A deep learning-based time-series behavior detection method and system

Abstract

The application belongs to the field of time sequence behavior detection, and provides a time sequence behavior detection method and system based on deep learning, aiming at the problems of complex structure, large amount of calculation and difficult deployment of the existing time sequence behavior detection method, and by referring to the yolo target detection algorithm, the time sequence behavior detection task is designed as three parts of a feature extraction module, a feature fusion module and a prediction module, through the feature blocks based on different scales, average pooling operation is carried out in the space direction, and the start time and the end time of the behavior are predicted in the time sequence direction, so that the application of the scene general target recognition and behavior analysis is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of time series behavior detection, and particularly relates to a time series behavior detection method and system based on deep learning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] At present, time series behavior detection has important significance in the scene general target recognition and behavior analysis, which involves locating the occurrence time interval of a specific behavior in a video sequence and identifying the behavior type. The existing methods for time series behavior detection include sliding window method, segmentation proposal, multi-scale pyramid network and graph convolution method, but these methods have the following technical defects:

[0004] The sliding window method slides a fixed size window on the time axis, and then uses a pre-trained behavior recognition model to classify each window. It is simple and easy to implement, but the calculation amount is large, and there may be more overlapping detection results.

[0005] The segmentation proposal method first generates potential behavior segment proposals from the video, and then classifies the proposals. For example, the Temporal Actionness Grouping (TAG) method generates proposals by calculating the behavior heat; the BSN (Boundary-Sensitive Network) method generates proposals by using boundary sensitivity.

[0006] The multi-scale pyramid network considers the behavior features at different time scales, and is suitable for processing behaviors of different lengths. For example, MultiScale-TRN (Temporal Relation Networks) decomposes the video into multiple scales and uses the TRN model to learn the behavior relationship between different scales. This method helps to capture the diversity of behaviors.

[0007] The graph convolution network uses graph structure to model the behavior segments with strong correlation. For example, T-CNN (Temporal Convolutional Networks) uses a graph convolution network to model the video segments and capture the correlation between different time segments. This kind of method has advantages in handling long time dependency in video.

[0008] Although the above-mentioned time sequence behavior detection method has achieved good results in many practical applications, there are still some defects, one of which is the high complexity of the implementation method. Due to the high dimension and long time sequence characteristics of video data, special data processing and model structure are needed for processing, which increases the complexity of the implementation method, and also makes the calculation complexity of the model higher and the deployment more difficult. SUMMARY

[0009] In order to solve at least one technical problem in the above background art, the present application provides a time sequence behavior detection method and system based on deep learning, which solves the problems of complex structure, large calculation amount and difficult deployment of the existing time sequence behavior detection method by performing average pooling operation in the spatial direction based on different scale feature blocks and predicting the start time and end time of the behavior in the time sequence direction.

[0010] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0011] The first aspect of the present application provides a time sequence behavior detection method based on deep learning, comprising the following steps:

[0012] obtaining a video sequence to be detected;

[0013] based on the video sequence to be detected and the trained time sequence behavior detection model, obtaining the category of the behavior, the boundary of the behavior and the quality of the behavior;

[0014] The construction process of the time sequence behavior detection model is:

[0015] based on the video sequence to be detected, performing feature extraction to obtain a plurality of feature blocks with different resolutions;

[0016] extending the plurality of feature blocks with different resolutions to obtain corresponding feature blocks with different scales;

[0017] based on the feature blocks with different scales, performing average pooling operation in the spatial direction, predicting the start time and end time of the behavior in the time sequence direction, and predicting the behavior in different length periods according to different scale three-dimensional feature blocks.

[0018] The second aspect of the present application provides a time sequence behavior detection system based on deep learning, comprising:

[0019] a data acquisition module for acquiring a video sequence to be detected;

[0020] a behavior detection module for obtaining the category of the behavior, the boundary of the behavior and the quality of the behavior based on the video sequence to be detected and the trained time sequence behavior detection model;

[0021] The construction process of the time sequence behavior detection model is as follows:

[0022] Feature extraction is performed based on the video sequence to be detected, and a plurality of feature blocks with different resolutions are obtained.

[0023] The plurality of feature blocks with different resolutions are expanded to obtain corresponding feature blocks of different scales.

[0024] Based on the feature blocks of different scales, average pooling operation is performed in the spatial direction, the start time and the end time of the behavior are predicted in the time sequence direction, and the behavior in different length periods is predicted based on the three-dimensional feature blocks of different scales.

[0025] The third aspect of the present application provides a computer readable storage medium.

[0026] A computer readable storage medium has a computer program stored thereon, and the program is executed by a processor to implement the steps of the deep learning-based time sequence behavior detection method according to the first aspect.

[0027] The fourth aspect of the present application provides a computer device.

[0028] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the program to implement the steps of the deep learning-based time sequence behavior detection method according to the first aspect.

[0029] Compared with the prior art, the present application has the following advantages:

[0030] The present application aims at the problems of complex structure, large amount of calculation and difficult deployment of the existing time sequence behavior detection method, and learns from the yolo target detection algorithm, and designs the time sequence behavior detection task as three parts of a feature extraction module, a feature fusion module and a prediction module, and performs average pooling operation in the spatial direction based on the feature blocks of different scales, and predicts the start time and the end time of the behavior in the time sequence direction, so as to realize the application of scene-based general target recognition and behavior analysis.

[0031] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be known by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0032] The drawings constituting a part of the specification of the present application are used to provide further understanding of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute improper limitation on the present application.

[0033] Figure 1is a flow chart of a temporal action detection method based on deep learning provided by an embodiment of the present application;

[0034] Figure 2 is a schematic diagram of the overall structure of a temporal action detection model provided by an embodiment of the present application;

[0035] Figure 3 is a schematic diagram of the specific structure of a temporal action detection model provided by an embodiment of the present application;

[0036] Figure 4 is a schematic diagram of the calculation process of a feature pyramid provided by an embodiment of the present application;

[0037] Figure 5 is a fusion schematic diagram of a feature fusion module provided by an embodiment of the present application. DETAILED DESCRIPTION

[0038] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0040] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of the features, steps, operations, devices, components and / or combinations thereof.

[0041] Temporal action detection refers to detecting and locating the time period and type of behavior occurrence in a video sequence. Unlike traditional behavior recognition tasks, temporal action detection requires accurate positioning and classification of the time of behavior occurrence. Temporal action detection has important significance in practical applications, such as video surveillance, sports competition analysis, etc. In video surveillance, temporal action detection can automatically identify abnormal events such as theft, attack, etc., thereby realizing intelligent security; in sports competition analysis, temporal action detection can automatically identify scoring behavior, foul behavior, etc. in the game, thereby helping referees make more fair decisions.

[0042] Therefore, temporal action detection is an important research direction in the field of computer vision, and its research significance lies in improving the automated processing capability of video data, thereby realizing more intelligent and efficient video analysis applications.

[0043] Embodiment one

[0044] As Figure 1 shown, the present application provides a deep learning-based time sequence behavior detection method, comprising the following steps:

[0045] S1, sample collection: collect video sequences from monitoring videos, label behaviors of the video sequences, and construct a time sequence behavior detection dataset.

[0046] S2, time sequence behavior detection model design: referring to the yolo series target detection algorithm, the time sequence behavior detection task is designed into three parts of feature extraction module, feature fusion module and prediction module.

[0047] S3, time sequence behavior detection model training: using the dataset made in S1, the behavior recognition model designed in S2 is trained.

[0048] S4, time sequence behavior detection model inference: inputting the image into the behavior recognition model trained in S3, the model performs inference operation on the input video data, and outputs the result of video behavior recognition.

[0049] In S1, video clips containing different behaviors are collected from monitoring videos, and data related to the video are recorded, such as timestamp, camera position, behavior label, etc.

[0050] Next, the collected data is cleaned and labeled by artificial means, redundant data and abnormal data are removed, and the start time, end time and type of behavior in the video are labeled.

[0051] In order to improve the efficiency of model training, the video sequence length is sampled to 128 frames by using equal interval sampling method in this embodiment, and the resolution of video frame is scaled to 224x224.

[0052] In S2, the time sequence behavior detection model of the present application is composed of three modules, namely feature extraction module, feature fusion module and prediction module. Its structure diagram is as Figures 2-3 shown:

[0053] The feature extraction module is used to convert the input video sequence into a higher level abstract feature representation, so as to perform subsequent tasks.

[0054] In this embodiment, 3D Resnet50 neural network is used to extract features from the input video sequence.

[0055] 3D ResNet is a deep learning network applied to video recognition tasks, which is based on the classic ResNet architecture (Residual Network) and introduces 3D convolution to process video data. This allows the network to capture information in the spatial and temporal dimensions of the video, enabling effective recognition of video content.

[0056] Compared to traditional 2D ResNet, 3D ResNet uses 3D convolution kernels (width, height, and time), which allows it to directly process consecutive video frames and learn spatial-temporal features. In 3D ResNet, each convolutional layer uses 3D convolution operations, considering both the spatial structure and temporal continuity of the video. In addition, 3D ResNet inherits the main features of ResNet, using residual connections to solve the gradient vanishing and representation bottleneck problems, allowing the network to train deeper architectures more easily.

[0057] The feature fusion module is used to transfer information and fuse features of different sizes extracted by the feature extraction module to improve the model's ability to recognize targets.

[0058] The different levels of features extracted by the feature extraction module have different receptive fields and semantic information.

[0059] The shallower layers provide higher resolution and local detail information after feature fusion, while the deeper layers provide richer semantic information. Therefore, information transfer and feature fusion between different feature layers can effectively improve the accuracy of target detection.

[0060] In this embodiment, the fusion algorithm used in the feature fusion module is to extend the FPN algorithm to a 3DFPN algorithm, i.e., extend the 2D 1x1 convolution kernel in FPN to a 3D 1x1x1 convolution kernel, and extend the 2D upsampling layer operation to a 3D upsampling layer.

[0061] The input of this module is three different scale feature blocks output by 3D ResNet, and the output is three different scale feature blocks that fuse features of different levels.

[0062] As shown in Figure 4 the specific process of feature extraction using a feature pyramid is as follows:

[0063] First, the model 3D ResNet performs convolution operations on the input image to generate three different resolution feature blocks: the first resolution feature block C3, the second resolution feature block C4, and the third resolution feature block C5.

[0064] Then, a 1x1x1 convolution is used to change the number of feature channels of the third resolution feature block C5 to obtain the first scale feature block P5.

[0065] Finally, the first scale feature block P5 is up-sampled and added to the second resolution feature block C4 to obtain the second scale feature block P4.

[0066] Since the second resolution feature block C4 and the first scale feature block P5 have different numbers of feature channels, the number of feature channels of C4 is changed to be equal to that of P5 using a 1x1x1 convolution before the two are added.

[0067] Similarly, P4 is up-sampled and added to the first resolution feature block C3 to obtain the third scale feature block P3.

[0068] P3, P4 and P5 are used as feature blocks for fusing features of different layers for subsequent prediction of detection results.

[0069] The detection module includes a behavior category prediction module, a boundary prediction module and a mask prediction module.

[0070] The behavior category prediction module outputs the category to which the behavior belongs, the boundary prediction module outputs the boundary of the behavior, and the mask prediction module generates a mask indicating the occurrence of the behavior, which is used to represent the quality of the predicted behavior.

[0071] The feature fusion module outputs three feature blocks of different scales, each of which needs to be used for prediction of a detection result.

[0072] Among them, the large-scale feature block is used to predict a short-period behavior, and the small-scale feature block is used to predict a long-period behavior.

[0073] Since the temporal behavior detection only needs to predict the start time and end time of the behavior in the time sequence direction, an average pooling operation can be performed in the spatial direction to change the scale of the input feature block to Tx1x1xC, where T represents the length of the feature block in the time sequence direction, and C represents the number of feature channels. This operation can effectively reduce the calculation amount of the prediction layer.

[0074] In order to improve the prediction accuracy, a 1x1x1 convolution layer is used to decouple the prediction result, and a detection head for decoupling the prediction result is shown as follows: Figure 5

[0075] Figure 5 Among them, the behavior category prediction outputs the score of each behavior belonging to each category among all categories, and the interval is 0-1. The highest score is taken as the recognition result of the behavior category.

[0076] ​The boundary prediction branch outputs the behavior start position and the length of the behavior in the sequence.

[0077] The mask prediction branch is used to predict the mask of the behavior, and each video frame in the video sequence is predicted as a 0-1 value.

[0078] In S3, the behavior recognition model designed in S2 is trained using the time sequence behavior detection dataset constructed in S1.

[0079] In the training stage, the model first performs forward operation on the input video sequence and outputs the prediction value of the time sequence behavior detection.

[0080] In this embodiment, the simOTA algorithm is used to realize the allocation of labels in the training process, then the behavior class loss function, the boundary loss function and the mask loss function are used to calculate the loss generated in the model training process, and finally the model parameters are optimized based on the loss using the optimizer.

[0081] The total model loss value is represented as:

[0082] L=L hm +λ size L size +λ mask L mask

[0083] In the formula, L hm represents the behavior class loss; L mask represents the behavior quality loss; L size represents the boundary prediction loss; λ mask and λ size are weight coefficients of the behavior class prediction loss L mask and the boundary prediction loss L size .

[0084] In this embodiment, λ mask is set to 1 and λ size is set to 0.1.

[0085] In the formula, the behavior class loss function L hm is a focal loss, and its formula is:

[0086]

[0087] In the formula, N is the number of objects to be detected in the image; Y xyc_p is the behavior class predicted by the behavior prediction network; Y xyc_g is the class labeled in the video, and when Y xyc_g is 1, the loss of the positive sample is calculated, and when Y xyc_gis 0, the loss of negative samples is calculated; and and are used to adjust the loss of positive samples and the loss of negative samples generated during model training, to solve the problem of imbalance between positive samples and negative samples in the target detection task.

[0088] In this embodiment, = 2 and = 4.

[0089] Boundary prediction loss function L size is a DIOU loss function, and its calculation formula is:

[0090]

[0091] In the formula, b, b gt respectively represent the center points of the predicted frame and the real frame, and represents the Euclidean distance between the two center points. represents the diagonal distance of the smallest closed region that can contain the predicted frame and the real frame.

[0092] Center point bias loss function L off is:

[0093]

[0094] In the formula, N represents the number of frames of the video sequence, O mask-p represents the mask value predicted in the time sequence direction of the video sequence; and O mask is the mask real label of the video sequence.

[0095] In S4, the video sequence to be recognized is sampled and scaled, and the sampled and scaled video sequence is input into the time sequence behavior detection model trained in S3, and the model outputs the class of the behavior, the position where the behavior occurs, and the mask of the behavior.

[0096] When predicting, the proportion of the values greater than 0.6 in the mask branch prediction result is calculated in the behavior region predicted by the boundary prediction branch, and the value is the quality score of the predicted behavior.

[0097] The confidence of the behavior prediction result is obtained by multiplying the predicted value of the behavior class prediction branch by the quality score of the behavior.

[0098] When the confidence value is greater than the set confidence threshold, the behavior is input to the NMS post-processing algorithm, and the result output by the NMS is the final time sequence behavior detection result.

[0099] In this embodiment, the confidence threshold is 0.6.

[0100] The NMS (Non-Maximum Suppression) algorithm is a post-processing method used in the detection task, which aims to eliminate redundant candidate frames and only keep the frames that are most likely to contain the detection target.

[0101] In this embodiment, the prediction interval of the behavior is taken as a detection frame for NMS calculation.

[0102] The following steps are included:

[0103] All candidate frames are sorted according to their probability of containing the detection target, from high to low.

[0104] The candidate frame with the highest probability is selected as the reserved frame, and the frame is deleted from the candidate frame list.

[0105] The remaining candidate frames are calculated for their Intersection over Union (IoU) with the selected reserved frame.

[0106] If the IoU of a candidate frame with the selected reserved frame is greater than a certain threshold (0.7 is used in this embodiment), the candidate frame is deleted from the candidate frame list.

[0107] Repeat S2-S4 until the candidate frame list is empty or the number of reserved frames reaches the set maximum number.

[0108] Through this process, the final reserved frame list can be obtained, in which each frame has a high probability and low overlap between them, which can effectively avoid redundant detection.

[0109] Embodiment Two

[0110] This embodiment provides a deep learning-based time-series behavior detection system, characterized in that it comprises:

[0111] A data acquisition module for acquiring a video sequence to be detected;

[0112] A behavior detection module for obtaining the class of behavior, the boundary of behavior, and the quality of behavior based on the video sequence to be detected and the trained time-series behavior detection model;

[0113] Wherein, the construction process of the time-series behavior detection model is:

[0114] Based on the video sequence to be detected, feature extraction is performed to obtain multiple feature blocks with different resolutions;

[0115] The multiple feature blocks with different resolutions are expanded to obtain corresponding feature blocks with different scales;

[0116] Based on the feature blocks with different scales, average pooling operation is performed in the spatial direction, the start time and end time of the behavior are predicted in the time-series direction, and the behavior in different periods is predicted according to the three-dimensional feature blocks with different scales.

[0117] Embodiment Three

[0118] The embodiment provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement steps in the deep learning based time sequence behavior detection method in the embodiment one.

[0119] Embodiment four

[0120] The embodiment provides a computer device, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the processor implements steps in the deep learning based time sequence behavior detection method in the embodiment one when executing the program.

[0121] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can take the form of a hardware embodiment, a software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0122] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device implemented in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by the one or more blocks.

[0123] These computer program instructions can also be stored in a computer readable storage medium capable of guiding the computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction means, which implements the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0124] These computer program instructions can also be loaded into a computer or other programmable data processing device to make a series of operation steps executed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide a means for implementing the flowcharts and / or block diagrams.Figure 1 one or more processes and / or functions specified in one or more blocks. Figure 1 one or more blocks or any combination thereof.

[0125] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a computer readable storage medium, and when the program is executed, the processes of the above-mentioned embodiment methods can be included. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), a random access memory (RAM), or the like.

[0126] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A temporal behavior detection method based on deep learning, characterized in that, Includes the following steps: Obtain the video sequence to be detected; Based on the video sequence to be detected and the trained temporal behavior detection model, the category, boundary, and quality of the behavior are obtained. The construction process of the temporal behavior detection model is as follows: Feature extraction is performed on the video sequence to be detected to obtain multiple feature blocks with different resolutions; Multiple feature blocks with different resolutions are expanded to obtain corresponding feature blocks of different scales; Based on feature blocks of different scales, average pooling is performed in the spatial direction to predict the start and end times of behavior in the temporal direction, and behavior within different length periods is predicted based on 3D feature blocks of different scales. When training the temporal behavior detection model, the simOTA algorithm is used to assign labels during the training process. The loss generated during model training is calculated using the behavior category loss function, boundary loss function, and mask loss function, respectively. Based on this loss, an optimizer is used to optimize the model parameters.

2. The temporal behavior detection method based on deep learning as described in claim 1, characterized in that, The 3DResnet50 neural network is used to extract features from the video sequence to be detected.

3. The temporal behavior detection method based on deep learning as described in claim 1, characterized in that, When expanding multiple feature blocks with different resolutions, the method used is to expand the FPN algorithm into a 3D FPN algorithm, that is, to expand the 2D 1×1 convolution kernel in FPN into a 3D 1×1×1 convolution kernel, and to expand the 2D upsampling layer operation into a 3D upsampling layer.

4. The temporal behavior detection method based on deep learning as described in claim 1, characterized in that, The step of expanding multiple feature blocks with different resolutions to obtain corresponding feature blocks of different scales specifically includes: By changing the number of feature channels in the third-resolution feature block, the first-scale feature block can be obtained. An upsampling operation is performed on the first-scale feature block, and the upsampled feature block is added to the second-resolution feature block to obtain the second-scale feature block; The second-scale feature block is upsampled and then added to the first-resolution feature block to obtain the third-scale feature block. The first-scale feature block, the second-scale feature block, and the third-scale feature block are used as three types of feature blocks that fuse features from different layers.

5. The temporal behavior detection method based on deep learning as described in claim 1, characterized in that, The predicted value of the behavior category is multiplied by the quality score of the behavior to obtain the confidence level of the prediction result. When the confidence level is greater than the set threshold, the behavior is input into the NMS post-processing algorithm. The output of NMS is the final result of temporal behavior detection.

6. The temporal behavior detection method based on deep learning as described in claim 5, characterized in that, The final temporal behavior detection results are obtained by using the NMS post-processing algorithm, specifically including: Sort all candidate boxes according to the probability that they contain the detected target, from high to low; Select the candidate box with the highest probability, keep it as the candidate box, and delete it from the candidate box list. For the remaining candidate boxes, calculate their overlap with the selected retained boxes. If the overlap between a candidate box and a selected retained box is greater than the overlap threshold, then remove the candidate box from the candidate box list. Otherwise, continue until the candidate box list is empty or the number of retained boxes reaches the set maximum number.

7. A temporal behavior detection system based on deep learning, characterized in that, include: The data acquisition module is used to acquire the video sequence to be detected; The behavior detection module is used to obtain the behavior category, behavior boundary, and behavior quality based on the video sequence to be detected and the trained temporal behavior detection model. The construction process of the temporal behavior detection model is as follows: Feature extraction is performed on the video sequence to be detected to obtain multiple feature blocks with different resolutions; Multiple feature blocks with different resolutions are expanded to obtain corresponding feature blocks of different scales; Based on feature blocks of different scales, average pooling is performed in the spatial direction to predict the start and end times of behavior in the temporal direction, and behavior within different length periods is predicted based on 3D feature blocks of different scales. When training the temporal behavior detection model, the simOTA algorithm is used to assign labels during the training process. The loss generated during model training is calculated using the behavior category loss function, boundary loss function, and mask loss function, respectively. Based on this loss, an optimizer is used to optimize the model parameters.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the deep learning-based temporal behavior detection method as described in any one of claims 1-6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the deep learning-based temporal behavior detection method as described in any one of claims 1-6.

Images