Methods, devices and electronic equipment for identifying fighting behavior

Abstract

This invention discloses a method, apparatus, and electronic device for identifying fighting behavior. The invention acquires image frames from a surveillance video stream, inputs these frames into a pedestrian detection model for detection, obtains pedestrian detection results, crops the results to a fixed size, normalizes them to obtain image features, inputs these features into a human keypoint detection model for detection, obtains multiple human keypoints, and obtains a fighting category probability P1 through the classification branch of the human keypoint detection model. Based on the distribution of human keypoints, a keypoint logical score is calculated. The pedestrian detection results are stacked in chronological order into a fixed-length feature, input into a temporal classification model for detection, and obtain a fighting category probability P2. Based on the keypoint logical score, the fighting category probability P1, and the fighting category probability P2, a fighting score is calculated. An alarm for fighting behavior is triggered based on the fighting score, enabling accurate determination of whether a fight has occurred.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, and electronic device for recognizing fighting behavior. Background Technology

[0002] Existing security monitoring systems typically report anomalies through real-time feedback from monitoring personnel. However, this manual monitoring method suffers from drawbacks such as unreliability and high manpower consumption. In rail transit scenarios, where passenger flow is high, the failure to promptly alert authorities to fights can lead to congestion and even stampedes. Therefore, a timely alarm method for detecting fights can provide immediate response and crucial evidence for maintaining public order.

[0003] Patent CN113111733A discloses a fight behavior recognition method based on pose flow. This method first uses a deep learning-based human keypoint detection algorithm to obtain a sequence of human keypoints, then combines this with traditional image tracking algorithms to track the human body. By logically analyzing the human pose queue in consecutive frames of a video stream, it determines whether a fight has occurred. This pose flow-based fight behavior recognition method can avoid false alarms or missed alarms caused by false detections in single-frame images. However, pedestrians are non-rigid objects, and their poses change diversely and have high degrees of freedom during a fight. There are often non-ideal situations such as partial occlusion and blurring during a fight. If a direct keypoint model is used, there will be many false detections of keypoints, affecting the accuracy of fight behavior recognition. Summary of the Invention

[0004] The purpose of this invention is to provide a method, device, and electronic device for identifying fighting behavior, in order to solve the problem that the existing technology uses a direct key point model, which has a large number of false key point detections, affecting the accuracy of fighting behavior identification.

[0005] In a first aspect, the present invention provides a method for identifying fighting behavior, comprising:

[0006] Acquire image frames from the surveillance video stream;

[0007] The image frame is input into the pedestrian detection model for detection to obtain pedestrian detection results;

[0008] The pedestrian detection results are cropped to a fixed size and normalized to obtain image features;

[0009] The image features are input into the human key point detection model for detection to obtain multiple human key points, and the probability P1 of the fighting category is obtained through the classification branch of the human key point detection model.

[0010] Calculate the key point logic score based on the distribution of the key points on the human body;

[0011] The pedestrian detection results are stacked into a fixed-length feature in chronological order and input into a temporal classification model for detection to obtain the probability P2 of the fighting category.

[0012] Calculate the fighting score based on the key point logic score, the fighting category probability P1, and the fighting category probability P2.

[0013] An alarm will be triggered based on the fight score.

[0014] Further, the image frame is input into a pedestrian detection model for detection to obtain pedestrian detection results, including:

[0015] The image frame is input into the pedestrian detection model to obtain the detection box of each pedestrian and the corresponding feature vector of the entire image.

[0016] The center point of the detection box is calculated based on the pedestrian detection box.

[0017] Furthermore, the image features are input into a human keypoint detection model for detection to obtain multiple human keypoints. The probability P1 of the fighting category is then obtained through the classification branch of the human keypoint detection model, including:

[0018] The image features are input into the human key point detection model for detection to obtain multiple human key points;

[0019] Based on the image features and human key points, the classification branch of the human key point detection model is used to distinguish whether the human body is in a fighting posture, and the fighting category probability P1 is obtained.

[0020] Furthermore, after inputting the image features into a human keypoint detection model for detection and obtaining multiple human keypoints, the method further includes:

[0021] Unnecessary human body key points are removed in the fight behavior recognition.

[0022] Furthermore, based on the distribution of the key points on the human body, a key point logical score is calculated, including:

[0023] Based on the distribution of the key points of the human body, calculate the angle θ1 between the upper arm and forearm, the angle θ2 between the upper arm and body, and the angle θ3 between the thigh.

[0024] Calculate the key point logic score using the following formula:

[0025] f(x)=1 / 180×θ1+1 / 90×θ2+1 / 180×θ3;

[0026] In the formula, f(x) is the key point logic score; θ1 is the angle between the left and right upper arms and forearms, θ1 is 0°~180°; θ2 is the angle between the left and right upper arms and the body, θ2 is 0°~180°; θ3 is the angle between the left and right thighs, θ3 is 0°~180°.

[0027] Further, based on the key point logic score, the fight category probability P1, and the fight category probability P2, a fight score is calculated, including:

[0028] The score for the fight is calculated based on h(x) = p1 × f(x) + p2;

[0029] In the formula, h(x) is the fighting score, P1 is the fighting category probability obtained by the human key point detection model, and P2 is the fighting category probability obtained by the temporal classification model.

[0030] Furthermore, based on the aforementioned fight score, an alarm is triggered to report the fight, including:

[0031] Determine whether the fighting score is greater than or equal to a preset threshold;

[0032] If the fighting score is less than a preset threshold, it is determined that there is no fighting behavior, and the fighting behavior recognition of the image frame continues.

[0033] If the fighting score is greater than or equal to a preset threshold, it is determined that a fighting has occurred.

[0034] Generate an alarm and return the current frame image.

[0035] Furthermore, if the fighting score is greater than or equal to a preset threshold, and a fighting behavior is determined to have occurred, the method further includes:

[0036] Determine whether the duration of the fight meets the time requirement T1;

[0037] If the duration of the fight meets the time requirement T1, increment the cumulative fight value by 1.

[0038] Determine whether the cumulative fighting value is greater than or equal to a set threshold;

[0039] If the cumulative fighting value is greater than or equal to a set threshold, an alarm is generated and the current frame image is returned;

[0040] If the cumulative fighting value is less than a set threshold, continue to identify fighting behavior in the image frame.

[0041] In a second aspect, the present invention provides a fighting behavior recognition device, comprising:

[0042] The acquisition unit is used to acquire image frames from the monitoring video stream.

[0043] The pedestrian detection unit is used to input the image frame into the pedestrian detection model for detection and obtain pedestrian detection results.

[0044] The cropping unit is used to crop the pedestrian detection results to a fixed size, normalize them, and then obtain image features.

[0045] The key point detection unit is used to input the image features into the human key point detection model for detection, obtain multiple human key points, and obtain the fighting category probability P1 through the classification branch of the human key point detection model.

[0046] The key point logic score calculation unit is used to calculate the key point logic score based on the distribution of the key points of the human body.

[0047] The temporal classification unit is used to stack the pedestrian detection results into a fixed-length feature in chronological order, input the temporal classification model for detection, and obtain the probability P2 of the fighting category.

[0048] The fight score calculation unit is used to calculate the fight score based on the key point logic score, the fight category probability P1, and the fight category probability P2.

[0049] An alarm unit is used to issue an alarm for fighting behavior based on the fighting score.

[0050] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0051] The beneficial effects of this invention: The fighting behavior recognition method, device, and electronic device provided by this invention acquire image frames from a surveillance video stream, input the image frames into a pedestrian detection model for detection, obtain pedestrian detection results, crop the pedestrian detection results to a fixed size, normalize them to obtain image features, input the image features into a human keypoint detection model for detection, obtain multiple human keypoints, and obtain a fighting category probability P1 through the classification branch of the human keypoint detection model. Based on the distribution of the human keypoints, a keypoint logical score is calculated. The pedestrian detection results are stacked into a fixed-length feature in chronological order and input into a temporal classification model for detection to obtain a fighting category probability P2. Based on the keypoint logical score, the fighting category probability P1, and the fighting category probability P2, a fighting score is calculated. The fighting score is then used to further... This invention provides an alarm for fighting behavior. It employs a pedestrian detection model and a human keypoint detection model, which can effectively locate keypoint information even in non-ideal situations such as partial occlusion or blurring during a fight. A temporal classification model is used, which has excellent extraction performance for time-series features. Furthermore, the temporal classification model fully utilizes existing features extracted by the pedestrian detection model, improving model utilization, classification accuracy, and reducing complexity. This invention proposes a multi-result adaptive weighting scheme, adaptively combining the keypoint fighting judgment result and the temporal classification result. This reduces the unreasonableness caused by manually set parameters and effectively improves the results. This invention combines local features of human keypoints and global features of temporal classification to determine whether a fight has occurred, fully considering both local and global information, and can accurately determine whether a fight has occurred. Attached Figure Description

[0052] To further understand the features and technical content of this invention, please refer to the following detailed description and accompanying drawings. However, the drawings are provided for reference and illustration only and are not intended to limit the invention.

[0053] Attached Figure

[0054] Figure 1 This is a flowchart of the fighting behavior recognition method of the present invention;

[0055] Figure 2 Here is the flowchart for S102;

[0056] Figure 3 Here is the flowchart for S104;

[0057] Figure 4 This is a flowchart of an optional implementation of S108;

[0058] Figure 5 This is a flowchart of another optional implementation of S108;

[0059] Figure 6 A diagram showing the locations of 12 key points on the human body as defined;

[0060] Figure 7 Schematic diagram for example 1 of fighting action recognition;

[0061] Figure 8 Schematic diagram for example 2 of fighting action recognition;

[0062] Figure 9 This is a block diagram of the fighting behavior recognition device of the present invention;

[0063] Figure 10 This is a block diagram of the electronic device of the present invention. Implementation

[0064] To further illustrate the technical means and effects of the present invention, the following detailed description is provided in conjunction with the preferred embodiments of the present invention and their accompanying drawings.

[0065] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0067] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to implement and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be implemented without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0068] Please see Figure 1 This invention provides a method for identifying fighting behavior, comprising:

[0069] S101, acquire image frames from the monitoring video stream.

[0070] S102, the image frame is input into the pedestrian detection model for detection to obtain the pedestrian detection result.

[0071] Please see Figure 2 S102 may specifically include:

[0072] S1021, The image frame is input into the pedestrian detection model to obtain the detection box of each pedestrian and the corresponding feature vector of the entire screen.

[0073] S1022, Calculate the center point of the detection frame based on the pedestrian detection frame.

[0074] Specifically, the pedestrian detection model can use a deep neural network detection model (YOLOv5).

[0075] S103, the pedestrian detection results are cropped to a fixed size and normalized to obtain image features.

[0076] Generally, fighting involves large movements of the hands and feet, making it difficult for pedestrian detection models to include the entire person within the detection frame, especially for actions such as reaching out or raising the leg at a large angle. Therefore, the pedestrian detection results are cropped to a fixed size and normalized before being input into the human body key point detection model in subsequent steps.

[0077] S104, the image features are input into the human key point detection model for detection to obtain multiple human key points, and the probability P1 of the fighting category is obtained through the classification branch of the human key point detection model.

[0078] Please see Figure 3 S104 may specifically include:

[0079] S1041, The image features are input into the human key point detection model for detection to obtain multiple human key points.

[0080] The output of the Human Keypoint Detection Model (HRnet) is 12 keypoints, such as... Figure 6 As shown.

[0081] S1042, remove unnecessary human body key points in the fight behavior recognition.

[0082] Remove key human body points that are not very useful for judging fights to reduce the requirements on model capabilities and the amount of labeled data.

[0083] S1043, Based on the image features and human body key points, the classification branch of the human body key point detection model is used to distinguish whether the human body is in a fighting posture, and the fighting category probability P1 is obtained.

[0084] To reduce computational load, the original model structure was modified, reducing some intermediate layers and effectively improving speed. Simultaneously, the keypoint model was fully utilized to simplify subsequent keypoint judgment logic. A classification branch was added to the original model to distinguish whether a person is in a fighting posture. This branch uses keypoint information to assist classification, improving accuracy without excessively increasing inference time. This solution employs the YOLOv5 detection model and the HRNet keypoint detection model, demonstrating good prediction performance for occlusion and simplifying the judgment logic.

[0085] S105, Calculate the key point logic score based on the distribution of the key points of the human body.

[0086] Specifically, based on the distribution of the key points of the human body, the angles θ1 between the upper and lower arms and forearms, θ2 between the upper and lower arms and the body, and θ3 between the left and right thighs are calculated.

[0087] The fighting behavior in this invention only involves punching and kicking. During a fight, the elbow mainly has two states: bent and straight. The state of the arm is determined by combining θ1 and θ2. Generally, during a fight, the arm alternates between straight and bent; changes in angle can be used to determine if a fight is taking place. When a fighter raises their leg significantly, θ3 will be greater than the normal angle.

[0088] Calculate the key point logic score using the following formula:

[0089] f(x)=1 / 180×θ1+1 / 90×θ2+1 / 180×θ3;

[0090] In the formula, f(x) is the key point logic score; θ1 is the angle between the left and right upper arms and forearms, θ1 is 0°~180°; θ2 is the angle between the left and right upper arms and the body, θ2 is 0°~180°; θ3 is the angle between the left and right thighs, θ3 is 0°~180°.

[0091] S106, the pedestrian detection results are stacked into a fixed-length feature in chronological order, and input into the temporal classification model for detection to obtain the probability P2 of the fighting category.

[0092] Fighting is a continuous action within a time period, and its recognition requires careful consideration of the changes in action over that time. Pedestrian detection models have excellent feature extraction capabilities, and to reduce the difficulty of feature extraction in temporal classification models, the features extracted by pedestrian detection models can be fully utilized. The features extracted by pedestrian detection models contain features at multiple scales, including both local and global features. These features are stacked in chronological order into a fixed-length feature set and then input into a temporal classification model (Bi-LSTM). Because the temporal classification model utilizes the features extracted by pedestrian detection models, its accuracy is higher than that of directly using original image features. Furthermore, the temporal classification model effectively captures temporal features, avoiding misclassifications of actions similar to fighting. Fighting actions generally lack a clear pattern over a period, while actions like dancing exhibit certain patterns. This approach uses the Bi-LSTM temporal classification model, which demonstrates excellent performance in extracting time-series features.

[0093] S107, Calculate the fighting score based on the key point logic score, the fighting category probability P1, and the fighting category probability P2.

[0094] Specifically, the score for the fight is calculated based on h(x) = p1 × f(x) + p2;

[0095] In the formula, h(x) is the fighting score, P1 is the fighting category probability obtained by the human key point detection model, and P2 is the fighting category probability obtained by the temporal classification model.

[0096] This invention proposes the above-mentioned multi-result adaptive weighting scheme, which adaptively combines the key point conflict judgment result and the time series classification result, reducing the unreasonableness caused by manually set parameters and effectively improving the results.

[0097] S108, trigger an alarm for fighting behavior based on the fighting score.

[0098] After the preceding steps, the fight score can be used to estimate whether fighting has occurred within a certain time interval, and then decide whether to issue an alarm. Please refer to [link / reference]. Figure 7 and Figure 8 , Figure 7 (a) Figure 7 (b) Figure 7 (c) are images of the arm bending, reaching out to strike, and alarm from the fighting action recognition example 1. Figure 8 (a) Figure 8 (b) Figure 8 (c) are images of the arm bending, reaching out to strike, and alarm from the fighting action recognition example 2.

[0099] Please see Figure 4 S108 may specifically include:

[0100] S1081, determine whether the fighting score is greater than or equal to a preset threshold.

[0101] S1082, if the fighting score is less than a preset threshold, it is determined that there is no fighting behavior, and the fighting behavior recognition of the image frame continues.

[0102] S1083, if the fighting score is greater than or equal to a preset threshold, it is determined that a fighting behavior has occurred.

[0103] S1084 generates an alarm and returns the current frame image.

[0104] Please see Figure 5 As a preferred embodiment, after S1083, the following steps may also be included:

[0105] S1085, determine whether the duration of the fighting behavior meets the time requirement T1.

[0106] S1086, If the duration of the fight meets the time requirement T1, increment the cumulative value of the fight by 1.

[0107] If the duration of the fight does not meet the time requirement T1, then S1085 is executed again to continue timing the duration of the fight.

[0108] S1087, determine whether the cumulative fighting value is greater than or equal to a set threshold.

[0109] If the cumulative fighting value is greater than or equal to the set threshold, then execute S1084, generate an alarm and return the current frame image.

[0110] S1088, if the cumulative fighting value is less than a set threshold, continue to identify fighting behavior in the image frame.

[0111] As can be seen from the above embodiments, the fighting behavior recognition method provided by the present invention acquires image frames from a monitoring video stream, inputs the image frames into a pedestrian detection model for detection, obtains pedestrian detection results, crops the pedestrian detection results to a fixed size, normalizes them to obtain image features, and inputs the image features into a human key point detection model for detection to obtain multiple human key points. The present invention uses a pedestrian detection model and a human key point detection model, which can also locate key point information well even in non-ideal situations such as partial occlusion and blurring during the fighting process.

[0112] This invention obtains the fighting category probability P1 through the classification branch of the human keypoint detection model. Based on the distribution of the human keypoints, the keypoint logical score is calculated. The pedestrian detection results are stacked into a fixed-length feature in chronological order and input into a temporal classification model for detection to obtain the fighting category probability P2. This invention uses a temporal classification model, which has a good extraction effect on time series features. Furthermore, the temporal classification model makes full use of the existing features extracted by the pedestrian detection model, which can improve the utilization of the model, improve the classification accuracy of the temporal model, and reduce the complexity of the temporal classification model.

[0113] This invention calculates a fighting score based on the key point logical score, the fighting category probability P1, and the fighting category probability P2, and then issues a fighting behavior alarm based on the fighting score. This invention proposes a multi-result adaptive weighting scheme, adaptively combining the key point fighting judgment result and the temporal classification result. This reduces the unreasonableness caused by manually set parameters and effectively improves the results. This invention combines local features of human key points and global features of temporal classification to determine whether fighting behavior exists, fully considering both local and global information, and can accurately determine whether fighting behavior exists.

[0114] Please see Figure 9 The present invention provides a fighting behavior recognition device, comprising:

[0115] Acquisition unit 11 is used to acquire image frames from the monitoring video stream;

[0116] Pedestrian detection unit 12 is used to input the image frame into the pedestrian detection model for detection and obtain pedestrian detection results;

[0117] The cropping unit 13 is used to crop the pedestrian detection results to a fixed size, normalize them, and then obtain image features.

[0118] The key point detection unit 14 is used to input the image features into the human key point detection model for detection, obtain multiple human key points, and obtain the fighting category probability P1 through the classification branch of the human key point detection model.

[0119] The key point logic score calculation unit 15 is used to calculate the key point logic score based on the distribution of the key points of the human body.

[0120] The temporal classification unit 16 is used to stack the pedestrian detection results into a fixed-length feature in chronological order, input the temporal classification model for detection, and obtain the fighting category probability P2.

[0121] The fight score calculation unit 17 is used to calculate the fight score based on the key point logic score, the fight category probability P1, and the fight category probability P2.

[0122] Alarm unit 18 is used to issue an alarm for fighting behavior based on the fighting score.

[0123] Please see Figure 10 It includes: a memory 100 and a processor 200, wherein the memory 100 stores a computer program, and when the computer program is executed by the processor 200, the processor 200 performs the steps of the method described above.

[0124] This invention also provides a storage medium storing a computer program. When executed by a processor, the computer program implements some or all of the steps in the various embodiments of the fighting behavior recognition method provided by this invention. The storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0125] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0126] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the description of the fighting behavior recognition device embodiment is relatively simple since it is basically similar to the method embodiment, and the relevant parts can be referred to the description in the method embodiment.

[0127] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention.

Claims

1. A method for identifying fighting behavior, characterized by, The method comprises the following steps: obtaining an image frame of a monitoring video stream; inputting the image frame into a pedestrian detection model for detection to obtain a pedestrian detection result; cropping the pedestrian detection result to a fixed size, and obtaining an image feature after normalization; inputting the image feature into a human key point detection model for detection to obtain multiple human key points, and obtaining a fighting category probability P1 through a classification branch of the human key point detection model; calculating a key point logic score according to the distribution of the human key points, wherein the key point logic score is calculated according to an included angle θ1 between left and right upper arms and forearms, an included angle θ2 between left and right upper arms and bodies, and an included angle θ3 between left and right thighs; calculating the key point logic score according to the following formula: ； In the formula, is the key point logic score; θ1 is the angle between the upper arm and forearm of the left and right hands, θ1 is 0°~180°; θ2 is the angle between the upper arm and body of the left and right hands, θ2 is 0°~180°; θ3 is the angle between the left and right thighs, θ3 is 0°~180°; stacking the pedestrian detection results in chronological order into a fixed length feature, and inputting the fixed length feature into a time sequence classification model for detection to obtain a fighting category probability P2; According to the key point logic score, the fighting class probability P1, and the fighting class probability P2, a fighting score is calculated; wherein, according to a fighting score is calculated; In the formula, P1 is a fighting category probability obtained by the human key point detection model, and P2 is a fighting category probability obtained by the time series classification model. determining whether the fighting score is greater than or equal to a preset threshold value; if the fighting score is less than the preset threshold value, it is determined that there is no fighting behavior, and the image frame continues to be recognized for fighting behavior; if the fighting score is greater than or equal to the preset threshold value, it is determined that there is a fighting behavior; determining whether the duration of the fighting behavior meets a time requirement T1; if the duration of the fighting behavior meets the time requirement T1, the fighting cumulative value is increased by 1; determining whether the fighting cumulative value is greater than or equal to a set threshold value; if the fighting cumulative value is greater than or equal to the set threshold value, an alarm is generated and the current frame picture is returned; if the fighting cumulative value is less than the set threshold value, the image frame continues to be recognized for fighting behavior.

2. The fighting behavior recognition method according to claim 1, wherein The method comprises the following steps: inputting the image frame into a pedestrian detection model to obtain a pedestrian detection result, comprising: inputting the image frame into the pedestrian detection model to obtain a detection frame of each pedestrian and a corresponding whole picture feature vector; 3. The fighting behavior recognition method according to claim 1, wherein calculating a center point of the detection frame according to the detection frame. The method comprises the following steps: inputting the image feature into a human key point detection model for detection to obtain multiple human key points, and obtaining a fighting category probability P1 through a classification branch of the human key point detection model, comprising:

4. The fighting behavior recognition method according to claim 3, characterized by, inputting the image feature into the human key point detection model for detection to obtain multiple human key points; distinguishing whether the human body is in a fighting posture by using the classification branch of the human key point detection model according to the image feature and the human key points to obtain the fighting category probability P1.

5. A fighting behavior recognition device characterized by comprising: After inputting the image feature into the human key point detection model for detection to obtain multiple human key points, the method further comprises: removing unnecessary human key points in fighting behavior recognition. The method comprises the following steps: an obtaining unit, configured to obtain an image frame of a monitoring video stream; a pedestrian detection unit, configured to input the image frame into a pedestrian detection model for detection to obtain a pedestrian detection result; a cropping unit, configured to crop the pedestrian detection result to a fixed size, and obtain an image feature after normalization; a key point detection unit, configured to input the image feature into a human key point detection model for detection to obtain multiple human key points, and obtain a fighting category probability P1 through a classification branch of the human key point detection model; A key point logic score calculation unit is configured to calculate a key point logic score according to the distribution of the human body key points; wherein, according to the distribution of the human body key points, an angle θ1 between the left and right upper arms and forearms, an angle θ2 between the left and right upper arms and the body, and an angle θ3 between the left and right thighs are calculated; The key point logic score is calculated according to the following formula: ； wherein is the keypoint logic score; θ1 is the angle between the upper arm and forearm of the left and right hands, θ1 is 0°~180°; θ2 is the angle between the upper arm and body of the left and right hands, θ2 is 0°~180°; θ3 is the angle between the left and right thighs, θ3 is 0°~180°; A time sequence classification unit is configured to stack the pedestrian detection results in time sequence into fixed length features, input a time sequence classification model for detection, and obtain a fighting category probability P2; The fighting score calculation unit is configured to calculate a fighting score according to the key point logical score, the fighting category probability P1, and the fighting category probability P2. The fighting score calculation unit is configured to calculate the fighting score according to the following formula: The fighting score calculation unit is configured to calculate a fighting score according to the key point logical score, the fighting category probability P1, and the fighting category probability P2. The fighting score calculation unit is configured to calculate the fighting score according to the following formula: In the formula, P1 is a fighting category probability obtained by the human key point detection model, and P2 is a fighting category probability obtained by the time series classification model. An alarm unit is configured to judge whether the fighting score is greater than or equal to a preset threshold; if the fighting score is less than the preset threshold, it is determined that there is no fighting behavior, and the image frame is continuously subjected to fighting behavior recognition; if the fighting score is greater than or equal to the preset threshold, it is determined that there is a fighting behavior; it is judged whether the fighting behavior duration meets a time requirement T1; if the fighting behavior duration meets the time requirement T1, the fighting cumulative value is added by 1; it is judged whether the fighting cumulative value is greater than or equal to a set threshold; if the fighting cumulative value is greater than or equal to the set threshold, an alarm is generated and the current frame picture is returned; if the fighting cumulative value is less than the set threshold, the image frame is continuously subjected to fighting behavior recognition.

6. An electronic device, comprising: Comprise: A memory and a processor, the memory stores a computer program, the computer program is executed by the processor, so that the processor executes the steps of the method as claimed in any one of claims 1-4.

Images