A method and system for detecting dangerous behavior of a person in a station scenario

Abstract

This invention relates to the field of rail transit safety technology and discloses a method and system for detecting dangerous human behavior in station scenarios. The method includes: receiving video streams from multiple cameras and preprocessing the images; extracting human dynamic features from the images and fusing them with a pre-established station semantic map to obtain a heterogeneous map; performing spatiotemporal fusion reasoning in the heterogeneous map to output a dangerous intent score; aggregating the dangerous intent scores into a candidate event sequence within a sliding time window, calculating the causal continuity probability and determining the behavioral evolution chain, and classifying the event risk level according to the risk index; and matching corresponding alarm and handling strategies according to the event risk level. This invention jointly models the posture features, movement trajectory, and station semantic region constraints of a person, and simultaneously constructs a causal risk judgment mechanism oriented towards temporal behavioral evolution, enabling real-time identification, early warning, and coordinated handling of dangerous behaviors.

Description

Technical Field

[0001] This invention relates to the field of rail transit safety technology, and in particular to a method and system for detecting dangerous human behavior in station scenarios. Background Technology

[0002] Station scenarios are characterized by high population density, well-defined spatial boundaries, high risk levels, and rapid behavioral changes. Existing methods based on general object detection or single-frame action recognition often only focus on whether a person is present or moving, lacking comprehensive modeling of the relationships between people and the semantics of the platform area, track boundaries, safety lines, turnstile status, and train arrival status. Therefore, they are prone to the following problems: 1. Dangerous behaviors are difficult to distinguish from normal behaviors: For example, actions such as briefly stopping, bending down to pick up an object, turning around quickly, or standing near a safety line may appear similar to dangerous behaviors in a single frame, which can easily lead to false alarms.

[0003] 2. Lack of constraints on the semantic regions of the station: Existing methods usually do not explicitly model the spatial semantics of platform edges, waiting areas, restricted areas, track areas, turnstiles, etc., which makes it impossible for the system to determine whether a person's behavior is occurring in a high-risk area.

[0004] 3. Difficulty in describing the evolution of dangerous behaviors: Many dangerous behaviors are not completed instantaneously, but rather evolve gradually from approaching the boundary to crossing the boundary or entering the restricted area. If only single-frame recognition is used, it is difficult to capture the chain of behavioral evolution, leading to missed detections.

[0005] The core of the dangerous behavior detection system in railway stations lies in the accurate real-time identification and risk assessment of complex human behaviors. Essentially, it is a spatiotemporal modeling task oriented towards strong semantic constraints. The spatial structure of the platform area is highly structured; boundaries such as safety lines, restricted areas, and turnstiles have clear risk semantics. The interaction between individuals and these areas determines the level of danger of their behavior.

[0006] However, existing general visual recognition methods still face significant challenges when processing station scene data. On the one hand, single-frame action recognition ignores the temporal evolution of behavior. Dangerous behaviors often undergo a multi-stage evolution process, such as approaching, probing, and crossing the line. Relying solely on single-frame features makes it difficult to accurately distinguish between dangerous and normal behaviors, leading to a large number of false alarms and false negatives. On the other hand, existing methods lack explicit modeling of the semantic regions of the station, failing to associate the spatial location of a person with the semantic relationship of high-risk areas such as platform edges and track areas. This results in the system lacking the ability to perceive whether a behavior occurs in a dangerous area. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a method and system for detecting dangerous human behavior in a station setting. It jointly models the posture features and movement trajectory of a person with the semantic region constraints of the station, and simultaneously constructs a causal risk assessment mechanism oriented towards the evolution of temporal behavior, enabling real-time identification, early warning, and coordinated handling of dangerous behaviors.

[0008] The technical solution adopted in this invention is as follows: A method for detecting dangerous human behavior in a train station setting includes: Receive video streams from multiple cameras and preprocess the images, including noise reduction, enhancement, and correction operations; Human dynamic features are extracted from the image and fused with a pre-established station semantic map to obtain a heterogeneous map; spatiotemporal fusion reasoning is performed on the heterogeneous map to output a dangerous intent score; Within a sliding time window, dangerous intent scores are aggregated into candidate event sequences, the probability of causal continuity is calculated, and the behavioral evolution chain is determined. Event risk levels are then classified based on the risk index. Match the corresponding alarm and handling strategies based on the event risk level, and output the dangerous behavior category and location, event risk level, and alarm and handling strategies.

[0009] Furthermore, the method for establishing the station semantic map includes: The station's hazardous and critical functional areas are divided into different semantic areas, including the platform edge, safety line, waiting area, turnstile area, stairwell, escalator entrance, and track access restricted area. Assigning risk weights to each semantic region The risk weight The calculation methods include:

[0010] in, semantic region Basic risk weights; For scene dynamic adjustment coefficients; This is the Sigmod activation function; for t The train's speed vector at the moment of arrival; This is the normal vector of the region boundary.

[0011] Furthermore, the extraction of human dynamic features from the image and its fusion with a pre-established station semantic map to obtain a heterogeneous map includes: Based on a preset target detection and pose estimation network, the system outputs human dynamic features in real time. These features include detection boxes, skeleton key points, trajectory coordinates, movement speed, and body and head orientation angles. By combining human dynamic features with a pre-established semantic map of the station, a heterogeneous graph of human body, semantic region, and boundary is constructed, and interaction weights are calculated through a spatiotemporal attention mechanism. This enables multimodal feature alignment.

[0012] Furthermore, the interaction weights The calculation methods include:

[0013] in, for t Query vector of dynamic human features at any given time, superscript Indicates the transpose operation; semantic region The key vector, semantic region The key vector, This represents the number of semantic regions. For feature dimensions; This is the spatial attenuation coefficient; For the human body's core To the regional boundary The Euclidean distance.

[0014] Furthermore, the process of performing spatiotemporal fusion reasoning on the heterogeneous graph and outputting a dangerous intent score includes: By employing spatiotemporal attention mechanisms or graph structure relationship learning methods, the association between the human body, semantic regions, boundaries, and time segments is learned, and a danger intent score is output. The danger intent score The calculation methods include:

[0015] Among them, the danger intent score The closer the value is to 1, the stronger the dangerous intent. and For learnable parameters, superscript Indicates the transpose operation; It is a characteristic that integrates human posture and movement; Indicates feature splicing; The number of semantic regions; semantic region Risk weights.

[0016] Furthermore, the aggregation of dangerous intent scores into a sequence of candidate events within a sliding time window includes: Aggregating dangerous intent scores across consecutive frames within a sliding time window Aggregates to form a candidate event sequence :

[0017] in, The length of the sliding time window; For displacement increment; This represents the rate of change of the orientation angle.

[0018] Furthermore, the calculation of causal continuity probability and determination of behavioral evolution chains includes: Candidate event sequence After the system is built, causal relationship determination is performed. By combining preset rule constraints with a time series model, it is determined whether there is causal continuity between events and the probability of causal continuity is calculated. Thus, the behavioral evolution chain is determined; the probability of causal continuity The calculation methods include:

[0019] in, This is the state transition probability matrix; For the first Discrete behavioral states of the step; Prior to the scene rules; The time interval between two adjacent behavioral state segments; This is the maximum reasonable evolution time window.

[0020] Furthermore, the classification of event risk levels based on risk indices includes: Risky behaviors are classified into multiple levels of risk, and a multi-level risk index is output by combining causal continuity, mean intention, and duration of stay. The multi-level risk index The calculation methods include:

[0021] in, The mean of the danger intent score within the sliding time window; For causal continuity probability; This is the residence time decay coefficient; The weighting coefficients are configurable and satisfy the following conditions: .

[0022] Furthermore, the matching of corresponding alarm and handling strategies based on the event risk level includes: Predefined multi-level risk thresholds for multi-level risk indices calculated in real time. Select the highest priority level as the final output:

[0023] in, It is the final output discrete control command, used to specify the specific security linkage strategy to be executed at the current moment; This is the threshold for grading; This is the corresponding handling strategy vector; By executing alarm and handling strategies of corresponding levels through various terminals, a complete closed loop of real-time security management of the station is achieved.

[0024] A system for detecting dangerous human behavior in a train station setting includes: The video perception and preprocessing module is configured to receive video streams captured by multiple cameras and preprocess the images, including noise reduction, enhancement and correction operations. The station area semantic constraint spatiotemporal behavior recognition module is configured to extract human dynamic features from images and fuse them with a pre-established station semantic map to obtain a heterogeneous map; perform spatiotemporal fusion reasoning in the heterogeneous map and output a dangerous intent score; The multi-event causal risk assessment and coordinated response module is configured to aggregate dangerous intent scores into candidate event sequences within a sliding time window, calculate the probability of causal continuity and determine the behavioral evolution chain, and classify event risk levels based on risk indices. The output and application module is configured to match the corresponding alarm and handling strategies according to the event risk level, and output the dangerous behavior category and location, event risk level, and alarm and handling strategies.

[0025] The beneficial effects of this invention are as follows: 1. This invention integrates human dynamic features with station semantic maps to conduct heterogeneous graph spatiotemporal reasoning. It can combine the safety attributes of different areas of the station to make targeted judgments on human behavior, effectively distinguishing between normal passage, short stay, and item picking and dangerous behaviors with safety hazards, significantly improving the accuracy of identifying dangerous human behavior in station scenarios.

[0026] 2. This invention uses a sliding time window to aggregate behavioral information and determines the behavioral evolution chain through causal continuity probability, which breaks through the limitations of single-frame image judgment. It can filter out a large amount of meaningless instantaneous action noise and reduce the occurrence of false alarms. At the same time, it can capture the progressive development of dangerous behavior processes and avoid the problem of missed detection due to the lack of obvious characteristics in the early stage of behavior.

[0027] 3. This invention can classify events based on risk indices and automatically match corresponding alarm and handling strategies. It can simultaneously output information on the type, location, and risk level of dangerous behaviors, providing clear and specific decision-making basis for station safety management personnel and promoting the transformation of station safety management from passive post-event tracing to proactive in-event intervention.

[0028] 4. Based on universal video acquisition equipment and adaptable station semantic map construction, this invention can adapt to complex on-site environments such as different station structures, different camera installation angles, and multiple people obstructing each other. It does not require large-scale modification of the existing monitoring system and has good engineering deployment compatibility and promotion application value. Attached Figure Description

[0029] Figure 1 This is an overall architecture diagram of a dangerous human behavior detection system for a train station setting, according to Embodiment 2 of the present invention.

[0030] Figure 2 This is a schematic diagram of the station area semantic constraint spatiotemporal behavior recognition module in Embodiment 2 of the present invention.

[0031] Figure 3 This is a schematic diagram of the multi-event causal risk determination and linkage handling module in Embodiment 2 of the present invention. Detailed Implementation

[0032] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0033] Example 1 This embodiment provides a method for detecting dangerous human behavior in a train station setting, including: Receive video streams from multiple cameras and preprocess the images, including noise reduction, enhancement, and correction operations; Human dynamic features are extracted from the image and fused with a pre-established station semantic map to obtain a heterogeneous map; spatiotemporal fusion reasoning is performed on the heterogeneous map to output a dangerous intent score; Within a sliding time window, dangerous intent scores are aggregated into candidate event sequences, the probability of causal continuity is calculated, and the behavioral evolution chain is determined. Event risk levels are then classified based on the risk index. Match the corresponding alarm and handling strategies based on the event risk level, and output the dangerous behavior category and location, event risk level, and alarm and handling strategies.

[0034] It should be noted that the method in this embodiment adopts a processing flow of video perception—regional constraint—temporal reasoning—risk determination—linked alarm. After receiving video streams from one or more cameras, the semantic structure of the human body, key areas and station in the picture is first analyzed. Then, combined with the behavioral evolution characteristics in the time dimension, the dangerous behavior category, risk level, occurrence location and alarm strategy are output, as detailed below.

[0035] I. Spatiotemporal Behavior Recognition under Semantic Constraints in Station Areas To address the shortcomings of existing methods that lack explicit modeling of the semantic regions of stations and cannot determine whether a person's behavior occurs in a high-risk area, this embodiment proposes a spatiotemporal behavior recognition method with semantic constraints in station areas. This method pre-establishes a semantic map of the station scene, dividing the platform edge, safety line, restricted area, and track area into different semantic regions and assigning them risk weights. It then uses a spatiotemporal attention mechanism to jointly model the relationship between a person's posture, movement trajectory, and regional semantics, ensuring that the behavior recognition results are combined with strong semantic constraints of the scene. The specific details are as follows.

[0036] (1) Semantic region dynamic risk weight modeling A semantic map of the station scene is pre-built, dividing the platform edge, safety line, waiting area, turnstile area, stairwell, escalator entrance, and track restriction zone into different semantic regions, and assigning a risk weight to each region. .

[0037] Preferably, the risk weight The calculation methods include:

[0038] in, semantic region Basic risk weights; For scene dynamic adjustment coefficients; This is the Sigmod activation function; for t The train's speed vector at the moment of arrival; This is the normal vector of the region boundary.

[0039] (2) Human spatiotemporal feature extraction and graph spatiotemporal fusion reasoning Based on a pre-defined target detection and pose estimation network, real-time human dynamic features are output. These features include bounding boxes, skeleton keypoints, trajectory coordinates, movement speed, and body and head orientation angles. These human dynamic features are then combined with a pre-built station semantic map to construct a heterogeneous graph of human body, semantic region, and boundary. Interaction weights are calculated using a spatiotemporal attention mechanism. This enables multimodal feature alignment.

[0040] Preferably, the interaction weight The calculation methods include:

[0041] in, for t Query vector of dynamic human features at any given time, superscript Indicates the transpose operation; semantic region The key vector, semantic region The key vector, The number of semantic regions; For feature dimensions; This is the spatial attenuation coefficient; For the human body's core To the regional boundary The Euclidean distance.

[0042] (3) Determination of dangerous intent By employing spatiotemporal attention mechanisms or graph structure relationship learning methods, the association between the human body, semantic regions, boundaries, and time segments is learned, and a danger intent score is output. .

[0043] Preferably, the danger intent scoring The calculation methods include:

[0044] Among them, the danger intent score The closer the value is to 1, the stronger the dangerous intent. and For learnable parameters, superscript Indicates the transpose operation; It is a characteristic that integrates human posture and movement; Indicates feature splicing; The number of semantic regions; semantic region Risk weights.

[0045] For example: if a person continues to approach the edge of the platform and increases in speed, it can be judged as a risk of collision; if a person crosses the safety line and enters the restricted area, it can be judged as a risk of crossing the line / intrusion; if a person approaches the boundary during the train's entry into the station and is accompanied by abnormal movements such as leaning or rapidly waving their arms, it can be judged as a high-risk approach event.

[0046] It should be noted that by using the semantic constraints of station area spatiotemporal behavior recognition, it is possible to explicitly identify in which area the action occurs, in which direction it is directed, and with which boundaries it interacts, thereby achieving semantic understanding of dangerous behaviors in station scenarios.

[0047] II. Risk Assessment and Coordinated Response for Multiple Events To address the shortcomings of single-frame recognition in capturing the evolutionary chain of dangerous behaviors, leading to missed detections, this embodiment designs a multi-event causal risk assessment and coordinated response method. This method aggregates the detection results of consecutive frames within a sliding time window to form a candidate event sequence. It then determines the causal continuity between events by combining rule constraints with a temporal model, achieving complete identification of progressive dangerous behavior chains such as approaching—probing—crossing the line—entering the restricted area. Specific details are as follows.

[0048] The first step is behavior fragment aggregation, which aggregates the detection results of consecutive frames within a sliding time window to form candidate event sequences, such as approaching an edge—stopping—leaning forward—crossing a safety line.

[0049] Preferably, the danger intent score of consecutive frames is aggregated within a sliding time window. Aggregates to form a candidate event sequence :

[0050] in, The length of the sliding time window (usually 15 to 30 frames). For displacement increment; This represents the rate of change of the orientation angle.

[0051] Candidate event sequence After the system is built, causal relationship determination is performed. By combining preset rule constraints with a time series model, it is determined whether there is causal continuity between events. For example, bending down to pick up an object is not necessarily dangerous in itself, but if it occurs near the safety line and is accompanied by a forward shift of the human body's center of gravity, it can be classified as a high-risk event.

[0052] Preferably, the probability of causal continuity The calculation methods include:

[0053] in, This is the state transition probability matrix; For the first Discrete behavioral states of the step; Prior to the scene rules; The time interval between two adjacent behavioral state segments; This represents the maximum reasonable evolution time window. Probabilistic calculations can filter out pseudo-continuous sequences caused by jitter or irrelevant actions.

[0054] Next, risk classification is performed, categorizing risky behaviors into multiple risk levels. For example: low risk is approaching the boundary but not crossing it; medium risk includes pseudo-continuous lingering, repeated probing, and abnormal running; high risk includes crossing the safety line, entering restricted areas, climbing railings, and prolonged lingering after a fall; and emergency risk is high-risk approaching actions when a train is entering or stopping at a station. Combining causal continuity, the mean of intent, and the duration of lingering, a multi-level risk index is then output. .

[0055] Preferably, the multi-level risk index The calculation methods include:

[0056] in, The mean of the danger intent score within the sliding time window; For causal continuity probability; This is the residence time decay coefficient; The weighting coefficients are configurable and satisfy the following conditions: .

[0057] After risk classification, it can be linked with platform broadcasts, screen prompts, edge alarm lights, back-end monitoring platforms and manual duty terminals to form a closed loop for classified handling.

[0058]

[0059] in, It is the final output discrete control command, used to specify the specific security linkage strategy to be executed at the current moment; It is a multi-level risk index; This is the threshold for grading; This is the corresponding handling strategy vector.

[0060] It should be noted that, to address the shortcomings of existing systems in lacking structured risk classification and coordinated response capabilities, this embodiment constructs a multi-level risk output and coordinated alarm mechanism. Specifically, behaviors are categorized into four risk levels: low, medium, high, and emergency. Upon triggering the corresponding threshold, the system automatically triggers platform broadcasts, screen prompts, edge warning lights, and the backend monitoring platform, forming a complete closed loop from hazard identification to response, significantly improving the efficiency of real-time station safety management. This embodiment elevates hazardous behavior identification from a single classification to a continuous decision-making process encompassing event evolution, risk assessment, and coordinated response, significantly reducing false alarms and improving the response speed to real safety incidents.

[0061] Example 2 This embodiment is based on embodiment 1: This embodiment provides a dangerous behavior detection system for people in a train station setting, including: The video perception and preprocessing module is configured to receive video streams captured by multiple cameras and preprocess the images, including noise reduction, enhancement and correction operations. The station area semantic constraint spatiotemporal behavior recognition module is configured to extract human dynamic features from images and fuse them with a pre-established station semantic map to obtain a heterogeneous map; perform spatiotemporal fusion reasoning in the heterogeneous map and output a dangerous intent score; The multi-event causal risk assessment and coordinated response module is configured to aggregate dangerous intent scores into candidate event sequences within a sliding time window, calculate the probability of causal continuity and determine the behavioral evolution chain, and classify event risk levels based on risk indices. The output and application module is configured to match the corresponding alarm and handling strategies according to the event risk level, and output the dangerous behavior category and location, event risk level, and alarm and handling strategies.

[0062] like Figure 1 As shown, this system is a closed-loop safety detection system from bottom-level perception to high-level decision-making. Its operation begins with the video perception and preprocessing stage. Multiple camera video streams are acquired and preprocessed, including denoising, enhancement, and correction, providing a high-quality image foundation for subsequent analysis. Subsequently, the data enters the core station area semantic constraint spatiotemporal behavior recognition module. This module simultaneously extracts features such as human skeleton points, movement trajectories, and speeds, and deeply integrates them with a station semantic map containing information such as platform edges and safety lines. It uses a spatiotemporal attention mechanism to perform reasoning in the heterogeneous map, thereby outputting scores reflecting dangerous intent and semantically represented behavioral fragments. These fragments are then sent to the multi-event causal risk judgment and linkage response module, where they are aggregated into candidate event sequences within a sliding time window. The system filters noise and determines the behavioral evolution chain by calculating the causal continuity probability, ultimately classifying events into four levels: low, medium, high, and urgent based on the risk index. Finally, the system automatically matches linkage response strategies based on the classification results, achieving real-time closed-loop safety management through alarm terminals, dispatch platforms, platform broadcasts, and the monitoring center.

[0063] like Figure 2 As shown, the station area semantic constraint spatiotemporal behavior recognition module constructs an end-to-end network architecture from low-level visual perception to high-level semantic reasoning. Its core logic lies in strongly coupling dynamic human motion features with the physical semantics of the station space. In the feature extraction stage, this module receives a continuous sequence of video frames (frame t to frame t-N+1) and, through parallel target detection, pose estimation, and target tracking sub-networks, extracts in real time features including bounding box positions, key points of the human skeleton, and motion speed. Displacement increment Orientation Angle Human spatiotemporal feature vectors (including) To endow the system with the ability to understand the station environment, the module simultaneously introduces regional semantic coding paths, mapping key spatial elements such as platform edges, safety lines, waiting areas, track areas, and turnstile areas into structured regional semantic vectors. (dimension is) The module dynamically adjusts the risk weights of each region based on the train's arrival speed vector. Upon entering the core inference layer, the module utilizes a spatiotemporal attention mechanism to construct a heterogeneous graph of "human body-semantic region-boundary," based on the geometric distance from the human body's centroid to the boundary of a specific region. Interaction weights are calculated to achieve deep alignment between human dynamic features and environmental semantic features. Finally, the high-dimensional features are concatenated and fused. The data is fed into a prediction network, which combines nonlinear mapping to output a dangerous intent score. And the corresponding risk level, thereby transforming simple pixel-level motion into behavior judgment results with station scene constraints.

[0064] like Figure 3 As shown, the multi-event causal risk assessment and coordinated response module adopts a hierarchical processing flow, transforming discrete behavioral fragments into structured risk events and coordinated response strategies. The data input layer receives continuous behavioral fragment detection results from preceding modules, including intent scores. Displacement increment Rate of change of orientation angle In the sliding time window The detection results of consecutive frames in the internal buffer are used to extract temporal features. Then, the causal continuity probability is calculated. This filters out pseudo-continuous sequences caused by detection jitter or irrelevant actions. Then, it considers the mean intent score within the window. Continuity of causality and length of stay Calculate the risk index Based on preset thresholds, risks are categorized into four levels: low risk, medium risk, high risk, and emergency risk, and corresponding response strategy vectors are dynamically matched. The station implements hierarchical alarms and coordinated responses through various terminals (broadcast, screen prompts, edge alarm lights, back-end dispatch platform, and manual duty terminals) to achieve a complete closed loop for real-time safety management.

[0065] Example 3 This embodiment is based on embodiment 1: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the dangerous behavior detection method for a train station scene described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0066] Example 4 This embodiment is based on embodiment 1: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the dangerous behavior detection method for a person in a station scene as described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0067] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0068] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A method for detecting dangerous human behavior in a train station setting, characterized in that, include: Receive video streams from multiple cameras and preprocess the images, including noise reduction, enhancement, and correction operations; Human dynamic features are extracted from the image and fused with a pre-established station semantic map to obtain a heterogeneous map; Perform spatiotemporal fusion reasoning on heterogeneous graphs and output a dangerous intent score; Within a sliding time window, dangerous intent scores are aggregated into candidate event sequences, the probability of causal continuity is calculated, and the behavioral evolution chain is determined. Event risk levels are then classified based on the risk index. Match the corresponding alarm and handling strategies based on the event risk level, and output the dangerous behavior category and location, event risk level, and alarm and handling strategies.

2. The method for detecting dangerous human behavior in a train station setting according to claim 1, characterized in that, The method for establishing the station semantic map includes: The station's hazardous and critical functional areas are divided into different semantic areas, including the platform edge, safety line, waiting area, turnstile area, stairwell, escalator entrance, and track access restricted area. Assigning risk weights to each semantic region The risk weight The calculation methods include: in, semantic region Basic risk weights; For scene dynamic adjustment coefficients; This is the Sigmod activation function; for t The train's speed vector at the moment of arrival; This is the normal vector of the region boundary.

3. The method for detecting dangerous human behavior in a train station setting according to claim 1, characterized in that, The process of extracting human dynamic features from the image and fusing them with a pre-established station semantic map to obtain a heterogeneous map includes: Based on a preset target detection and pose estimation network, the system outputs human dynamic features in real time. These features include detection boxes, skeleton key points, trajectory coordinates, movement speed, and body and head orientation angles. By combining human dynamic features with a pre-established semantic map of the station, a heterogeneous graph of human body, semantic region, and boundary is constructed, and interaction weights are calculated through a spatiotemporal attention mechanism. This enables multimodal feature alignment.

4. The method for detecting dangerous human behavior in a train station setting according to claim 3, characterized in that, The interaction weight The calculation methods include: in, for t Query vector of dynamic human features at any given time, superscript Indicates the transpose operation; semantic region The key vector, semantic region The key vector, The number of semantic regions; For feature dimensions; This is the spatial attenuation coefficient; For the human body's core To the regional boundary The Euclidean distance.

5. The method for detecting dangerous human behavior in a train station setting according to claim 4, characterized in that, The process of performing spatiotemporal fusion reasoning on heterogeneous graphs and outputting a dangerous intent score includes: By employing spatiotemporal attention mechanisms or graph structure relationship learning methods, the association between the human body, semantic regions, boundaries, and time segments is learned, and a danger intent score is output. The danger intent score The calculation methods include: Among them, the danger intent score The closer the value is to 1, the stronger the dangerous intent. and For learnable parameters, superscript Indicates the transpose operation; It is a characteristic that integrates human posture and movement; Indicates feature splicing; The number of semantic regions; semantic region Risk weights.

6. The method for detecting dangerous human behavior in a train station setting according to claim 1, characterized in that, The aggregation of dangerous intent scores into a candidate event sequence within a sliding time window includes: Aggregating dangerous intent scores across consecutive frames within a sliding time window Aggregates to form a candidate event sequence : in, The length of the sliding time window; For displacement increment; This represents the rate of change of the orientation angle.

7. The method for detecting dangerous human behavior in a train station setting according to claim 6, characterized in that, The calculation of causal continuity probability and determination of behavioral evolution chains includes: Candidate event sequence After the system is built, causal relationship determination is performed. By combining preset rule constraints with a time series model, it is determined whether there is causal continuity between events and the probability of causal continuity is calculated. Thus, the behavioral evolution chain is determined; the probability of causal continuity The calculation methods include: in, This is the state transition probability matrix; For the first Discrete behavioral states of the step; Prior to the scene rules; The time interval between two adjacent behavioral state segments; This is the maximum reasonable evolution time window.

8. The method for detecting dangerous human behavior in a train station setting according to claim 1, characterized in that, The classification of event risk levels based on risk indices includes: Risky behaviors are classified into multiple levels of risk, and a multi-level risk index is output by combining causal continuity, mean intention, and duration of stay. The multi-level risk index The calculation methods include: in, The mean of the danger intent score within the sliding time window; For causal continuity probability; This is the residence time decay coefficient; The weighting coefficients are configurable and satisfy the following conditions: .

9. A method for detecting dangerous human behavior in a train station setting according to claim 1, characterized in that, The matching of corresponding alarm and handling strategies based on the event risk level includes: Predefined multi-level risk thresholds for multi-level risk indices calculated in real time. Select the highest priority level as the final output: in, It is the final output discrete control command, used to specify the specific security linkage strategy to be executed at the current moment; This is the threshold for grading; This is the corresponding handling strategy vector; By executing alarm and handling strategies of corresponding levels through various terminals, a complete closed loop of real-time security management of the station is achieved.

10. A system for detecting dangerous human behavior in a train station setting, characterized in that, include: The video perception and preprocessing module is configured to receive video streams captured by multiple cameras and preprocess the images, including noise reduction, enhancement and correction operations. The station area semantic constraint spatiotemporal behavior recognition module is configured to extract human dynamic features from images and fuse them with a pre-established station semantic map to obtain a heterogeneous map; perform spatiotemporal fusion reasoning in the heterogeneous map and output a dangerous intent score; The multi-event causal risk assessment and coordinated response module is configured to aggregate dangerous intent scores into candidate event sequences within a sliding time window, calculate the probability of causal continuity and determine the behavioral evolution chain, and classify event risk levels based on risk indices. The output and application module is configured to match the corresponding alarm and handling strategies according to the event risk level, and output the dangerous behavior category and location, event risk level, and alarm and handling strategies.

Images