Hidden social model-based fight behavior identification method

Abstract

A fight behavior identification method based on an implicit social model relates to the technical field of behavior identification and detection, and comprises the following steps: step S01, extracting continuous image frames from a monitoring video stream; s02, training an action target detector by adopting a fight image data set with marked action candidate box coordinates and category information; step S03, performing image enhancement preprocessing on the image frame, and then sending the image frame into a human body target key point detection model adopting a YOLOPose human body key point extraction model; and step S04, carrying out preprocessing operation and the like on the extracted human body key point information. According to the method, on the basis of a YOLOPose and ActionNet video understanding framework, the social information feature extraction network is embedded into the framework while the inter-frame time sequence and the single-frame space information are analyzed, and the interaction information between people is extracted in a feature form, so that the fighting behavior and the occurrence position of the fighting behavior under the monitoring video are accurately identified.

Description

Technical Field

[0001] The present invention relates to the technical field of behavior recognition and detection, and in particular to a fighting behavior recognition method based on an implicit social model. Background Art

[0002] With the development of society and the advancement of technology, public safety has become an important factor in urban management. In order to maintain good social order, prevent the occurrence of violent incidents, and respond to emergencies in a timely and effective manner, more and more cities have begun to adopt advanced video surveillance technology and artificial intelligence systems to improve the level of safety management in public areas. It is particularly important to take effective measures to prevent and quickly respond to such incidents in public places. Fighting behavior recognition technology, as an emerging technical means, has shown great potential in this field.

[0003] Fighting behavior recognition technology is mainly based on computer vision and deep learning principles. By analyzing the video stream captured by the camera, it automatically detects fighting behaviors in the picture. When a suspected fighting behavior pattern is identified, the system will immediately trigger an alarm and notify security personnel or the police to intervene in time, thereby minimizing the harm caused by violent incidents. Deploying fighting behavior recognition algorithms in public places is not only a reflection of technological progress, but also an important measure to ensure public safety and maintain social stability.

[0004] Existing methods for identifying fighting behaviors are mainly divided into two categories: the first category is based on multi-frame videos, and analyzes and identifies fighting behaviors in videos by designing video understanding models, or identifies target actions by designing skeleton extraction and skeleton behavior recognition models. This type of method can simultaneously analyze the temporal information between video frames and the spatial information of a single frame, but the network can only output the category of fighting behaviors and human skeleton information, and cannot locate the specific location where the fighting behavior occurs in the video. And because this type of method requires video analysis, it occupies a large amount of resources, has a slow processing speed, and cannot meet real-time requirements. The second category is based on single-frame images, and detects the category and location of fighting behaviors in the image in the form of target detection. This type of method does not combine the relationship between video frames, and simply detects fighting behaviors from the perspective of images. Compared with the first type of method, it is faster and has fewer parameters, but it has more false alarms and will identify many actions similar to the target fighting behavior. Secondly, fighting behaviors often occur in two-person or multi-person scenes, and existing methods cannot extract interaction information between people, resulting in false alarms and missed alarms when identifying fighting behaviors. Summary of the invention

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a fighting behavior recognition method based on an implicit social model. Based on the YOLOPose and ActionNet video understanding framework, a social information feature extraction network is embedded therein while analyzing the inter-frame temporal sequence and single-frame spatial information, and the interaction information between people is extracted in the form of features, thereby realizing accurate recognition of fighting behaviors and their occurrence locations in surveillance videos.

[0006] In order to achieve the above object, the main technical solutions adopted by the present invention include: A fighting behavior recognition method based on an implicit social model comprises the following steps: Step S01: Extracting continuous image frames from the surveillance video stream F t ( t =1,2,..., f );in, f is the number of consecutive image frames; Step S02, using a fighting image dataset with action candidate box coordinates and category information labeled "fighting" and "non-fighting" to train an action target detector using YOLOv7 as a target detection model; Step S03: image frame F t The image is pre-processed for image enhancement and then sent to the human target key point detection model using the YOLOPose human key point extraction model to obtain the coordinates of the human skeleton points; at the same time, the image frame F t Send it to the action target detector in step S02 to obtain the image frame F t The target detection result of the fighting behavior in the target detection frame includes the image position information of the fighting behavior and the fighting behavior category, wherein the behavior category includes fighting and normal; Step S04: Combine the target detection frame bounding box coordinates of the fighting action in step S03, perform preprocessing operations on the extracted human key point information, and set a fixed calculation frame length of N , by the visible key points K n ( n =1,2,...,17), calculate the fixed frame length according to formula (1) M The speed of movement of individual v i ( i =1,2,..., M ), (1) in, forN Intraframe n The pixel displacement set of each key point of the human body, is the displacement length of the key point with the largest displacement in the set; the displacement speeds of the key points of the human body with three speed levels of low speed, medium speed and high speed are defined. M The movement speeds of individual behaviors are divided into three clusters with different speed magnitudes, and we get N Intra M The motion information of each behavioral individual is recorded as m , m =( m 1, m 2, m 3), m Indicates that M Individual behavior in N The set of visible key points in the frame; Step S05: construct a social information feature extraction network, which is composed of a spatial convolution module and a temporal convolution module. The spatial feature output by the spatial convolution module is V spatial , the temporal features output by the temporal convolution module are V temporal , by V spatial and V temporal Convolution fusion obtains spatiotemporal features V st ; Step S06: Set up a social information quantification module to quantify the spatiotemporal features in step S05 into scores; the social information quantification module uses a fully connected network based on a logistic regression algorithm to train the weights of the fully connected network. V st Mapping to score , is the social information score for different speed magnitude sets; Step S07, using a pseudo 3D video understanding module with the ActionNet network as the basic architecture to extract the action category of continuous image frames, the pseudo 3D video understanding module is composed of a spatiotemporal excitation module, a channel excitation module and a motion excitation module, and the three modules are connected in parallel to form the backbone network of the video understanding network. After the video understanding training, the video level behavior recognition task is completed to obtain the predicted behavior category; Step S08, embedding the social information quantification module in step S06 into the pseudo 3D video understanding module in step S07 to form a fighting behavior recognition model based on an implicit social model, and training the fighting behavior recognition model based on the implicit social model in a multi-loss feedback form; Step S09: using the fighting behavior recognition model based on the implicit social model trained in step S08 to determine whether there is fighting behavior in the video frame to be detected.

[0007] Furthermore, in step S02, the fighting image dataset consists of RGB images and skeleton images, and the fighting image dataset includes fighting behavior images captured from the public datasets UBI-Fights and Hockey-Fights datasets, as well as collected indoor fighting behavior images.

[0008] Furthermore, in step S03, the image frame F t The image enhancement preprocessing includes random cropping, flipping, and random color perturbation operations on the image.

[0009] Furthermore, in step S04, the speed threshold set [0, 0.01, 0.1, 1.2] is used to set the current N Intra M The movement speed of the human body is divided into three clusters with different speed magnitudes, among which the human body N The intra-frame motion speed is obtained by calibrating the camera, establishing the mapping relationship between the image pixel coordinates and the world coordinates of the corresponding points in space, and obtaining the target human body's N The actual displacement within the frame time is used to calculate the human body movement speed, and the human body targets are classified according to the speed value interval. The key points of the human body with different speed magnitudes after classification are given by formula (2): (2) in, C x , C y are the upper left and lower right coordinate points of the image position box where the fighting behavior occurs, marked by the target detection box. W , H are the width and height of the coordinate frame respectively. box conf , class conf are the confidence values ​​of the target detection box and the detection category respectively, , Respectively n The horizontal and vertical coordinate values ​​of key points of the individual body, Indicates n Confidence values ​​of individual body key points.

[0010] Furthermore, in step S05, the spatiotemporal features V st According to formula (3)-(5), (3) (4) (5) in, V spatial is the spatial feature output by the spatial convolution module, V temporal represents the temporal features output by the temporal convolution module, V st for V spatial and V temporal The spatiotemporal features after convolution fusion, Conv 2D is 2D convolution, and RELU is the activation function in artificial neural networks.

[0011] Furthermore, in step S06, the loss function of the social information quantification module is formula (6); (6) in, S p represents the prediction score, S t represents the true score, m Indicates the number of samples.

[0012] Furthermore, in step S07, the pseudo 3D video understanding module is inserted into the ResNet-50 backbone network to form the basic architecture of the video understanding module.

[0013] Furthermore, in step S08, the video classification loss function of the pseudo 3D video understanding module uses FocalLoss, that is, formula (7), (7) in, is the loss value of the video classification loss function, is the weight parameter of easy-to-classify samples, is a hyperparameter. For positive samples, equal , for negative samples, equal , According to formula (8), (8) in, The probability that the model predicts the sample category to be "fighting" is is the true category value of the predicted category.

[0014] Furthermore, in step S09, the method for determining whether there is a fighting behavior in the video frame to be detected is that the video classification module outputs the "fighting" category, and a target box of the "fighting" category appears in a single frame of video, and the target box, i.e., the category prediction probability value, is greater than the threshold set by the model reasoning, which is set to 0.5 and 0.65 respectively during reasoning. Video frames that meet the above conditions are judged as having a "fighting" behavior.

[0015] Furthermore, the training loss function of the fighting behavior recognition model based on the implicit social model is formula (9), (9) in, , are the loss weights for classification and regression tasks respectively, , They are the video classification loss function and the social information score prediction module loss function respectively.

[0016] The beneficial effects of the present invention are: 1. The present invention uses the change form of human key points in multiple consecutive frames to model speed information, reflecting the change of inter-frame speed when fighting occurs, and enhancing the ability of the network to identify fighting behavior; 2. The present invention combines the structure of the fusion of human skeleton point motion speed information and social relationship feature extraction network, extracts the social relationship of human bodies with similar moving speeds in multiple frames based on convolutional neural network, and realizes effective modeling for fighting behaviors with fast speed changes; 3. The present invention adopts a lightweight and real-time pseudo 3D video understanding module to enhance the network's ability to extract timing information; 4. The present invention adopts the Focal Loss loss function to solve the model training problem caused by sample imbalance. It also quantifies social information into scores, and jointly learns with the video understanding module, combining classification tasks and regression tasks to improve the accuracy of fighting behavior recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of the structure of a social information feature extraction network of the present invention; Figure 2 Schematic diagram of the structure of the pseudo 3D video understanding module of the present invention; Figure 3 A schematic diagram of the basic architecture of the video understanding module of the present invention; Figure 4 It is a schematic diagram of the basic architecture of the fighting behavior identification method model based on implicit social model of the present invention. DETAILED DESCRIPTION

[0018] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0019] The present invention provides a fighting behavior recognition method based on an implicit social model, comprising the following steps: Step S01: Extracting continuous image frames from the surveillance video stream F t ( t =1,2,..., f, );in, f is the number of consecutive image frames; Step S02, using a fight image dataset with action candidate frame coordinates and category information labeled "fight" and "non-fight", training an action target detector with YOLOv7 as the target detection model to achieve image-level prediction of fighting behavior. Specifically, the action target detector uses the YOLOv7 target detection model as the basic architecture, and uses fighting images to pre-train the YOLOv7 model. The trained model is used to obtain the target detection result of the fighting behavior in the image frame, including the image position information of the fighting behavior marked by the target detection frame and the fighting behavior category. The above two types of information are used to pre-process the human body motion information and assist in building a social information feature extraction network.

[0020] The fight image dataset contains images collected from public datasets and image collections from video stream image frames, consisting of RGB images and skeleton images. Specifically, the fight image dataset contains fighting behavior images captured from the public datasets UBI-Fights and Hockey-Fights datasets, as well as collected indoor fighting behavior images. More specifically, the fight image dataset consists of 3486 RGB images and 1754 skeleton images, of which the training set consists of 4200 samples and the test set consists of 1040 samples. The data annotations contain the coordinates and category information of the action candidate boxes of "fighting" and "non-fighting".

[0021] Step S03: image frame F t Perform image enhancement preprocessing, specifically, the image frame F t The image enhancement preprocessing includes random cropping, flipping, and random color perturbation operations on the image, and then feeding it into the human target key point detection model using the YOLOPose human key point extraction model to obtain the coordinates of the human skeleton points. Specifically, the image size of the input human key point detection model is 640, the human key point confidence threshold is set to 0.5, and the candidate box intersection ratio threshold is set to 0.65. At the same time, the motion target detector is used to detect the image frame. Ft To perform target action recognition, the image frame F t Send it to the action target detector in step S02 to obtain the image frame F t The target detection result of fighting behavior in the target detection box includes the image location information of the fighting behavior marked by the target detection box and the fighting behavior category, and the behavior category includes fighting and normal. Specifically, the image size input to the action target detector is 640, the action category confidence threshold is set to 0.55, and the candidate box intersection-to-union ratio threshold is set to 0.55.

[0022] After the training of the fighting behavior recognition model based on the implicit social model of the present invention is completed, the monitoring video stream used for reasoning and use is the video stream captured by remote monitoring acquired in real time. The target detector and the human target key point detection model are respectively the YOLOv7 target detection model and the YOLOPose human key point extraction model, which are responsible for processing the collected video surveillance data. The former is used to output the detection results of the fighting category, and the latter is used to output the human key points.

[0023] Step S04: Combine the target detection frame bounding box coordinates of the fighting action in step S03, perform preprocessing operations on the extracted human key point information, and set a fixed calculation frame length of N Specifically, the fixed calculation frame length is set to 16, and the visible key points K n ( n =1,2,...,17), calculate the fixed frame length according to formula (1) M The speed of movement of individual v i ( i =1,2,..., M ), (1) in, for N Intraframe n The pixel displacement set of each key point of the human body, is the displacement length of the key point with the largest displacement in the set; the displacement speeds of the key points of the human body are defined in three speed levels: low speed, medium speed, and high speed. Specifically, the speed threshold set is set to [0, 0.01, 0.1, 1.2]. That is, if v i If the value of is greater than 0 and less than or equal to 0.01, it is low speed; if v i If the value of is greater than 0.01 and less than or equal to 0.1, it is medium speed; if v iIf the value is greater than 0.1 and less than or equal to 1.2, it is high speed. M The movement speeds of individual behaviors are divided into three clusters with different speed magnitudes, and we get N Intra M The motion information of M behavior individuals in 16 frames is obtained, which is recorded as m , m =( m 1, m 2, m 3), m Indicates that M Individual behavior in N The set of visible key points in the frame, where the speed of the human body in N frames is obtained by calibrating the camera, establishing the mapping relationship between the image pixel coordinates and the world coordinates of the corresponding points in space, obtaining the actual displacement of the target human body in N frames, calculating the human body speed, and classifying the human body targets according to the speed value range. The human key points of different speed magnitudes after classification are given by formula (2): (2) in, C x , C y are the upper left and lower right coordinate points of the image position box where the fighting behavior occurs, marked by the target detection box. W , H are the width and height of the coordinate frame respectively. box conf , class conf are the confidence values ​​of the target detection box and the detection category respectively, , Respectively n The horizontal and vertical coordinate values ​​of key points of the individual body, Indicates n Confidence value of human key points. The human key point extraction model YOLOPose is used to detect 17 joints of the human body and output the coordinate position and confidence of each joint point.

[0024] Step S05: construct a social information feature extraction network to extract the temporal and spatial characteristics of the human body key point movement speed changes and individual changes from the obtained skeletal movement information. Figure 1 As shown in FIG. 1 , the social information feature extraction network is composed of a spatial convolution module and a temporal convolution module. The spatial feature output by the spatial convolution module is V spatial , which represents the spatial features output by the spatial convolution module; the temporal features output by the temporal convolution module are V temporal, represents the temporal features output by the temporal convolution module, by V spatial and V temporal Convolution fusion obtains spatiotemporal features V st .

[0025] Specifically, in step S05, the spatiotemporal features V st According to formula (3)-(5), (3) (4) (5) in, V spatial is the spatial feature output by the spatial convolution module, V temporal represents the temporal features output by the temporal convolution module, V st for V spatial and V temporal The spatiotemporal features after convolution fusion, Conv 2D is 2D convolution, the convolution kernel size is 3, the padding size is 1, and RELU is the activation function in the artificial neural network.

[0026] Step S06: Set up a social information quantification module to quantify the spatiotemporal features in step S05 into scores; the social information quantification module uses a fully connected network based on a logistic regression algorithm to train the weights of the fully connected network. V st Mapping to score , is the social information score of different speed magnitude sets. In this step, the weights of the fully connected network are trained to achieve inference prediction of the social score, where the dimension of the fully connected network is set to (2048, 1).

[0027] Specifically, in step S06, the loss function of the social information quantification module is formula (6); (6) The loss calculation method uses the mean square error loss function to measure the average difference between the model's predicted value and the true value. S p represents the prediction score, S t represents the true score, m Indicates the number of samples.

[0028] Step S07: Use the pseudo 3D video understanding module with the ActionNet network as the basic architecture to extract the action categories of the continuous image frames. Specifically, in step S07, the pseudo 3D video understanding module is inserted into the ResNet-50 backbone network to form the basic architecture of the video understanding module. The schematic diagram of the inserted pseudo 3D video understanding module is as follows: Figure 2 As shown in Figure 2, the basic architecture of the video understanding module is as follows: Figure 3 shown.

[0029] The pseudo 3D video understanding module is composed of a spatiotemporal excitation module, a channel excitation module and a motion excitation module. The three modules are connected in parallel to form the backbone network of the video understanding network. After the video understanding training, the video-level behavior recognition task is completed to obtain the predicted behavior category.

[0030] Step S08, embedding the social information quantification module in step S06 into the pseudo 3D video understanding module in step S07 to form a fighting behavior recognition model based on an implicit social model, and training the fighting behavior recognition model based on an implicit social model in the form of multi-loss feedback; the network training process is realized by feedback of the loss function and updating of the model parameters.

[0031] Specifically, in step S08, the video classification loss function of the pseudo 3D video understanding module uses FocalLoss, that is, formula (7), (7) in, is the loss value of the video classification loss function, is the weight parameter of easy-to-classify samples, is a hyperparameter. For positive samples, equal , for negative samples, equal ,in, and The values ​​of the two hyperparameter variables are set to 0.25 and 2.0 respectively; According to formula (8), (8) in, The probability that the model predicts the sample category to be "fighting" is in the range of , is the probability that the sample class is “normal”; is the true category value of the predicted category.

[0032] Step S09: using the fighting behavior recognition model based on the implicit social model trained in step S08 to determine whether there is fighting behavior in the video frame to be detected.

[0033] Specifically, in step S09, the method for determining whether there is a fighting behavior in the video frame to be detected is that the video classification module outputs the "fighting" category, and a target box of the "fighting" category appears in a single frame of video, and the target box, i.e., the category prediction probability value, is greater than the threshold set by the model reasoning, which is set to 0.5 and 0.65 respectively during reasoning. The video frame that meets the above conditions is judged as having a "fighting" behavior.

[0034] Specifically, the training loss function of the fighting behavior recognition model based on the implicit social model is formula (9), (9) in, , are the loss weights for classification and regression tasks, respectively, which are set to 0.6 and 0.4. Their values ​​can also be flexibly adjusted according to the training conditions of the two models. , The loss function of the video classification and the loss function of the social information score prediction module are respectively. The basic architecture diagram of the fighting behavior recognition method model based on the implicit social model of the present invention is as follows Figure 4 shown.

[0035] The method for determining whether there is a fighting behavior in the video frame to be detected is as follows: the video classification module outputs the "fighting" category, and a target box of the "fighting" category appears in a single frame of video, and the target box, that is, the category prediction probability value, is greater than the threshold set by the model inference, which is set to 0.5 and 0.65 respectively during inference. Video frames that meet the above conditions are judged to have "fighting" behavior.

[0036] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. Alterations, modifications, substitutions and variations of the above embodiments by a person skilled in the art are all within the scope of the present invention.

Claims

1. A fighting behavior recognition method based on implicit social model, characterized by: The steps include: Step S01: Extracting continuous image frames from the surveillance video stream F t ( t =1,2,..., f );in, f is the number of consecutive image frames; Step S02, using a fighting image dataset with action candidate box coordinates and category information labeled "fighting" and "non-fighting" to train an action target detector using YOLOv7 as a target detection model; Step S03: image frame F t The image is pre-processed for image enhancement and then sent to the human target key point detection model using the YOLOPose human key point extraction model to obtain the coordinates of the human skeleton points; at the same time, the image frame F t Send it to the action target detector in step S02 to obtain the image frame F t The target detection result of the fighting behavior in the target detection frame includes the image position information of the fighting behavior and the fighting behavior category, wherein the behavior category includes fighting and normal; Step S04: Combine the target detection frame bounding box coordinates of the fighting action in step S03, perform preprocessing operations on the extracted human key point information, and set a fixed calculation frame length of N , by the visible key points K n ( n =1,2,...,17), calculate the fixed frame length according to formula (1) M The speed of movement of individual v i ( i =1,2,..., M ), (1) in, for N Intraframe n The pixel displacement set of each key point of the human body, is the displacement length of the key point with the largest displacement in the set; the displacement speeds of the key points of the human body are defined at three speed levels: low speed, medium speed, and high speed. M The movement speeds of individual behaviors are divided into three clusters with different speed magnitudes, and we get N Intra M The motion information of each behavioral individual is recorded as m , m =( m 1, m 2, m 3), m Indicates that M Individual behavior in N The set of visible key points in the frame; Step S05: construct a social information feature extraction network, which is composed of a spatial convolution module and a temporal convolution module. The spatial feature output by the spatial convolution module is V spatial , the temporal features output by the temporal convolution module are V temporal , by V spatial and V temporal Convolution fusion obtains spatiotemporal features V st ; Step S06: Set up a social information quantification module to quantify the spatiotemporal features in step S05 into scores; the social information quantification module uses a fully connected network based on a logistic regression algorithm to train the weights of the fully connected network. V st Mapping to score , is the social information score for different speed magnitude sets; Step S07, using a pseudo 3D video understanding module with the ActionNet network as the basic architecture to extract the action category of continuous image frames, the pseudo 3D video understanding module is composed of a spatiotemporal excitation module, a channel excitation module and a motion excitation module, and the three modules are connected in parallel to form the backbone network of the video understanding network. After the video understanding training, the video level behavior recognition task is completed to obtain the predicted behavior category; Step S08, embedding the social information quantification module in step S06 into the pseudo 3D video understanding module in step S07 to form a fighting behavior recognition model based on an implicit social model, and training the fighting behavior recognition model based on the implicit social model in a multi-loss feedback form; Step S09: using the fighting behavior recognition model based on the implicit social model trained in step S08 to determine whether there is fighting behavior in the video frame to be detected.

2. A fighting behavior identification method based on implicit social model according to claim 1, characterized in that: In step S02, the fighting image dataset consists of RGB images and skeleton images, and the fighting image dataset includes fighting behavior images captured from the public datasets UBI-Fights and Hockey-Fights datasets, as well as collected indoor fighting behavior images.

3. A fighting behavior identification method based on implicit social model according to claim 1, characterized in that: In step S03, the image frame F t The image enhancement preprocessing includes random cropping, flipping, and random color perturbation operations on the image.

4. The method for identifying fighting behavior based on an implicit social model according to claim 1, characterized in that: In step S04, the speed threshold set [0, 0.01, 0.1, 1.2] is used to set the current N Intra M The movement speed of the human body is divided into three clusters with different speed magnitudes, among which the human body N The intra-frame motion speed is obtained by calibrating the camera, establishing the mapping relationship between the image pixel coordinates and the world coordinates of the corresponding points in space, and obtaining the target human body's N The actual displacement within the frame time is used to calculate the human body movement speed, and the human body targets are classified according to the speed value interval. The key points of the human body with different speed magnitudes after classification are given by formula (2): （2） in, C x , C y are the upper left and lower right coordinate points of the image position box where the fighting behavior occurs, marked by the target detection box. W , H are the width and height of the coordinate frame respectively. box conf , class conf are the confidence values ​​of the target detection box and the detection category respectively, , Respectively n The horizontal and vertical coordinate values ​​of key points of the individual body, Indicates n Confidence values ​​of individual body key points.

5. The method for identifying fighting behavior based on an implicit social model according to claim 1, characterized in that: In step S05, the spatiotemporal features V st According to formula (3)-(5), （3） （4） （5） in, V spatial is the spatial feature output by the spatial convolution module, V temporal represents the temporal features output by the temporal convolution module, V st for V spatial and V temporal The spatiotemporal features after convolution fusion, Conv 2D is 2D convolution, and RELU is the activation function in artificial neural networks.

6. The method for identifying fighting behavior based on an implicit social model according to claim 1, characterized in that: In step S06, the loss function of the social information quantification module is formula (6); （6） in, S p represents the prediction score, S t represents the true score, m Indicates the number of samples.

7. The method for identifying fighting behavior based on an implicit social model according to claim 1, characterized in that: In step S07, the pseudo 3D video understanding module is inserted into the ResNet-50 backbone network to form the basic architecture of the video understanding module.

8. The method for identifying fighting behavior based on an implicit social model according to claim 1, characterized in that: In step S08, the video classification loss function of the pseudo 3D video understanding module uses Focal Loss, that is, formula (7): （7） in, is the loss value of the video classification loss function, is the weight parameter of easy-to-classify samples, is a hyperparameter. For positive samples, equal , for negative samples, equal , According to formula (8), （8） in, The probability that the model predicts the sample category to be "fighting" is, is the true category value of the predicted category.

9. The method for identifying fighting behavior based on an implicit social model according to claim 1, characterized in that: In step S09, the method for determining whether there is a fighting behavior in the video frame to be detected is that the video classification module outputs the "fighting" category, and a target box of the "fighting" category appears in a single frame of video, and the target box, i.e., the category prediction probability value, is greater than the threshold set by the model reasoning, which is set to 0.5 and 0.65 respectively during reasoning. Video frames that meet the above conditions are judged to have a "fighting" behavior.

10. The method for identifying fighting behavior based on an implicit social model according to claim 1, characterized in that: The training loss function of the fighting behavior recognition model based on the implicit social model is formula (9), （9） in, , are the loss weights for classification and regression tasks respectively, , They are the video classification loss function and the social information score prediction module loss function respectively.

Images