Method and system for extracting behavior features of monitoring video based on spatio-temporal graph convolution network

Abstract

The present application belongs to the technical field of computer vision and image data processing, and relates to a monitoring video behavior feature extraction method and system based on a space-time graph convolution network. The method comprises the following steps: first, extracting human body skeleton key points from the monitoring video and performing scale normalization based on the torso length; second, calculating a vertical control deviation value reflecting the degree of vertical gravity control loss of the human body and a trajectory oscillation entropy reflecting the degree of disorder of the motion trajectory; third, using the vertical control deviation value and the trajectory oscillation entropy to construct an adaptive weight function to generate a behavior anomaly weighting coefficient; and finally, weighting the to-be-identified skeleton features using the behavior anomaly weighting coefficient and inputting the space-time graph convolution network for identification. The present application can effectively distinguish behaviors that are similar in spatial geometric shape but different in physical nature, thereby reducing the false positive rate in industrial scenarios.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image data processing technology, specifically relating to a method and system for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent video surveillance systems have been widely applied in industrial safety production and public safety early warning. Among numerous action recognition technologies, the analysis method based on human skeletal key points has gradually become the mainstream technology in this field due to its strong robustness to changes in lighting and background complexity. Currently, action recognition based on skeletal key points mainly relies on spatiotemporal graph convolutional networks. The core process of this technology is to extract human joint coordinates through pose estimation algorithms, construct a spatiotemporal graph, and aggregate features in spatial and temporal dimensions using convolutional kernels, thereby achieving the classification and recognition of human actions.

[0003] While existing technologies perform adequately in general scenarios, they still face significant challenges in complex industrial safety monitoring applications. The main difficulty lies in the striking similarity in spatial geometry between high-risk abnormal behaviors and normal daily behaviors. For example, a worker's controlled squatting or bending motion during work is almost identical in its final static posture after a sudden fainting and uncontrolled fall. This high degree of visual confusion makes methods relying solely on spatial skeletal geometry ineffective in distinguishing between the two.

[0004] The fundamental reason for the aforementioned recognition difficulties lies in the fact that existing spatiotemporal graph convolutional network models distribute feature extraction weights relatively evenly across time series and primarily focus on changes in the geometric positions of key points, while neglecting the underlying physical and mechanical essence of the movements. From a kinematic perspective, falling is a gravity-driven, uncontrolled acceleration process, accompanied by drastic changes in vertical velocity and a disordered trajectory; while squatting is a controlled deceleration process where muscles counteract gravity, exhibiting clear motion logic and regularity. Current technologies lack explicit modeling and analysis of this force state and motion control logic, leading to a high likelihood of false alarms when dealing with behaviors that appear similar but are fundamentally different, severely impacting the practicality and reliability of industrial monitoring systems. Summary of the Invention

[0005] The purpose of this invention is to propose a method and system for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks, in order to solve the technical problem that existing behavior recognition technologies based on spatiotemporal graph convolutional networks cannot effectively distinguish behaviors with similar spatial geometric shapes but different physical essences because they ignore the physical kinematics.

[0006] To address the above problems, the technical solution of the surveillance video behavior feature extraction method based on spatiotemporal graph convolutional networks proposed in this invention is as follows:

[0007] A method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks includes the following steps:

[0008] Human skeleton key points are extracted from the surveillance video stream, and a normalized coordinate system is constructed based on the human torso length benchmark to convert the original pixel coordinates of the human skeleton key points into normalized coordinates.

[0009] The difference between the acceleration of each joint in the human body in the vertical direction and the preset standard gravity reference value is calculated based on the normalized coordinates, and the vertical runaway deviation value reflecting the degree of gravity runaway is determined by combining the detection confidence level.

[0010] The velocity vectors of each joint of the human body are calculated based on the normalized coordinates. Based on the directional changes and velocity magnitudes of the velocity vectors at adjacent time points, the trajectory oscillation entropy, which reflects the degree of motion disorder, is determined.

[0011] An adaptive weighting function is constructed using the trajectory oscillation entropy and the vertical runaway deviation value. The abnormal behavior weighting coefficient is calculated and then fused with the skeleton feature matrix to be identified, which is used as the input of the spatiotemporal graph convolutional network for behavior recognition.

[0012] Furthermore, the human torso length reference is determined by calculating the Euclidean distance between the cervical joint and the center point of the pelvis in the same frame.

[0013] Furthermore, the process of converting the original pixel coordinates of the key points of the human skeleton into normalized coordinates specifically includes:

[0014] Using the center point of the pelvis in each frame as the origin of the coordinate system, the original x-coordinate and original y-coordinate of the center point of the pelvis are subtracted from the original x-coordinate and original y-coordinate of the key points of the human skeleton, respectively. The results of the subtraction are then divided by the length reference of the human torso to obtain the normalized x-coordinate and normalized y-coordinate of the key points of the human skeleton.

[0015] Furthermore, the formula for calculating the vertical runaway deviation value is as follows:

[0016]

[0017] In the formula, Indicates the first Frame number Vertical runaway deviation value of each joint Indicates the first Frame number Normalized vertical acceleration of each joint This represents the preset normalized standard gravity reference value. Indicates the first Frame number Detection confidence of each joint It is the first minimum constant. It is a natural constant.

[0018] Furthermore, the normalized vertical acceleration is obtained through second-order difference operations, specifically including:

[0019] The normalized vertical acceleration is obtained by performing a second-order difference operation using the normalized ordinates of the previous frame, the current frame, and the next frame, and then dividing the result by the square of the normalized time step.

[0020] Furthermore, the formula for calculating the trajectory oscillation entropy is:

[0021]

[0022] In the formula, Indicates the first Each joint in the time window The trajectory oscillation entropy within, and They represent the first The velocity components of the velocity vector in the horizontal and vertical directions at any given moment. Indicates the first The square of the magnitude of the velocity vector at time step [time value] It is the second minimum constant.

[0023] Furthermore, the velocity components of the velocity vector in both the horizontal and vertical directions are obtained through first-order difference operations, specifically including:

[0024] The velocity component in the horizontal direction is obtained by subtracting the normalized horizontal coordinate of the previous frame from the normalized horizontal coordinate of the current frame; the velocity component in the vertical direction is obtained by subtracting the normalized vertical coordinate of the previous frame from the normalized vertical coordinate of the current frame.

[0025] Furthermore, the formula for calculating the weighting coefficient of the behavioral anomaly is as follows:

[0026]

[0027] In the formula, Indicates the first Frame number The weighted coefficient for abnormal behavior of each joint. It is a regulating factor.

[0028] Further, the fusion of the behavioral anomaly weighting coefficients with the skeletal feature matrix to be identified includes:

[0029] The abnormal behavior weighting coefficients are multiplied element-wise with the skeletal feature matrix to be identified to generate a weighted skeletal feature matrix.

[0030] The technical solution of the surveillance video behavior feature extraction system based on spatiotemporal graph convolutional networks proposed in this invention is as follows:

[0031] The surveillance video behavior feature extraction system based on spatiotemporal graph convolutional networks includes a processor and a memory. The memory stores a computer program. When the computer program is run by the processor, it implements the surveillance video behavior feature extraction method based on spatiotemporal graph convolutional networks described in any of the above technical solutions.

[0032] The beneficial effects of this invention are as follows: This invention utilizes two core physical indicators, vertical runaway deviation value and trajectory oscillation entropy, to accurately capture the essential characteristics of gravity acceleration and disordered motion trajectory during abnormal falls, thereby realizing the distinction between controlled motion and runaway motion from the perspective of force state and motion control logic, and significantly reducing the false alarm rate in safety monitoring.

[0033] This invention, by designing an adaptive weighting function that incorporates detection confidence and velocity modulus, can automatically shield against the influence of low-quality data and environmental noise. During the calculation process, the algorithm uses detection confidence to correct vertical runaway deviation, effectively reducing the interference of keypoint drift caused by occlusion or blurring on the judgment results. Simultaneously, the calculation process of trajectory oscillation entropy can filter out minute sensor jitter in static states, ensuring that the system only focuses on feature changes with actual motion significance, thereby guaranteeing stable operation of the system in complex industrial environments.

[0034] This invention eliminates the influence of camera shooting distance, focal length differences, and human height differences on motion data by constructing a normalized coordinate system based on the human body's own scale. This makes the calculated physical indicators such as acceleration no longer limited to specific pixel units, thus ensuring that the preset physical threshold standards remain effective in different monitoring perspectives and scenarios, and has good versatility and universality.

[0035] This invention pre-calculates physical characteristics and generates anomaly weighting coefficients, enabling the neural network to focus its attention on key nodes that both conform to the laws of free fall and exhibit highly chaotic trajectory characteristics. This not only improves the convergence effect of the model but also retains lightweight computational characteristics, ultimately achieving accurate monitoring of high-risk abnormal behaviors in industry. Attached Figure Description

[0036] Figure 1 This is a flowchart of the steps of the surveillance video behavior feature extraction method based on spatiotemporal graph convolutional network provided in the embodiments of the present invention;

[0037] Figure 2 This is a schematic diagram of the trajectory oscillation entropy value change curves of different behavioral samples in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram comparing the weighted coefficients of behavioral anomalies for different behavioral samples in an embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0040] A specific embodiment of the surveillance video behavior feature extraction method based on spatiotemporal graph convolutional networks proposed in this invention:

[0041] like Figure 1 As shown, the method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks specifically includes the following steps:

[0042] S1. Extract key points of the human skeleton from the surveillance video stream, and construct a normalized coordinate system based on the length of the human torso, and transform the original pixel coordinates of the key points of the human skeleton into normalized coordinates.

[0043] In this step, a surveillance video stream is first acquired via a camera, and each frame of the video stream is processed using a pre-trained human pose estimation algorithm. The human pose estimation algorithm employs one of OpenPose, AlphaPose, or HRNet. For the first frame of the video sequence... The system detects and outputs the first frame of the image. Original pixel coordinates of each skeletal key point and the corresponding detection confidence level .in, The horizontal pixel coordinates are... The vertical pixel coordinates are typically in the range of [0,1]; the skeletal key points include the head, neck, shoulder, elbow, wrist, hip, knee, and ankle.

[0044] To eliminate the impact of camera shooting distance, focal length differences, and variations in human height on motion feature calculations, this embodiment constructs a human torso length benchmark based on human body scale for standardization. The neck joint and pelvic center point in the same frame are selected as reference points. The pelvic center point is taken as the midpoint between the coordinates of the left and right hip joints. The Euclidean distance between the neck joint and the pelvic center point is calculated as the human torso length benchmark for the current frame. .

[0045] Using the pelvic center point of each frame as the origin of the normalized coordinate system, the original abscissa and ordinate of the pelvic center point are subtracted from the original abscissa and ordinate of the key points of the human skeleton, respectively. The results of the subtraction are then divided by the length reference of the human torso to obtain the normalized abscissa and normalized ordinate of the key points of the human skeleton.

[0046] For example, suppose that in frame 10, the pixel coordinates of a person's right wrist are detected as follows: The coordinates of the neck are The coordinates of the center point of the pelvis are The standard length of the human torso The normalized x-coordinate of the right wrist is then... The normalized ordinate is The final normalized coordinates are .

[0047] Through this step, the motion data of all skeletal points are mapped to a unified metric space.

[0048] S2. Calculate the difference between the acceleration of each joint of the human body in the vertical direction and the preset standard gravity reference value based on the normalized coordinates, and determine the vertical runaway deviation value reflecting the degree of gravity runaway by combining the detection confidence level.

[0049] This step aims to extract the uncontrolled characteristics hidden in human movement. In normal controlled movements such as squatting, the human body exerts force through muscles to counteract gravity, resulting in a relatively gradual change in vertical acceleration. However, in uncontrolled movements such as fainting and falling, the human body loses control of its limbs, and the vertical movement will exhibit acceleration characteristics close to free fall.

[0050] Based on this, this step first calculates the normalized vertical acceleration in the vertical direction. The second-order difference operation is performed using the normalized ordinates of the previous frame, the current frame, and the next frame, and the result is divided by the square of the normalized time step. The specific calculation formula is as follows:

[0051]

[0052] in, To normalize the time step, we take 1 in this embodiment. This formula reflects the second derivative of the positional change of the human joint in the vertical direction.

[0053] Then, the vertical runaway deviation value is calculated, which combines the acceleration deviation and the detection confidence level. The calculation formula is as follows:

[0054]

[0055] In the formula, Indicates the first Frame number Vertical runaway deviation value of each joint.

[0056] This is the physical deviation term. Using a preset normalized standard gravity reference value, and based on extensive experimental statistics, the reference value for the acceleration of a human body in free fall under normalized space is typically within... In this embodiment, the value is preferably set to 0.1. If the actual acceleration is close to this value, the difference will approach 0, indicating that it meets the characteristics of an uncontrolled fall.

[0057] This is a confidence penalty term. Among them, It is a natural constant; To prevent extremely small constants with a denominator of 0, for example, take... ; For the first Frame number The confidence level of the detection of each joint.

[0058] When the detection confidence level When the confidence level is high, close to 1, the confidence penalty term is small, and the vertical runaway deviation value mainly depends on the physical deviation. When the detection confidence level is very low, close to 0, the confidence penalty term becomes large, which amplifies the vertical runaway deviation value, thereby avoiding misjudging low-quality detection data as high-precision runaway features.

[0059] This step, by calculating the vertical runaway deviation value, can accurately assess whether the current motion conforms to the laws of free fall, and automatically reduces the interference of noise data through the confidence term.

[0060] S3. Calculate the velocity vectors of each joint of the human body based on the normalized coordinates, and determine the trajectory oscillation entropy that reflects the degree of motion disorder based on the directional changes and velocity magnitudes of the velocity vectors at adjacent time points.

[0061] This step aims to quantify the degree of disorder in human limb movements. Given that abnormal behavior is usually accompanied by disordered movement characteristics such as random abrupt changes in limb direction, while the trajectory of normal behavior is relatively smooth and continuous, we first calculate the velocity components of the velocity vector along the horizontal axis by subtracting the normalized coordinates of the previous frame from the normalized coordinates of the current frame. and the velocity components of the velocity vector along the vertical axis And further calculate to obtain the first The square of the magnitude of the velocity vector at any given time , .

[0062] Based on this, a cosine similarity reflecting the degree of directional change is constructed using the ratio of the dot product of the velocity vectors at adjacent time points to the product of their magnitudes. This cosine similarity is then weighted and summed using the magnitudes of the velocity vectors to calculate the trajectory oscillation entropy. The specific formula for calculating the trajectory oscillation entropy is as follows:

[0063]

[0064] In the formula, Indicates the first Each joint in the time window The trajectory oscillation entropy within;

[0065] It is the cosine similarity of velocity vectors at adjacent time points;

[0066] It reflects the degree of change in direction. If the direction remains unchanged, the cosine value is 1, and this term is 0; if the direction reverses or changes abruptly, this term increases significantly.

[0067] As a velocity-weighted term, it means that even if the direction changes greatly, if the speed is extremely slow, such as a slight tremor when stationary, this term will reduce the trajectory oscillation entropy by multiplying it by the velocity magnitude, thereby effectively filtering background noise.

[0068] It is the second smallest constant to prevent division by zero errors.

[0069] To further explain the calculation mechanism and physical significance of trajectory oscillation entropy, a specific comparative analysis is conducted through two typical motion scenarios.

[0070] The first motion scenario involves a sudden change in direction. Assuming... The velocity vector at time step is , The velocity vector changes abruptly at time step. That is, reverse motion has occurred, and the velocity magnitudes at both moments are 1. Substituting into the above formula, we get:

[0071] ; The contribution of 2 to the trajectory oscillation entropy at that moment indicates that a violent oscillation occurred at that moment.

[0072] The second type of motion scenario is smooth motion, assuming the... Time and the The velocity vector at time t remains constant. That is, it is moving at a constant velocity in a straight line. At this time:

[0073] The contribution of the trajectory oscillation entropy at that moment is 0, which indicates that the trajectory is in a smooth state.

[0074] This example aims to explain the calculation principle of each term within the summation symbol in the formula, i.e., how to calculate the instantaneous oscillation contribution value through the dot product and modulus, rather than demonstrating the accumulation process over the entire time window. In practical algorithm applications, time windows are used to assess the overall degree of disorder over a period of time. Typically, 5 to 10 frames are used.

[0075] Through the above calculation process, the algorithm can not only effectively assess the degree of confusion in the direction of limb movement, but also automatically shield the slight jitter caused by sensor noise in a static state through the weighting mechanism of velocity modulus, thus ensuring the accuracy of feature extraction.

[0076] S4. Construct an adaptive weight function using the trajectory oscillation entropy and the vertical runaway deviation value, calculate the behavior anomaly weighting coefficient, and fuse the behavior anomaly weighting coefficient with the skeleton feature matrix to be identified, and use it as the input of the spatiotemporal graph convolutional network for behavior recognition.

[0077] This step involves fusing the two physical indices mentioned above to generate weighting coefficients for the input layer of the spatiotemporal graph convolutional network. This allows the spatiotemporal graph convolutional network to focus on key nodes that both conform to the laws of gravity runaway and exhibit chaotic trajectory characteristics.

[0078] Definition of the first Frame number The weighted coefficient for behavioral abnormalities of each joint is: This coefficient is directly proportional to the trajectory oscillation entropy and inversely proportional to the vertical runaway deviation. The specific calculation formula is as follows:

[0079]

[0080] in, This is an adjustment factor used to control the sensitivity to vertical runaway deviation; in this embodiment, the value ranges from 5.0 to 10.0.

[0081] After obtaining the weighted coefficient matrix, it is fused with the original skeletal feature matrix. The fusion method is as follows: the calculated behavioral anomaly weighting coefficients are used to multiply the skeletal feature matrix to be identified element-wise to generate a weighted skeletal feature matrix. Finally, the fused skeletal feature matrix is ​​input into a spatiotemporal graph convolutional network. Since the input features have been filtered and enhanced by physical rules, the spatiotemporal graph convolutional network can more easily capture the essential features of fall behavior, thus exhibiting high accuracy and robustness in distinguishing between rapid crouching and accidental falls.

[0082] The following combination Figure 2 and Figure 3 The technical solution and technical effects of the present invention will be further explained.

[0083] Figure 2 The curves showing the changes in trajectory oscillation entropy values ​​for different behavioral samples on the same timeline are presented. The solid line represents abnormal fall samples, and the dashed line represents normal squatting samples. A significant spike appears in the solid line near the 30th frame after the action occurs, with a value much higher than the baseline, directly reflecting the drastic and disordered abrupt change in the direction of limb movement at the moment of impact. In contrast, the dashed line exhibits low-amplitude irregular fluctuations, reflecting the natural physiological swaying of the human body when squatting in a real-world scenario, as well as minor sensor noise. Its overall value remains low, forming a stark contrast to the towering spike of the solid line, thus verifying the ability of trajectory oscillation entropy to distinguish the degree of motion disorder.

[0084] Figure 3 The diagram illustrates a comparison curve of the weighted coefficients for different behavioral samples calculated using the formula of this invention. Solid lines represent the weights of abnormal behavior features, dashed lines represent the weights of normal behavior features, and the dotted lines represent the judgment threshold. The solid line forms a large peak at a critical moment, significantly exceeding the judgment threshold, indicating that the algorithm assigns a very high weight to this action at the instant of the fall to activate the alarm system. While the dashed line contains velocity and displacement changes in the original data, it is strongly suppressed by the vertical runaway deviation of the denominator in the formula, keeping the entire curve firmly below the threshold. This clearly demonstrates that this solution can effectively filter out normal operational behaviors with similar spatial postures, achieving the design goal of reducing the false alarm rate.

[0085] Specific embodiments of the surveillance video behavior feature extraction system based on spatiotemporal graph convolutional networks proposed in this invention are as follows:

[0086] The surveillance video behavior feature extraction system based on spatiotemporal graph convolutional networks includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it can implement the surveillance video behavior feature extraction method based on spatiotemporal graph convolutional networks in the above embodiments.

[0087] The surveillance video behavior feature extraction system based on spatiotemporal graph convolutional networks also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0088] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks, characterized in that, Includes the following steps: Human skeleton key points are extracted from the surveillance video stream, and a normalized coordinate system is constructed based on the human torso length benchmark to convert the original pixel coordinates of the human skeleton key points into normalized coordinates. The difference between the acceleration of each joint in the human body in the vertical direction and the preset standard gravity reference value is calculated based on the normalized coordinates, and the vertical runaway deviation value reflecting the degree of gravity runaway is determined by combining the detection confidence level. The velocity vectors of each joint of the human body are calculated based on the normalized coordinates. Based on the directional changes and velocity magnitudes of the velocity vectors at adjacent time points, the trajectory oscillation entropy, which reflects the degree of motion disorder, is determined. An adaptive weighting function is constructed using the trajectory oscillation entropy and the vertical runaway deviation value. The abnormal behavior weighting coefficient is calculated and then fused with the skeleton feature matrix to be identified, which is used as the input of the spatiotemporal graph convolutional network for behavior recognition.

2. The method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, The baseline for the length of the human torso is determined by calculating the Euclidean distance between the cervical joint and the center point of the pelvis in the same frame.

3. The method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks according to claim 2, characterized in that, The process of converting the original pixel coordinates of the key points of the human skeleton into normalized coordinates specifically includes: Using the center point of the pelvis in each frame as the origin of the coordinate system, the original x-coordinate and original y-coordinate of the center point of the pelvis are subtracted from the original x-coordinate and original y-coordinate of the key points of the human skeleton, respectively. The results of the subtraction are then divided by the length reference of the human torso to obtain the normalized x-coordinate and normalized y-coordinate of the key points of the human skeleton.

4. The method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, The formula for calculating the vertical runaway deviation value is as follows: In the formula, Indicates the first Frame number Vertical runaway deviation value of each joint Indicates the first Frame number Normalized vertical acceleration of each joint This represents the preset normalized standard gravity reference value. Indicates the first Frame number Detection confidence of each joint, It is the first minimum constant. It is a natural constant.

5. The method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks according to claim 4, characterized in that, The normalized vertical acceleration is obtained through second-order difference operations, specifically including: The normalized vertical acceleration is obtained by performing a second-order difference operation using the normalized ordinates of the previous frame, the current frame, and the next frame, and then dividing the result by the square of the normalized time step.

6. The method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks according to claim 4, characterized in that, The formula for calculating the trajectory oscillation entropy is: In the formula, Indicates the first Each joint in the time window The trajectory oscillation entropy within, and They represent the first The velocity components of the velocity vector in the horizontal and vertical directions at any given moment. Indicates the first The square of the magnitude of the velocity vector at time step [time] It is the second minimum constant.

7. The method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks according to claim 6, characterized in that, The velocity components of the velocity vector in both the horizontal and vertical directions are obtained through first-order difference operations, specifically including: The velocity component in the horizontal direction is obtained by subtracting the normalized horizontal coordinate of the previous frame from the normalized horizontal coordinate of the current frame; the velocity component in the vertical direction is obtained by subtracting the normalized vertical coordinate of the previous frame from the normalized vertical coordinate of the current frame.

8. The method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks according to claim 6, characterized in that, The formula for calculating the weighting coefficient for behavioral abnormalities is as follows: In the formula, Indicates the first Frame number Weighted coefficients for abnormal behavior of each joint It is a regulating factor.

9. The method for extracting behavioral features from surveillance videos based on spatiotemporal graph convolutional networks according to claim 8, characterized in that, The process of fusing the behavioral anomaly weighting coefficients with the skeletal feature matrix to be identified includes: The abnormal behavior weighting coefficients are multiplied element-wise with the skeletal feature matrix to be identified to generate a weighted skeletal feature matrix.

10. A surveillance video behavior feature extraction system based on spatiotemporal graph convolutional networks, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when run by the processor, implements the surveillance video behavior feature extraction method based on spatiotemporal graph convolutional networks as described in any one of claims 1-9.

Images