Dangerous behavior early warning method and system based on image recognition

Abstract

The invention relates to the technical field of image processing, and discloses a dangerous behavior early warning method and system based on image recognition, and the method comprises the steps: obtaining a to-be-processed video stream, and obtaining the three-dimensional space coordinates of a joint point and the type, position and range information of a scene object; constructing a dynamic space-time hypergraph sequence; nodes of the dynamic space-time hypergraph are composed of joint points and scene objects, and the hyperedge set comprises internal limb hyperedges and interactive hyperedges; inputting the dynamic space-time hypergraph sequence into a space-time hypergraph attention network to generate attitude-scene fusion features; inputting the attitude-scene fusion features into a sequence prediction model, and outputting a time sequence risk curve; and when the current score of the time sequence risk curve is higher than a first preset threshold value and the current local time gradient of the time sequence risk curve is higher than a second preset threshold value, generating early warning information. According to the method, the limitation of a traditional graph model in expressing simultaneous interaction of a plurality of body parts and one or more objects is overcome, and an alarm can be reliably given in advance.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method and system for early warning of dangerous behaviors based on image recognition. Background Technology

[0002] With the rapid development of computer vision technology and artificial intelligence, the use of video surveillance systems for the identification and early warning of dangerous behaviors in fields such as safe production, public security, and smart elderly care has become a research hotspot. Its core objective is to automatically detect and provide early warnings of potential dangerous events by analyzing video streams, such as worker violations, pedestrian fall hazards, or driver fatigue, thereby effectively preventing accidents and protecting the lives and property of people.

[0003] Existing technologies primarily rely on deep learning models, particularly architectures based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs). These methods for identifying and warning of dangerous behaviors typically begin by extracting human pose information from video frames, such as two-dimensional or three-dimensional skeletal joints. These joint sequence sequences are then fed into temporal models (such as LSTMs or GRUs) for action classification or anomaly detection. Some more advanced methods introduce graph convolutional networks (GCNs), modeling the human skeleton as a graph structure. By performing convolution operations on the graph, they capture the intrinsic relationships between joints, thereby enhancing the understanding of dynamic changes in human pose.

[0004] However, the aforementioned existing methods for identifying and warning of dangerous behaviors primarily focus on human posture itself. When multiple body parts of a person interact with one or more objects in the scene simultaneously, these models struggle to effectively capture and represent such complex, multi-faceted interactions. Moreover, the core objective of these existing technologies is "action classification" or "anomaly detection," meaning they typically only provide identification and warnings after a dangerous behavior has occurred. This reactive approach cannot effectively prevent accidents from happening. Summary of the Invention

[0005] This invention provides a dangerous behavior early warning method and system based on image recognition to solve the limitations of the prior art when multiple body parts interact with multiple objects at the same time, and the problem that identification and early warning can only be carried out after dangerous behavior occurs.

[0006] In a first aspect, the dangerous behavior early warning method based on image recognition of the present invention includes the following steps: The process involves acquiring the video stream to be processed, performing human joint detection and scene object segmentation on each frame to obtain the 3D spatial coordinates of the joints and the category, location, and range information of the scene objects. Based on the 3D spatial coordinates of the joints and the scene object information, a dynamic spatiotemporal hypergraph sequence is constructed. The nodes of the dynamic spatiotemporal hypergraph consist of joints and scene objects. The hyperedge set includes two types: one is the predefined internal limb hyperedge based on the human skeletal topology, and the other is the dynamically generated interactive hyperedge based on the 3D spatial proximity of the joints and scene objects. An interactive hyperedge is generated when the distance between the joint and the scene object is less than a preset distance threshold. The dynamic spatiotemporal hypergraph sequence is input into the spatiotemporal hypergraph attention network, and through information transmission and aggregation mechanisms... Adaptively learn and update the feature representations of nodes and hyperedges to generate pose-scene fusion features. The attention weights in the spatiotemporal hypergraph attention network characterize the contribution of different key points, scene objects, and their interactions to potential hazards. The pose-scene fusion features of T consecutive frames are input into the sequence prediction model, which outputs a sequence of hazard quantification scores for the next K time steps, forming a temporal risk curve. When the current score of the temporal risk curve is higher than a first preset threshold and the current local temporal gradient of the temporal risk curve is higher than a second preset threshold, a warning message containing the identifiers of human key points and scene objects associated with the interactive hyperedges that contribute the most to risk prediction is generated. The contribution is determined according to the attention weights in the spatiotemporal hypergraph attention network.

[0007] Preferably, the internal limb hyperedge is constructed by dividing a preset set of human joints into several limb parts according to physiological structure; the interactive hyperedge is dynamically generated by calculating the three-dimensional Euclidean distance between each human joint and the nearest point on the surface of each scene object. When the three-dimensional Euclidean distance is less than a preset distance threshold, an interactive hyperedge is constructed between the joint and the scene object.

[0008] Preferably, obtaining the three-dimensional spatial coordinates of the joint points includes: processing video frames using a two-dimensional pose estimation algorithm to obtain the two-dimensional joint point coordinates of the human body in the video frames; generating a depth map corresponding to the video frames using a depth estimation algorithm; and recovering the three-dimensional spatial coordinates of the joint points by combining the two-dimensional joint point coordinates with the depth values ​​of their corresponding positions on the depth map.

[0009] Preferably, the information transmission and aggregation mechanism in the spatiotemporal hypergraph attention network includes: after node features undergo a first learnable transformation, they are aggregated based on attention weights to update the hyperedge features they are connected to; after hyperedge features undergo a second learnable transformation, they are aggregated based on attention weights to update the node features they are connected to; the attention weights are dynamically calculated based on the node and hyperedge features to characterize the importance of their interrelationship.

[0010] Preferably, the sequence prediction model is a recurrent neural network or a variant thereof, used to learn the temporal dependency of the pose-scene fusion features; T for the consecutive T frames is 30, and K for the future K time steps is 10.

[0011] Preferably, the conditions for generating early warning information include: comparing the maximum predicted score in the hazard quantification score sequence over the next K time steps with a first preset threshold; calculating the linear regression slope of the hazard quantification score sequence to characterize the local time gradient; and generating early warning information when the maximum predicted score is higher than the first preset threshold and the local time gradient is higher than a second preset threshold.

[0012] Preferably, determining the interactive superedge that contributes the most to risk prediction includes: when an early warning is triggered, performing a time-weighted summation of the attention weights of each interactive superedge within the past T frames to obtain the cumulative contribution of each interactive superedge, wherein the time-weighted summation assigns a higher weight to frames closer to the current time; and selecting the interactive superedge with the highest cumulative contribution as the interactive superedge that contributes the most to risk prediction.

[0013] Preferably, the two-dimensional pose estimation algorithm is the HRNet algorithm.

[0014] Preferably, the depth estimation algorithm is the MiDaS algorithm.

[0015] Secondly, the dangerous behavior warning system based on image recognition of the present invention includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the above-mentioned dangerous behavior warning method based on image recognition.

[0016] The beneficial effects of this invention are as follows: By constructing a dynamic spatiotemporal hypergraph sequence containing human body joints and scene objects, this invention overcomes the limitations of traditional graph models in representing the simultaneous interaction of multiple body parts with one or more objects. Furthermore, by quantitatively predicting future hazards and forming a time-series risk curve, this invention achieves a shift from post-event detection to pre-event warning. By judging the absolute level of the risk score and its growth rate, it can issue warnings in advance and reliably. In addition, the warning information provided by this invention can directly indicate which specific body part and which scene object contributes the most to the risk, making the warning results provide clear guidance for risk tracing and precise prevention. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the dangerous behavior early warning method based on image recognition provided in an embodiment of the present invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0019] like Figure 1 As shown, an embodiment of the dangerous behavior early warning method based on image recognition provided by the present invention specifically includes the following steps: S1. Acquire the video stream to be processed, and perform human joint detection and scene object segmentation on each frame in the sequence to obtain the three-dimensional spatial coordinates of the human joints and the category, position and range information of the scene objects.

[0020] RGB video streams are obtained by calling the surveillance camera interface or reading video files. For each frame, a human body is identified using a target detection model such as YOLOv5, and then 25 two-dimensional skeletal joints of the human body are extracted using pose estimation algorithms such as HRNet.

[0021] The coordinates of the two-dimensional joints are input into a pre-trained 2D-to-3D upscaling network, such as VideoPose3D, to recover the three-dimensional spatial coordinates of each joint in the camera coordinate system. Simultaneously, an instance segmentation network, such as Mask R-CNN, is applied to the same frame to identify objects in the scene, such as a robotic arm, the ground, and a table, and to obtain the pixel-level mask, category label, and 3D bounding box information of each object, thereby determining the object's category, location, and extent.

[0022] S2. Based on the three-dimensional spatial coordinates of joints and scene object information of consecutive frames, construct a dynamic spatiotemporal hypergraph sequence; the nodes of the dynamic spatiotemporal hypergraph are composed of human joints and scene objects, and the hyperedge set includes two types: one is the internal limb hyperedge predefined based on the human skeleton topology, and the other is the interactive hyperedge dynamically generated according to the three-dimensional spatial proximity of human joints and scene objects. The interactive hyperedge is generated when the spatial distance between the joint and the scene object is less than a preset distance threshold.

[0023] For each frame, the human joints and scene objects obtained in the previous step are used as nodes in the graph. Internal limb hyperedges are defined using fixed definitions based on human anatomy; for example, the left shoulder, left elbow, and left wrist joints are defined as a hyperedge called "left arm," and the left hip, left knee, and left ankle are defined as "left leg." Interactive hyperedges are dynamically generated each frame. The 3D Euclidean distance between each joint (e.g., the right hand) and each object in the scene (e.g., a rotating grinding wheel) is calculated. If this 3D Euclidean distance is less than a set value of 30 centimeters, an interactive hyperedge containing the right hand joint and the grinding wheel object is created. If both hands simultaneously approach the grinding wheel, a higher-order hyperedge containing the left hand, right hand, and grinding wheel nodes is generated. This process is repeated across multiple consecutive frames, forming a hypergraph sequence that evolves over time.

[0024] S3. Input the dynamic spatiotemporal hypergraph sequence into the spatiotemporal hypergraph attention network. Through information transmission and aggregation mechanisms, adaptively learn and update the feature representations of nodes and hyperedges to generate pose-scene fusion features. The attention weights in the spatiotemporal hypergraph attention network represent the contribution of different key points, scene objects and their interactions to potential hazards.

[0025] Specifically, the constructed dynamic spatiotemporal hypergraph sequence is fed into a network composed of multiple stacked spatiotemporal hypergraph attention layers. In each layer, information transmission from node to hyperedge is first performed, where each key node and object node passes its feature vector to each of its associated hyperedges. The hyperedge uses an attention mechanism to aggregate all the node information it receives, assigning different weights to different nodes to form hyperedge features. Information transmission from hyperedge to node is then performed, with the hyperedge transmitting its new features back to all nodes it contains. Each node again uses the attention mechanism to aggregate information from different hyperedges and update its own features. After multiple layers of propagation, the final feature representation of each node integrates its own pose information, the attributes of neighboring objects, and higher-order interaction relationships. The attention weight matrix output by the network quantifies the contribution of each interaction to the final risk assessment.

[0026] S4. Input the attitude-scene fusion features of consecutive T frames into the sequence prediction model, and output the hazard quantification score sequence for the next K time steps to form a time-series risk curve.

[0027] More specifically, at each time step, the pose-scene fusion features output by the spatiotemporal hypergraph attention network, such as graph-level feature vectors obtained by average pooling all node features, are collected. After collecting feature vectors for T consecutive time steps, for example, T equals 30 frames, this feature sequence of length 30 is input into a sequence prediction model based on a gated recurrent unit (GRU) or Transformer encoder architecture. This sequence prediction model is trained to predict the risk scores for the next K time steps, for example, K equals 15 frames. These 15 consecutive scores output by the sequence prediction model constitute a quantitative prediction of future risk, which are connected to form a temporal risk curve.

[0028] S5. When the current score of the time-series risk curve is higher than the first preset threshold, and the local temporal gradient of the time-series risk curve at the current moment is higher than the second preset threshold, an early warning message is generated. The early warning message includes the identifiers of human body key points and scene objects associated with the interactive hyperedge that contributes the most to risk prediction. The contribution is determined according to the attention weights in the spatiotemporal hypergraph attention network.

[0029] The time-series risk curve is continuously monitored, assuming a first preset threshold of 0.8 and a second preset threshold of 0.1. If the score on the risk curve at the current time t is higher than 0.8, the difference between this point and the score at the previous time t-1, i.e., the local temporal gradient, is calculated. If the local temporal gradient is higher than 0.1, both conditions are met, triggering an alert. The system then backtracks to the spatiotemporal hypergraph attention network, searching for the interactive hyperedge with the highest attention weight in the layer that contributes most to the final prediction result. For example, if the interactive hyperedge connecting the right-hand joint to the cutting machine object is found to have the highest attention weight, an alert is generated: Warning, the risk of interaction between the right hand and the cutting machine has increased dramatically.

[0030] In an optional embodiment, the internal limb hyperedge is constructed by dividing a preset set of human joints into several limb parts according to physiological structure; the interactive hyperedge is dynamically generated by calculating the three-dimensional Euclidean distance between each human joint and the nearest point on the surface of each scene object. When the three-dimensional Euclidean distance is less than a preset distance threshold, an interactive hyperedge is constructed between the joint and the scene object.

[0031] Specifically, in a dataset containing 17 human joints, the left shoulder, left elbow, and left wrist joints can be grouped into an internal limb hyperedge called the left arm to represent the overall movement of the arm. For the construction of the interactive hyperedge, assuming there is a running device in the scene, the 3D Euclidean distance between each human joint and the device surface is continuously calculated. For example, at a certain moment, the closest distance between the right wrist joint and the device surface is calculated to be 0.05 meters, while the preset distance threshold is 0.1 meters. Since 0.05 meters is less than 0.1 meters, an interactive hyperedge connecting the right wrist node and the device node is created to represent the proximity relationship between the hand and the hazardous device.

[0032] In an optional embodiment, obtaining the three-dimensional spatial coordinates of human joints includes: processing video frames using a two-dimensional pose estimation algorithm to obtain the two-dimensional joint coordinates of the human body in the video frames; generating a depth map corresponding to the video frames using a depth estimation algorithm; and recovering the three-dimensional spatial coordinates of the joints by combining the two-dimensional joint coordinates with the depth values ​​of their corresponding positions on the depth map.

[0033] Analyzing a 1920×1080 resolution video frame using a 2D pose estimation algorithm such as HRNet, the 2D pixel coordinates of the right elbow joint are obtained as (800, 600). Simultaneously, the same video frame is processed using a depth estimation algorithm such as MiDaS to generate a depth map, where the value of each pixel represents the distance of that point from the camera in 3D space. By querying the depth value at coordinates (800, 600) on the depth map, this value is found to be 2.5 meters. Combining the camera's intrinsic parameter matrix, and using the 2D pixel coordinates (800, 600) and the depth value of 2.5 meters, the 3D spatial coordinates of the right elbow joint in the real world can be reconstructed through backprojection.

[0034] In an optional embodiment, the information transmission and aggregation mechanism in the spatiotemporal hypergraph attention network includes: after node features undergo a first learnable transformation, they are aggregated based on attention weights to update the hyperedge features they are connected to; after hyperedge features undergo a second learnable transformation, they are aggregated based on attention weights to update the node features they are connected to; the attention weights are dynamically calculated based on the node and hyperedge features to characterize the importance of their interrelationship.

[0035] In an internal limb hyperedge representing the left arm, three nodes are connected: the left shoulder, left elbow, and left wrist. During the information transmission phase from node to hyperedge, assuming the left wrist node has the largest range of motion, the attention weights dynamically calculated by the spatiotemporal hypergraph attention network might be 0.7 for the left wrist, 0.2 for the left elbow, and 0.1 for the left shoulder. The feature vectors of these three nodes are weighted and summed according to these weights to update the features of the left arm hyperedge, making it primarily reflect the wrist's movement state. Subsequently, during the information transmission phase from hyperedge to node, the left wrist node is simultaneously connected to both the left arm hyperedge and a hyperedge representing hand-machine interaction. If the hand is approaching the machine at this time, the attention weights calculated by the spatiotemporal hypergraph attention network might be 0.8 for the interaction hyperedge and 0.2 for the left arm hyperedge. The features of the two hyperedges are then aggregated based on these weights to update the features of the left wrist node, enabling it to integrate information from both limb movement and interaction with the external environment.

[0036] In an optional embodiment, the sequence prediction model is a recurrent neural network or a variant thereof, used to learn the temporal dependencies of the pose-scene fusion features; T for the consecutive T frames is 30, and K for the future K time steps is 10.

[0037] Specifically, a Long Short-Term Memory (LSTM) network is chosen as the sequence prediction model. This model receives pose-scene fusion features from the past 30 consecutive frames (T = 30 frames per second) as input. Assuming a video frame rate of 30 frames per second, this is equivalent to the sequence prediction model analyzing the complete action sequence within the past second. For example, the input is a sequence of 30 256-dimensional feature vectors. After learning and processing by the LSTM network, the sequence prediction model outputs a feature sequence for the next K = 10 time steps, predicting the evolution trend of human pose and scene interaction over the next approximately 0.33 seconds.

[0038] In an optional embodiment, the conditions for generating early warning information include: comparing the maximum predicted score in the hazard quantification score sequence over the next K time steps with a first preset threshold; calculating the linear regression slope of the hazard quantification score sequence to characterize the local temporal gradient; and generating early warning information when the maximum predicted score is higher than the first preset threshold and the local temporal gradient is higher than a second preset threshold.

[0039] Suppose the predicted danger score sequence for the next 10 frames is 0.60, 0.65, 0.70, 0.75, 0.82, 0.85, 0.88, 0.91, 0.90, 0.89. The highest predicted score in this sequence is 0.91. A first preset threshold is set at 0.85; since 0.91 is higher than 0.85, the first condition is met. Linear regression analysis is performed on these 10 score points, and the slope of the regression line is calculated to be 0.035, which represents the rate of increase in danger. A second preset threshold is set at 0.02; since 0.035 is higher than 0.02, the second condition is met. Because both conditions are met simultaneously, it is determined that this is an impending event with rapidly increasing danger, and an audible and visual alarm is triggered to issue a warning.

[0040] In an optional embodiment, determining the interactive superedge that contributes the most to risk prediction includes: when an alert is triggered, performing a time-weighted summation of the attention weights of each interactive superedge within the past T frames to obtain the cumulative contribution of each interactive superedge, wherein the time-weighted summation assigns a higher weight to frames closer to the current time; and selecting the interactive superedge with the highest cumulative contribution as the interactive superedge that contributes the most to risk prediction.

[0041] When an alert is triggered, data from the past 30 frames (T = 30 frames) is analyzed. Assume there are two interactive hyperedges in the scene: one for hand-rotating device interaction and the other for foot-ground interaction. The attention weights of these two hyperedges for each of the past 30 frames are extracted and summed temporally. For example, the attention weight of the most recent frame is multiplied by a coefficient of 1.0, the second most recent frame by 0.98, and so on, with older frames having lower weight coefficients. If the calculation shows that the cumulative contribution of the hand-rotating device interaction hyperedge is 22.5, and the cumulative contribution of the foot-ground interaction hyperedge is 3.1, then the interaction between the hand and the rotating device is determined to be the main cause of this risk alert, and this interaction relationship is highlighted on the monitoring interface to alert the operator.

[0042] The implementation principle of the image recognition-based dangerous behavior early warning method of this invention is as follows: This invention constructs a dynamic spatiotemporal hypergraph sequence containing human body joints and scene objects, which overcomes the limitations of traditional graph models in representing the simultaneous interaction of multiple body parts with one or more objects. Furthermore, by quantitatively predicting future risks and forming a time-series risk curve, this invention achieves a shift from post-event detection to pre-event warning. By judging the absolute level of the risk score and its growth rate, it can issue warnings reliably and in advance. In addition, the early warning information provided by this invention can directly indicate which specific body part and which scene object contributes the most to the risk, making the early warning results provide clearer guidance for risk tracing and precise prevention.

[0043] An embodiment of the dangerous behavior warning system based on image recognition provided by the present invention includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the dangerous behavior warning method based on image recognition in the above embodiment.

[0044] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for early warning of dangerous behaviors based on image recognition, characterized in that, include: The process involves acquiring the video stream to be processed, performing human joint detection and scene object segmentation on each frame, and obtaining the three-dimensional spatial coordinates of the joints as well as the category, position, and range information of the scene objects. A dynamic spatiotemporal hypergraph sequence is constructed based on the 3D spatial coordinates of key points and scene object information. The nodes of the dynamic spatiotemporal hypergraph consist of key points and scene objects. The hyperedge set includes two types: one is the internal limb hyperedge predefined based on the human skeleton topology, and the other is the interactive hyperedge dynamically generated based on the 3D spatial proximity of key points and scene objects. The interactive hyperedge is generated when the distance between a key point and a scene object is less than a preset distance threshold. The dynamic spatiotemporal hypergraph sequence is input into the spatiotemporal hypergraph attention network. Through information transmission and aggregation mechanisms, the feature representations of nodes and hyperedges are adaptively learned and updated to generate pose-scene fusion features. The attention weights in the spatiotemporal hypergraph attention network represent the contribution of different key points, scene objects and their interactions to potential hazards. The pose-scene fusion features of T consecutive frames are input into the sequence prediction model, which outputs a sequence of risk quantification scores for the next K time steps to form a temporal risk curve. When the current score of the temporal risk curve is higher than a first preset threshold and the current local temporal gradient of the temporal risk curve is higher than a second preset threshold, a warning message is generated that includes the identification of human key points and scene objects associated with the interactive hyperedge that contributes the most to the risk prediction. The contribution is determined according to the attention weights in the spatiotemporal hypergraph attention network.

2. The dangerous behavior early warning method based on image recognition according to claim 1, characterized in that, The internal limb hyperedge is constructed by dividing a preset set of human joints into several limb parts according to physiological structure; the interactive hyperedge is dynamically generated by calculating the three-dimensional Euclidean distance between each human joint and the nearest point on the surface of each scene object. When the three-dimensional Euclidean distance is less than a preset distance threshold, an interactive hyperedge is constructed between the joint and the scene object.

3. The dangerous behavior early warning method based on image recognition according to claim 1, characterized in that, The process of obtaining the three-dimensional spatial coordinates of the joint points includes: processing video frames using a two-dimensional pose estimation algorithm to obtain the two-dimensional joint coordinates of the human body in the video frames; generating a depth map corresponding to the video frames using a depth estimation algorithm; and recovering the three-dimensional spatial coordinates of the joint points by combining the two-dimensional joint coordinates with the depth values ​​of their corresponding positions on the depth map.

4. The dangerous behavior early warning method based on image recognition according to claim 1, characterized in that, The information transmission and aggregation mechanism in the spatiotemporal hypergraph attention network includes: after node features undergo a first learnable transformation, they are aggregated based on attention weights to update the hyperedge features they are connected to; after hyperedge features undergo a second learnable transformation, they are aggregated based on attention weights to update the node features they are connected to; the attention weights are dynamically calculated based on the node and hyperedge features to characterize the importance of their interrelationship.

5. The dangerous behavior early warning method based on image recognition according to claim 1, characterized in that, The sequence prediction model is a recurrent neural network or a variant thereof, used to learn the temporal dependency of the pose-scene fusion features; T is 30 for the consecutive T frames, and K is 10 for the K future time steps.

6. The dangerous behavior early warning method based on image recognition according to claim 1, characterized in that, The conditions for generating early warning information include: comparing the maximum predicted score in the hazard quantification score sequence over the next K time steps with a first preset threshold; calculating the linear regression slope of the hazard quantification score sequence to characterize the local time gradient; and generating early warning information when the maximum predicted score is higher than the first preset threshold and the local time gradient is higher than a second preset threshold.

7. The dangerous behavior early warning method based on image recognition according to claim 1, characterized in that, The process of determining the interactive hyperedge that contributes the most to risk prediction includes: when an alert is triggered, performing a time-weighted summation of the attention weights of each interactive hyperedge within the past T frames to obtain the cumulative contribution of each interactive hyperedge, wherein the time-weighted summation assigns a higher weight to frames closer to the current time; and selecting the interactive hyperedge with the highest cumulative contribution as the interactive hyperedge that contributes the most to risk prediction.

8. The dangerous behavior early warning method based on image recognition according to claim 3, characterized in that, The two-dimensional attitude estimation algorithm is the HRNet algorithm.

9. The dangerous behavior early warning method based on image recognition according to claim 3, characterized in that, The depth estimation algorithm is the MiDaS algorithm.

10. A dangerous behavior early warning system based on image recognition, characterized in that, It includes a memory and a processor. The memory stores computer instructions. When the processor executes the computer instructions, it implements the dangerous behavior warning method based on image recognition as described in any one of claims 1-9.

Images