable convolution respectively to obtain the first and second processed features.

[0144] In this embodiment, for OFsBlock, the two consecutive frames of character images with ID indices are concatenated and then input into the transformerEncoder and deformable convolution respectively to obtain features. and features .

[0145] S30404. The result obtained after resizing the first processed feature is concatted with the second processed feature to obtain the combined feature.

[0146] Following this embodiment, the feature Perform resize and Features are obtained by concatenating features .

[0147] S30405. Input the combined features into two convolutional blocks for processing to obtain the third and fourth processed features.

[0148] In this embodiment, the features The features are obtained by inputting them into two convolutional blocks respectively. and features The convolutional block consists of Depthwise, batch normalization, and ReLU activation functions.

[0149] S30406. The third and fourth processing features are upsampled respectively, and the results of the upsampling are added together to obtain the second added processing result feature.

[0150] In this embodiment, the features and features The features are obtained by upsampling each sample separately and then adding the upsampling results together. .

[0151] S30407. Perform concat processing on the features of the first addition result and the features of the second addition result to obtain intermediate result features.

[0152] In this embodiment, the features and characteristics Features are obtained by concat .

[0153] S30408. Input the intermediate result features into two convolutions for processing to obtain optical flow feature maps.

[0154] In this embodiment, the features The inputs are fed into a 3x3 convolution and a 5x5 convolution respectively. The output features of the 5x5 convolution are concatenated with dilated convolutions, then merged with the results of the 3x3 convolutions, and finally passed through a 1x1 convolution and amplification to obtain the features. This feature has the same size as the input size and has 2 channels.

[0155] S304. Input the key point heatmap and optical flow feature map into the video classification model for processing to obtain the human behavior recognition results.

[0156] In one embodiment, step S304 specifically includes the following steps: S3041-S3043.

[0157] S3041. Map the key points of the Gaussian heatmap on each channel of the key point heatmap to a single-channel graph to obtain a single-channel Gaussian heatmap.

[0158] In this embodiment, based on the obtained The key points of the Gaussian heatmap on each channel are mapped onto the single-channel heatmap to obtain the single-channel Gaussian heatmap. .

[0159] S3042. Scale the optical flow feature map to match the single-channel Gaussian thermal image. Figure 1 The same size is obtained and then concatted with it to obtain key point features containing optical flow information.

[0160] S3043. Input the key point features containing optical flow information into TIN for behavior classification to obtain human behavior recognition results.

[0161] For S3042 and S3043, in this embodiment, the optical flow feature map Scaling proportionally to Achieving consistent dimensions and then concatenating them yields keypoint features containing optical flow information. First, merge the frames corresponding to different people. The frame rate is controlled at 60 frames per second, and then input into a video classification model to obtain the behavior category result for each pedestrian. In this embodiment, TIN (Temporal Interlacing Network) is used for video classification. There are 5 video classification categories: jumping, running, sitting, standing, and walking.

[0162] In addition, the loss functions used in the human behavior prediction model include object detection loss function, keypoint loss function, optical flow loss function, and video behavior classification loss function, among which:

[0163] The loss function used for object detection is the YOLOv7 loss function, which mainly consists of classification loss, localization loss, and confidence loss function, namely:

[0164] .

[0165] The keypoint loss function used is the MSE loss function. Here, we will discuss the keypoint model features separately. , and features Calculate the MSE loss between the ground truth and the actual labeled ground truth, i.e.:

[0166] , here , = 0.3, .

[0167] Optical flow loss function is derived from loss as well as loss The system uniformly uses endpoint error (EPE) as the loss function for training, which represents the average Euclidean distance per pixel between the predicted flow vector and the ground truth.

[0168] The video behavior classification loss function uses the cross-entropy loss of classification for model training.

[0169] This invention combines multi-target tracking, keypoint detection, and optical flow estimation methods to effectively perform real-time behavior recognition of multiple moving human targets, exhibiting excellent compatibility with detecting large-scale human movement within a short timeframe. By fusing optical flow features and keypoint heatmap features, it enables more accurate action analysis for behavior recognition, solving the problem of poor training performance due to over-reliance on background, and improving the model's robustness and accuracy in different scenarios. Furthermore, it considers both spatial and temporal dimensions. Spatially, it incorporates supervision information from original pixel and keypoint positions; temporally, it considers both long-period video sequence information and short-period optical flow information, giving the model a better understanding of complex images.

[0170] Figure 3 This is a schematic block diagram of a multi-person behavior recognition device 100 provided in an embodiment of the present invention. Corresponding to the multi-person behavior recognition method described above, a specific embodiment of the present invention also provides a multi-person behavior recognition device 100. The multi-person behavior recognition device 100 includes units and modules for executing the above-described multi-person behavior recognition method, and the device can be configured in a server.

[0171] like Figure 3 As shown, the multi-person behavior recognition device 100 includes:

[0172] The acquisition unit 110 is used to acquire image data within the detection area.

[0173] The first processing unit 120 is used to perform frame-slicing processing on the image data to obtain multiple frames of static images.

[0174] The second processing unit 130 is used to input multiple frames of static images into the human behavior prediction model in the form of an image sequence for processing, so as to obtain human behavior recognition results.

[0175] In one embodiment, the second processing unit 130 includes:

[0176] The first processing module is used to input multiple frames of static images into the tracking model in the form of an image sequence for processing, so as to obtain a person image with an ID index.

[0177] In one embodiment, the first processing module includes:

[0178] The first processing submodule is used to input multiple frames of static images into the tracking model in the form of an image sequence in order to detect different human target boxes.

[0179] The second processing submodule is used to assign index IDs to the different detected human target boxes in order to obtain human images with ID indexes.

[0180] The second processing module is used to input different images of people with ID indexes into the key point detection model for processing to obtain a key point heatmap.

[0181] In one embodiment, the second processing module includes:

[0182] The third processing submodule is used to input different person images with id indexes into multiple concatenated ExtrackBlocks for processing to obtain person features.

[0183] In one embodiment, the third processing submodule includes:

[0184] The first convolution processing submodule is used to perform depthwise separable convolution and spatial separable convolution processing on different person images with id indices to obtain depthwise convolution features and spatial convolution features.

[0185] The first addition processing submodule is used to add the depthwise convolutional features and the spatial convolutional features to obtain the added result features.

[0186] The second convolution processing submodule is used to process the summed feature through the first convolution, the second convolution, the third convolution and the fourth convolution respectively to obtain the first convolution feature, the second convolution feature, the third convolution feature and the fourth convolution feature.

[0187] The attention processing submodule is used to input the first convolutional feature, the second convolutional feature, the third convolutional feature, and the fourth convolutional feature into the four attention mechanism branches for processing, so as to obtain the first attention branch feature, the second attention branch feature, the third attention branch feature, and the fourth attention branch feature, respectively.

[0188] The first fusion processing submodule is used to concat the first attention branch features, the second attention branch features, the third attention branch features, and the fourth attention branch features to obtain the combined features.

[0189] The first merging processing submodule is used to merge the combined features with the spatial convolution features to obtain the character features.

[0190] The fourth processing submodule is used to input the character features into RefineBlock for processing to obtain a key point heatmap.

[0191] In one embodiment, the fourth processing submodule includes:

[0192] The third convolution processing submodule is used to process human features through deformable convolution to obtain deformable convolution features.

[0193] The fifth processing submodule is used to input deformable convolutional features into SCSE for processing to obtain attention weight features.

[0194] The sixth processing submodule is used to input deformable convolutional features into a hybrid deep convolutional network and connect a sigmoid layer for processing to obtain attention feature maps.

[0195] The second addition processing submodule is used to add the result of multiplying the attention weight features and the attention feature map to the deformable convolution features to obtain the output features.

[0196] The magnification processing submodule is used to magnify the output features to their original size using bilinear interpolation to obtain a key point heatmap.

[0197] The third processing module is used to input the image of the person with the ID index into the optical flow model for processing to obtain the optical flow feature map.

[0198] In one embodiment, the third processing submodule includes:

[0199] The residual convolution processing submodule is used to input the two consecutive frames of character images with ID indices into the residual convolution for processing, so as to obtain the first residual convolution feature and the second residual convolution feature.

[0200] The seventh processing submodule is used to input the first residual convolution feature and the second residual convolution feature into the depthwise separable convolutional network and the spatially separable convolutional network respectively for processing, so as to obtain the first depthwise separable convolution feature, the first spatially separable convolution feature, the second depthwise separable convolution feature, and the second spatially separable convolution feature.

[0201] The first fusion submodule is used to concat the first depthwise separable convolutional features and the first spatially separable convolutional features to obtain the first combined features.

[0202] The second fusion submodule is used to concat the second depthwise separable convolutional feature and the second spatially separable convolutional feature to obtain the second combined feature.

[0203] The activation function processing submodule is used to process the result of multiplying the first and second combined features through the sigmoid activation function to obtain the activation function features.

[0204] The fourth convolution processing submodule is used to perform convolution processing on the first residual convolution features to obtain convolution processed features.

[0205] The third fusion submodule is used to concatenate the convolutional features with the activation function features to obtain merged features.

[0206] The first convolutional block processing submodule is used to input the merged features into the convolutional block for processing to obtain the convolutional block features.

[0207] The dilated convolution upsampling processing submodule is used to perform dilated convolution upsampling processing on the activation function features to obtain dilated convolution upsampling features.

[0208] The deconvolution upsampling processing submodule is used to perform deconvolution upsampling processing on the merged features to obtain deconvolution upsampling features.

[0209] The unpooling upsampling processing submodule is used to perform unpooling upsampling processing on the features of the convolutional blocks to obtain unpooling upsampling features.

[0210] The third addition processing submodule is used to add the values ​​of the dilated convolution upsampling feature, the deconvolution upsampling feature, and the unpooling upsampling feature to obtain the first addition processing result feature.

[0211] The eighth processing submodule is used to concat the two frames of character images with id indices and then input them into transformerEncoder and deformable convolution for processing to obtain the first processing feature and the second processing feature.

[0212] The fourth fusion processing submodule is used to concat the result obtained after resizing the first processed feature with the second processed feature to obtain a combined feature.

[0213] The second convolutional block processing submodule is used to input the combined features into two convolutional blocks for processing to obtain the third and fourth processed features.

[0214] The fourth addition processing submodule is used to upsample the third processing feature and the fourth processing feature respectively, and then add the results of the upsampling processing to obtain the second addition processing result feature.

[0215] The fifth fusion processing submodule is used to perform concat processing on the features of the first addition processing result and the features of the second addition processing result to obtain intermediate result features.

[0216] The ninth processing submodule is used to input the intermediate result features into two convolutions for processing to obtain optical flow feature maps.

[0217] The fourth processing module is used to input the key point heatmap and optical flow feature map into the video classification model for processing to obtain human behavior recognition results.

[0218] In one embodiment, the fourth processing module includes:

[0219] The mapping submodule is used to map the key points of the Gaussian heatmap on each channel of the key point heatmap to a single-channel map to obtain a single-channel Gaussian heatmap.

[0220] The scaling submodule is used to scale the optical flow feature map to a scale similar to that of a single-channel Gaussian thermal image. Figure 1 The same size is obtained and then concatted with it to obtain key point features containing optical flow information.

[0221] The classification submodule is used to input key point features containing optical flow information into TIN for behavior classification in order to obtain human behavior recognition results.

[0222] The aforementioned multi-person behavior recognition device can be implemented as a computer program, which can, for example... Figure 4 It runs on the computer device shown.

[0223] Please see Figure 4 , Figure 4 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 700 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0224] like Figure 4 As shown, the computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the multi-person behavior recognition method described above.

[0225] The computer device 700 can be a terminal or a server. The computer device 700 includes a processor 720, a memory, and a network interface 750 connected via a system bus 710, wherein the memory may include a non-volatile storage medium 730 and internal memory 740.

[0226] The non-volatile storage medium 730 can store an operating system 731 and a computer program 732. When the computer program 732 is executed, it enables the processor 720 to execute any multi-person behavior recognition method.

[0227] The processor 720 provides computing and control capabilities to support the operation of the entire computer device 700.

[0228] The internal memory 740 provides an environment for the execution of the computer program 732 in the non-volatile storage medium 730. When the computer program 732 is executed by the processor 720, the processor 720 can execute any multi-person behavior recognition method.

[0229] This network interface 750 is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 700 to which the present application is applied. The specific computer device 700 may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements. The processor 720 is used to run program code stored in memory to implement the following steps:

[0230] The first aspect is multi-person behavior recognition methods, including:

[0231] Acquire image data within the detection area;

[0232] The image data is sliced ​​into frames to obtain multiple static images;

[0233] Multiple static images are input as an image sequence into the human behavior prediction model for processing to obtain human behavior recognition results.

[0234] In one embodiment: the step of inputting multiple frames of static images as an image sequence into a human behavior prediction model for processing to obtain human behavior recognition results includes:

[0235] Multiple static images are input into the tracking model as an image sequence for processing to obtain human images with ID indexes;

[0236] Different images of people with ID indexes are input into a keypoint detection model for processing to obtain a keypoint heatmap;

[0237] The image of a person with an ID index is input into the optical flow model for processing to obtain an optical flow feature map;

[0238] The key point heatmap and optical flow feature map are input into the video classification model for processing to obtain the human behavior recognition results.

[0239] In one embodiment: the step of inputting different person images with ID indexes into a keypoint detection model for processing to obtain a keypoint heatmap includes:

[0240] Different images of people with ID indices are input into multiple concatenated ExtrackBlocks for processing to obtain the features of the people;

[0241] The character's features are input into RefineBlock for processing to obtain a key point heatmap.

[0242] In one embodiment: the step of inputting different person images with id indexes into multiple concatenated ExtrackBlocks for processing to obtain person features includes:

[0243] Different images of people with ID indices are processed by depthwise separable convolution and spatial separable convolution respectively to obtain depthwise convolution features and spatial convolution features.

[0244] The depthwise convolutional features and the spatial convolutional features are added together to obtain the summed features.

[0245] The summed features are processed by the first convolution, the second convolution, the third convolution, and the fourth convolution to obtain the first convolution feature, the second convolution feature, the third convolution feature, and the fourth convolution feature.

[0246] The first, second, third, and fourth convolutional features are respectively input into four attention mechanism branches for processing to obtain the first attention branch features, the second attention branch features, the third attention branch features, and the fourth attention branch features, respectively.

[0247] The features from the first attention branch, the second attention branch, the third attention branch, and the fourth attention branch are concatted to obtain combined features.

[0248] The combined features and spatial convolution features are merged to obtain the features of the person.

[0249] In one embodiment: the step of inputting the person's features into RefineBlock for processing to obtain a key point heatmap includes:

[0250] The features of the person are processed by deformable convolution to obtain deformable convolution features;

[0251] Deformable convolutional features are input into SCSE for processing to obtain attention weight features;

[0252] Deformable convolutional features are input into a hybrid deep convolutional network and connected to a sigmoid layer for processing to obtain attention feature maps;

[0253] The result of multiplying the attention weight features with the attention feature map is added to the deformable convolution features to obtain the output features.

[0254] The output features are enlarged to their original size using bilinear interpolation to obtain a keypoint heatmap.

[0255] In one embodiment: the step of inputting the image of a person with an ID index into an optical flow model for processing to obtain an optical flow feature map includes:

[0256] The two frames of character images with ID indices are input into residual convolution for processing to obtain the first residual convolution feature and the second residual convolution feature.

[0257] The first residual convolutional feature and the second residual convolutional feature are respectively input into the depthwise separable convolutional network and the spatially separable convolutional network for processing to obtain the first depthwise separable convolutional feature, the first spatially separable convolutional feature, the second depthwise separable convolutional feature, and the second spatially separable convolutional feature.

[0258] The first depthwise separable convolutional feature and the first spatially separable convolutional feature are concatted to obtain the first combined feature;

[0259] The second depthwise separable convolutional feature and the second spatially separable convolutional feature are concatted to obtain the second combined feature;

[0260] The result of multiplying the first and second combined features is then processed by the sigmoid activation function to obtain the activation function features.

[0261] The first residual convolutional feature is convolved to obtain the convolutional processed feature;

[0262] The convolutional features and the activation function features are concatenated to obtain merged features;

[0263] The merged features are input into the convolutional block for processing to obtain the convolutional block features;

[0264] The activation function features are subjected to dilated convolution upsampling to obtain dilated convolution upsampling features;

[0265] The merged features are subjected to deconvolution upsampling to obtain deconvolution upsampled features;

[0266] The convolutional block features are subjected to unpooling upsampling to obtain unpooling upsampling features;

[0267] The values ​​of the dilated convolution upsampling feature, the deconvolution upsampling feature, and the unpooling upsampling feature are added together to obtain the first summed feature.

[0268] The two frames of character images with ID indices are concatted and then input into transformerEncoder and deformable convolution respectively to obtain the first and second processed features.

[0269] The result of resizing the first processed feature is concatted with the second processed feature to obtain the combined feature;

[0270] The combined features are input into two convolutional blocks for processing to obtain the third and fourth processed features.

[0271] The third and fourth processing features are upsampled separately, and the results of the upsampling are added together to obtain the second summed feature.

[0272] The features of the first addition result and the features of the second addition result are concatted to obtain the intermediate result features;

[0273] The intermediate feature data are input into two convolutional layers for processing to obtain optical flow feature maps.

[0274] In one embodiment: the step of inputting the key point heatmap and optical flow feature map into the video classification model for processing to obtain human behavior recognition results includes:

[0275] Map the key points of the Gaussian heatmap on each channel of the key point heatmap to a single-channel heatmap to obtain a single-channel Gaussian heatmap.

[0276] The optical flow feature map was scaled proportionally to the single-channel Gaussian thermal image. Figure 1The same size is obtained and concatted with it to obtain key point features containing optical flow information;

[0277] Keypoint features containing optical flow information are input into TIN for behavior classification to obtain human behavior recognition results.

[0278] It should be understood that, in the embodiments of this application, the processor 720 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0279] Those skilled in the art will understand that Figure 4 The structure of the computer device 700 shown does not constitute a limitation on the computer device 700, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0280] In another embodiment of the present invention, a computer-readable storage medium is provided. This computer-readable storage medium may be a non-volatile computer-readable storage medium. The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the multi-person behavior recognition method disclosed in the embodiments of the present invention.

[0281] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0282] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Units with the same function may be grouped into one unit. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, or it may be an electrical, mechanical, or other form of connection.

[0283] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0284] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0285] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks.

[0286] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for recognizing multiple people's behavior, characterized in that, include: Acquire image data within the detection area; The image data is sliced ​​into frames to obtain multiple static images; Multiple static images are input into the human behavior prediction model as an image sequence for processing to obtain human behavior recognition results. The step of inputting multiple static images as an image sequence into the human behavior prediction model for processing to obtain human behavior recognition results includes: Multiple static images are input into the tracking model as an image sequence for processing to obtain human images with ID indexes; Different images of people with ID indexes are input into a keypoint detection model for processing to obtain a keypoint heatmap; The image of a person with an ID index is input into the optical flow model for processing to obtain an optical flow feature map; The key point heatmap and optical flow feature map are input into the video classification model for processing to obtain human behavior recognition results; The step of inputting the image of a person with an ID index into the optical flow model for processing to obtain an optical flow feature map includes: The two frames of character images with ID indices are input into residual convolution for processing to obtain the first residual convolution feature and the second residual convolution feature. The first residual convolutional feature and the second residual convolutional feature are respectively input into the depthwise separable convolutional network and the spatially separable convolutional network for processing to obtain the first depthwise separable convolutional feature, the first spatially separable convolutional feature, the second depthwise separable convolutional feature, and the second spatially separable convolutional feature. The first depthwise separable convolutional feature and the first spatially separable convolutional feature are concatted to obtain the first combined feature; The second depthwise separable convolutional feature and the second spatially separable convolutional feature are concatted to obtain the second combined feature; The result of multiplying the first and second combined features is then processed by the sigmoid activation function to obtain the activation function features. The first residual convolutional feature is convolved to obtain the convolutional processed feature; The convolutional features and the activation function features are concatenated to obtain merged features; The merged features are input into the convolutional block for processing to obtain the convolutional block features; The activation function features are subjected to dilated convolution upsampling to obtain dilated convolution upsampling features; The merged features are subjected to deconvolution upsampling to obtain deconvolution upsampled features; The convolutional block features are subjected to unpooling upsampling to obtain unpooling upsampling features; The values ​​of the dilated convolution upsampling feature, the deconvolution upsampling feature, and the unpooling upsampling feature are added together to obtain the first summed feature. The two frames of character images with ID indices are concatted and then input into transformerEncoder and deformable convolution respectively to obtain the first and second processed features. The result of resizing the first processed feature is concatted with the second processed feature to obtain the combined feature; The combined features are input into two convolutional blocks for processing to obtain the third and fourth processed features. The third and fourth processing features are upsampled separately, and the results of the upsampling are added together to obtain the second summed feature. The features of the first addition result and the features of the second addition result are concatted to obtain the intermediate result features; The intermediate feature data are input into two convolutional layers for processing to obtain optical flow feature maps.

2. The multi-person behavior recognition method according to claim 1, characterized in that, The step of inputting different person images with ID indexes into a keypoint detection model for processing to obtain a keypoint heatmap includes: Different images of people with ID indices are input into multiple concatenated ExtrackBlocks for processing to obtain the features of the people; The character's features are input into RefineBlock for processing to obtain a key point heatmap.

3. The multi-person behavior recognition method according to claim 2, characterized in that, The process of inputting different person images with ID indices into multiple concatenated ExtrackBlocks for processing to obtain person features includes: Different images of people with ID indices are processed by depthwise separable convolution and spatial separable convolution respectively to obtain depthwise convolution features and spatial convolution features. The depthwise convolutional features and the spatial convolutional features are added together to obtain the summed features. The summed features are processed by the first convolution, the second convolution, the third convolution, and the fourth convolution to obtain the first convolution feature, the second convolution feature, the third convolution feature, and the fourth convolution feature. The first, second, third, and fourth convolutional features are respectively input into four attention mechanism branches for processing to obtain the first attention branch features, the second attention branch features, the third attention branch features, and the fourth attention branch features, respectively. The features from the first attention branch, the second attention branch, the third attention branch, and the fourth attention branch are concatted to obtain combined features. The combined features and spatial convolution features are merged to obtain the features of the person.

4. The multi-person behavior recognition method according to claim 2, characterized in that, The process of inputting character features into RefineBlock for processing to obtain a key point heatmap includes: The features of the person are processed by deformable convolution to obtain deformable convolution features; Deformable convolutional features are input into SCSE for processing to obtain attention weight features; Deformable convolutional features are input into a hybrid deep convolutional network and connected to a sigmoid layer for processing to obtain attention feature maps; The result of multiplying the attention weight features with the attention feature map is added to the deformable convolution features to obtain the output features. The output features are enlarged to their original size using bilinear interpolation to obtain a keypoint heatmap.

5. The multi-person behavior recognition method according to claim 1, characterized in that, The step of inputting key point heatmaps and optical flow feature maps into a video classification model for processing to obtain human behavior recognition results includes: Map the key points of the Gaussian heatmap on each channel of the key point heatmap to a single-channel heatmap to obtain a single-channel Gaussian heatmap. The optical flow feature map is scaled to the same size as the single-channel Gaussian heatmap and then concatted with it to obtain key point features containing optical flow information. Keypoint features containing optical flow information are input into TIN for behavior classification to obtain human behavior recognition results.

6. A multi-person behavior recognition device, which, when operating, executes the multi-person behavior recognition method according to any one of claims 1-5, characterized in that, include: The acquisition unit is used to acquire image data within the detection area; The first processing unit is used to perform frame-slicing processing on the image data to obtain multiple frames of static images; The second processing unit is used to input multiple frames of static images into the human behavior prediction model in the form of an image sequence for processing, so as to obtain human behavior recognition results.

7. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the multi-person behavior recognition method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, the processor performs the steps of the multi-person behavior recognition method as described in any one of claims 1 to 5.

Images