Pedestrian interaction behavior recognition method in complex scene

Abstract

The application discloses a pedestrian interaction behavior recognition method in a complex scene, comprising: collecting pedestrian behavior videos in a preset period in a target scene, obtaining continuous trajectory data of each pedestrian and converting the continuous trajectory data into original trajectory data in a real world coordinate system, and obtaining target scene pedestrian motion trajectory data to be recognized through preprocessing; constructing initial recognition rules including initial preset rules and initial to-be-optimized rules; determining a dynamic optimization strategy of the initial to-be-optimized rules according to the target scene, optimizing the initial to-be-optimized rules, obtaining target dynamic recognition rules, and recognizing the target scene pedestrian motion trajectory data according to the rules to determine pedestrian interaction behaviors. The method can realize accurate recognition of pedestrian interaction behaviors, provides a complete pedestrian interaction behavior recognition method from data collection and processing to final behavior determination, and can provide efficient and accurate data support for urban traffic management, public security monitoring and social behavior research.

Description

Technical Field

[0001] This disclosure relates to the field of pedestrian interaction behavior recognition technology, and in particular to a method for recognizing pedestrian interaction behavior in complex scenarios. Background Technology

[0002] Pedestrian dynamics is of great significance for understanding and optimizing human activity spaces, especially in complex scenarios such as urban pedestrian zones, transportation hubs, and shopping malls. Among them, avoidance and overtaking behaviors are basic and common interaction patterns in pedestrian movement, occurring in opposite and same-direction movement scenarios, respectively. They involve multiple aspects such as pedestrians' perception of environmental information, cognitive judgment, and motion control, reflecting individuals' strategies for competing for and optimizing spatiotemporal resources.

[0003] In existing technologies, achieving automated and highly accurate recognition of pedestrian interaction behaviors in complex scenarios faces challenges. Overtaking and avoidance behaviors overlap and collinearize with other behaviors in terms of spatiotemporal trajectory features, leading to decreased inter-class separability and reduced feature discriminative power in traditional machine learning frameworks, resulting in high misclassification rates. This stems from the multimodal evolutionary characteristics of pedestrian interaction behaviors. Furthermore, existing pedestrian behavior analysis technologies suffer from methodological limitations. Traditional micro-simulation methods rely on idealized assumptions, making them difficult to adapt to real-world data and lacking generalization ability. Analysis methods based on trajectory feature engineering belong to shallow representation learning, failing to establish a quantitative description of the implicit dynamic mechanisms of pedestrian interactions and exhibiting insufficient behavior decoupling capabilities. Summary of the Invention

[0004] In view of this, the present disclosure provides a method for recognizing pedestrian interaction behavior in complex scenarios, which can solve the problems of poor accuracy in recognizing pedestrian interaction behavior in complex scenarios in the prior art.

[0005] In a first aspect, embodiments of this disclosure provide a method for recognizing pedestrian interaction behavior in complex scenarios, including:

[0006] Collect pedestrian behavior videos within a preset period in the target scene;

[0007] Continuous trajectory data for each pedestrian is obtained from the pedestrian behavior video; the continuous trajectory is a pixel-level motion trajectory.

[0008] The video pixel coordinate system of the continuous trajectory data is converted into a real-world coordinate system to obtain the original trajectory data;

[0009] The original trajectory data is preprocessed to obtain the pedestrian movement trajectory data of the target scene to be identified;

[0010] Construct initial recognition rules for pedestrian interaction behavior, wherein the initial recognition rules include initial preset rules and initial rules to be optimized;

[0011] The dynamic optimization strategy of the initial rule to be optimized is determined according to the target scenario, and the initial rule to be optimized is optimized based on the dynamic optimization strategy. The initial preset rule and the optimized initial rule to be optimized are used as the target dynamic identification rule.

[0012] Based on the target dynamic recognition rules, the pedestrian movement trajectory data of the target scene is identified to determine the pedestrian interaction behavior.

[0013] Secondly, embodiments of this disclosure provide a pedestrian interaction behavior recognition system for complex scenarios, including:

[0014] The acquisition module is used to acquire video of pedestrian behavior within a preset period in the target scene;

[0015] The continuous trajectory data acquisition module is used to acquire continuous trajectory data of each pedestrian from pedestrian behavior videos; the continuous trajectory is a pixel-level motion trajectory.

[0016] The raw trajectory data acquisition module is used to convert the video pixel coordinate system of continuous trajectory data into a real-world coordinate system to obtain the raw trajectory data.

[0017] The preprocessing module is used to preprocess the raw trajectory data to obtain the pedestrian movement trajectory data of the target scene to be identified;

[0018] The initial recognition rule construction module is used to construct the initial recognition rules for pedestrian interaction behavior. The initial recognition rules include initial preset rules and initial rules to be optimized.

[0019] The optimization module is used to determine the dynamic optimization strategy of the initial rule to be optimized based on the target scenario, and optimize the initial rule to be optimized based on the dynamic optimization strategy. The initial preset rule and the optimized initial rule to be optimized are used as the target dynamic recognition rule.

[0020] The recognition module is used to identify pedestrian movement trajectory data in the target scene according to the target dynamic recognition rules and determine pedestrian interaction behavior.

[0021] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution:

[0022] The computer device includes:

[0023] At least one processor; and,

[0024] A memory communicatively connected to the at least one processor; wherein,

[0025] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform any of the above-described pedestrian interaction behavior recognition methods in complex scenarios.

[0026] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions for causing a computer to execute any of the above-described pedestrian interaction behavior recognition methods in complex scenarios.

[0027] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0028] This application discloses a method for recognizing pedestrian interaction behavior in complex scenarios. It collects pedestrian behavior videos within a preset period in the target scene, extracts continuous trajectory data for each pedestrian from the videos (pixel-level motion trajectories), and provides a detailed data foundation for subsequent analysis by accurately recording the pedestrian's movement in the video frame. Through continuous trajectory data, information such as the pedestrian's walking path and stopping position can be observed intuitively. The method converts the video pixel coordinate system of the continuous trajectory data into a real-world coordinate system to obtain the original trajectory data. The video pixel coordinate system only reflects the relative position of the pedestrian in the video frame, while the real-world coordinate system provides the pedestrian's actual position information in the real scene, which is crucial for analyzing the actual interaction behavior of pedestrians, such as the actual distance between pedestrians and their walking direction. The method preprocesses the original trajectory data to obtain the pedestrian motion trajectory data of the target scene to be recognized. Preprocessing improves the data quality, making the trajectory data smoother and more complete, providing a better data foundation for accurately recognizing pedestrian interaction behavior. Finally, it constructs initial recognition rules for pedestrian interaction behavior, including initial preset rules and initial rules to be optimized. The scheme involves determining a dynamic optimization strategy for the initial rules to be optimized based on the target scenario, and then optimizing these rules. Both the initial preset rules and the optimized rules are used as the target dynamic recognition rules. Pedestrian movement trajectory data in the target scenario is identified based on these dynamic recognition rules to determine pedestrian interaction behavior. Optimizing the initial rules through the dynamic optimization strategy improves the accuracy and reliability of the rules, making them better suited to the characteristics of the target scenario. This dynamic optimization approach allows for continuous rule improvement; as data accumulates and algorithms improve, the rules can be continuously optimized to enhance recognition performance. Using the optimized target dynamic recognition rules to identify movement trajectory data allows for more accurate judgment of pedestrian interaction behavior, reducing false positives and false negatives. This scheme can dynamically adjust the recognition strategy according to the actual situation of the target scenario, improving accuracy and adaptability. Identifying movement trajectory data through the dynamic recognition strategy accurately determines pedestrian interaction information, which is significant for understanding pedestrian behavior patterns in complex scenarios and optimizing scenario design and management. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1This is a flowchart illustrating the pedestrian interaction behavior recognition method in complex scenarios provided in this embodiment of the disclosure.

[0031] Figure 2 A flowchart illustrating the method for acquiring raw trajectory data provided in this embodiment of the disclosure.

[0032] Figure 3 This is a flowchart illustrating a method for dynamically acquiring a dynamic recognition strategy for a target scene, as provided in an embodiment of this disclosure.

[0033] Figure 4 This is a schematic diagram illustrating the optimization process using the Bayesian algorithm provided in an embodiment of this disclosure.

[0034] Figure 5 This is a flowchart illustrating a method for determining pedestrian interaction behavior information provided in an embodiment of this disclosure.

[0035] Figure 6 A schematic diagram illustrating the overtaking behavior provided in an embodiment of this disclosure.

[0036] Figure 7 A schematic diagram illustrating obstacle avoidance behavior provided in an embodiment of this disclosure.

[0037] Figure 8 A schematic diagram of the pedestrian interaction behavior recognition framework provided in the embodiments of this disclosure.

[0038] Figure 9 This is a schematic diagram of a specific embodiment of a pedestrian interaction behavior recognition method.

[0039] Figure 10 A comparison chart of the recognition performance of different proxy models and acquisition functions provided in the embodiments of this disclosure.

[0040] Figure 11 A comparison chart of recognition performance using fixed parameters and Bayesian optimized parameters provided in the embodiments of this disclosure.

[0041] Figure 12 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0042] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0043] Reference Figure 1 This application discloses a method for recognizing pedestrian interaction behavior in complex scenarios, including:

[0044] S100 collects pedestrian behavior videos within a preset period in the target scene.

[0045] Specifically, high-definition cameras, surveillance networks, and other equipment are used to record video of the target scene from multiple angles over a long period of time to ensure that the pedestrian behavior information to be collected is covered; the video collection is the basis for subsequent analysis, and the preset cycle can be flexibly adjusted according to specific needs.

[0046] Furthermore, by using fixed-location surveillance cameras to collect data covering typical public places such as urban rail transit hubs and commercial complexes, the dynamic activities of pedestrians under different crowd densities and scene complexities can be fully recorded. With its spatiotemporal continuity and scene fidelity, surveillance video provides diverse and representative raw data for pedestrian interaction behavior analysis, ensuring that the research can capture the distribution of behavioral characteristics in real-world scenarios.

[0047] S200 acquires continuous trajectory data for each pedestrian from pedestrian behavior videos.

[0048] The continuous trajectory data includes coordinates in the video pixel coordinate system, and the continuous trajectory is a pixel-level motion trajectory.

[0049] Specifically, computer vision algorithms can be used to achieve real-time detection and continuous trajectory extraction of pedestrian targets, that is, to obtain the pixel-level motion trajectory of each pedestrian from pedestrian behavior videos. Pixel-level motion trajectories can accurately record the movement of pedestrians in the video frame, providing a detailed data foundation for subsequent analysis. Through continuous trajectory data, information such as the pedestrian's walking path and stopping position can be observed intuitively.

[0050] Pedestrian movement is characterized by low speed, high flexibility, and high density. Video detection, as a non-contact observation method, can continuously extract the spatial position of each pedestrian in different frames of surveillance video, thereby obtaining fine, individual-level continuous motion trajectory data. It is an effective means of trajectory extraction suitable for complex scenes. The core of this method lies in target recognition and multi-frame tracking of pedestrians in the video frame, thereby extracting the pixel-level motion trajectory of each pedestrian in the image. Commonly used implementations include joint detection and tracking methods based on deep learning, such as FairMOT and YOLO+DeepSORT. These methods can maintain high detection and tracking accuracy in complex scenes such as occlusion and trajectory intersection, providing a reliable pixel coordinate data foundation for subsequent behavior recognition.

[0051] S300 converts the video pixel coordinate system of continuous trajectory data into the real-world coordinate system to obtain the original trajectory data.

[0052] Video pixel coordinates can only reflect the relative position of pedestrians in the video frame, while real-world coordinates can provide the actual position information of pedestrians in the real scene. Converting video pixel coordinates to real-world coordinates ensures that trajectory data has a real spatial scale, so as to ensure that the thresholds obtained later are judgmental and operable.

[0053] S400 preprocesses the raw trajectory data to obtain the pedestrian movement trajectory data of the target scene to be identified.

[0054] Specifically, invalid data segments caused by occlusion or algorithm errors (such as trajectory breaks, drift points, and excessively short trajectories) can be removed, while motion trajectories that meet research needs and are of reliable quality can be retained, providing clean input for behavior recognition.

[0055] Preprocessing can improve data quality, making trajectory data smoother and more complete, thus providing a better data foundation for accurately identifying pedestrian interaction behavior.

[0056] S500 establishes initial recognition rules for pedestrian interaction behavior, which include initial preset rules and initial rules to be optimized.

[0057] S600 determines the dynamic optimization strategy of the initial rule to be optimized based on the target scenario, and optimizes the initial rule to be optimized based on the dynamic optimization strategy, and uses the initial preset rule and the optimized initial rule to be optimized as the target dynamic recognition rule.

[0058] The S700 identifies pedestrian movement trajectory data in a target scene based on target dynamic recognition rules and determines pedestrian interaction behavior.

[0059] Specifically, pedestrian interaction behavior includes the interaction individual's ID, event type, and the time of the entire process.

[0060] By identifying pedestrian movement trajectory data through dynamic target recognition rules, we can accurately determine the interactive behavior information between pedestrians, that is, accurately identify events that conform to behavioral patterns. This is of great significance for understanding pedestrian behavior patterns in complex scenarios and optimizing scenario design and management.

[0061] The pedestrian interaction behavior recognition method disclosed in this application forms a complete pedestrian interaction behavior recognition process through a series of steps from data acquisition, trajectory extraction, coordinate transformation, data preprocessing to rule construction and optimization. It can efficiently and accurately identify pedestrian interaction behaviors in complex scenarios. The solution has strong adaptability and flexibility, and can adjust rules and optimization strategies according to different target scenarios, making it suitable for various complex real-world scenarios. The entire solution is based on the analysis and processing of a large amount of pedestrian behavior video data. Through continuous optimization of recognition rules, it achieves data-driven recognition performance improvement, and has good development potential and application prospects.

[0062] In this embodiment, steps S100-S300 constitute the basic processing of pedestrian trajectory data, realizing the systematic transformation from raw video to structured trajectory data. Steps S400-S700, with the help of a data-driven model and rule engine, learn behavioral feature thresholds in dynamic scenarios based on labeled data, and combine them with real-time trajectory data to complete the intelligent recognition and decision output of pedestrian overtaking and avoidance interactions, thus constructing a complete behavior recognition chain. By combining data-driven threshold learning with rule-based logical judgment, the scene adaptability and accuracy of behavior recognition are significantly improved.

[0063] In this application, complex scenarios refer to real-world situations where pedestrian behavior patterns are highly diverse due to dynamic environmental changes, dense pedestrian flow, and diverse spatial structures, and data collection faces uncertainties such as noise interference and information loss, which significantly increases the difficulty of pedestrian interaction behavior recognition. Its core characteristics are the deep coupling of environmental dynamism, behavioral complexity, and data uncertainty. Typical scenarios include transportation hubs (such as subway station transfer passages and high-speed rail station waiting halls), commercial complexes (shopping mall atriums and pedestrian streets), and large event gathering places (sports stadium entrances and exits, exhibition venue passages), etc.

[0064] Reference Figure 2 The method for S300, which "converts the video pixel coordinate system of continuous trajectory data into a real-world coordinate system to obtain the original trajectory data," includes the following methods for acquiring the original trajectory data:

[0065] The S310 uses the Zhang Zhengyou calibration method to obtain camera intrinsic parameter information for collecting pedestrian behavior videos through a checkerboard target.

[0066] Among them, the camera intrinsic parameter information includes the camera distortion error D in the u direction. X The distortion error D of the camera in the v direction YThe camera's intrinsic parameters are determined by the position of the intersection of the camera's optical axis in the pixel coordinate system (U0, V0) and the camera's focal length f; the calibration error of the obtained camera intrinsic parameter information is less than 0.5 pixels. This method is applicable to various types of cameras, from ordinary consumer cameras to professional industrial cameras. Zhang Zhengyou's calibration method can be used for intrinsic parameter calibration, making it highly versatile and able to meet coordinate transformation requirements in different scenarios.

[0067] S320 calculates the camera's rotation matrix and translation vector based on the known world coordinate feature points and camera intrinsic parameter information, and establishes the transformation relationship between the world coordinate system and the camera coordinate system.

[0068] Specifically, based on feature points with known world coordinates (such as feature point data at different elevations, such as ground grid nodes, guardrail nodes, and handrail nodes in subway stations), the rotation matrix and translation vector of the camera are calculated to realize the rigid body transformation between the image plane and the physical space, that is, to establish the transformation relationship between the world coordinate system and the camera coordinate system.

[0069] Specifically, the transformation relationships include:

[0070] Where dx and dy represent the pixel physical dimensions in the u and v directions, respectively, and Z c Let A be the height of the cutting plane; let A be the camera intrinsic parameter, f / dx = f x f / dy = f y R and T are the extrinsic parameters of the camera, representing the rotation matrix and translation vector, respectively.

[0071] In practical applications, the position and orientation of the camera are constantly changing. This step takes into account the camera's rotation and translation, adapting to different camera installation positions and angles. Even if the camera moves or is adjusted in different scenarios, the relationship between the coordinate systems can be accurately established by solving the rotation matrix and translation vector.

[0072] S330, based on the transformation relationship, converts the video pixel coordinate system of continuous trajectory data into the real-world coordinate system to obtain the original trajectory data.

[0073] Through the above process, pixel coordinates x = (u, v) are converted into real-world 3D coordinates X = (x, y, z), ultimately constructing a trajectory dataset containing timestamps and spatial coordinates. (Considering only planar problems), it fully depicts the movement trajectory and interaction patterns of pedestrians in complex scenes, providing standardized data input for subsequent behavior modeling.

[0074] The method for acquiring raw trajectory data disclosed in this application, through the synergistic effect of the above three steps, can accurately convert the video pixel coordinate system into the real-world coordinate system, reducing coordinate transformation errors caused by factors such as inaccurate camera parameters and changes in camera pose, and improving the accuracy of the raw trajectory data. After converting the pedestrian trajectory data from the video pixel coordinate system to the real-world coordinate system, the system can perform pedestrian behavior analysis based on the real physical location, enhancing the system's practicality in real-world applications. In complex scenarios, pedestrian behavior analysis needs to take into account real physical spatial information. Through accurate coordinate transformation, a reliable foundation is provided for subsequent pedestrian interaction behavior recognition, behavior pattern analysis, etc., which helps to deeply understand the behavioral patterns of pedestrians in complex scenarios.

[0075] The S400 method for "preprocessing raw trajectory data to obtain pedestrian motion trajectory data of the target scene to be identified" includes:

[0076] S410: Obtain the trajectory length of each original trajectory data, and remove the original trajectory data whose trajectory length is less than the preset length threshold to obtain the first target trajectory dataset.

[0077] Preferably, if the length of a pedestrian's trajectory data is less than 4 seconds, it is considered a short trajectory. These short trajectories may represent transient pedestrians, tracking fragments, or noise. They are usually insufficient to contain a complete overtaking or avoidance interaction process, or the speed calculation may be unreliable due to too few frames. Therefore, the data is considered insufficient for reliable analysis, and the trajectory data is filtered.

[0078] S420, obtain the longitudinal displacement of each original trajectory data in the first target trajectory dataset, remove the original trajectory data whose longitudinal displacement is less than a preset longitudinal displacement threshold from the first target trajectory dataset, and obtain the second target trajectory dataset;

[0079] All the original trajectory data in the second target trajectory dataset are target motion trajectories.

[0080] Preferably, the longitudinal displacement of the trajectory is calculated, since overtaking and avoidance behaviors are usually accompanied by significant longitudinal movement. If the displacement is less than 1.5m, the trajectory is considered to have no significant movement, and the trajectory data is filtered out. This helps to filter out pedestrians who are nearly stationary, wandering in place, or mainly moving laterally.

[0081] The pedestrian trajectory dataset after initial screening, where the trajectory of pedestrian i is represented as:

[0082]

[0083] Let a complete sequence of trajectory coordinates be represented by z frames of data, forming a two-dimensional coordinate matrix:

[0084]

[0085] Among them, P i Let F be the trajectory of pedestrian i. i Let i be the set of time frames in which pedestrian i exists. Let x be the x-coordinate of pedestrian i in the f-th frame, where the x-axis is perpendicular to the pedestrian's walking direction. Let be the ordinate value of pedestrian i in frame f, where the y-axis represents the direction of pedestrian movement.

[0086] The method for constructing initial recognition rules for pedestrian interaction behavior in S500 specifically includes:

[0087] S510, Construct initial coarse recognition rules for pedestrian interaction behavior in the target scenario; the initial coarse recognition rules include initial coarse recognition preset rules and initial coarse recognition rules to be optimized.

[0088] Initial, coarse-based rule recognition leverages existing knowledge and experience to quickly establish a basic recognition framework for the pedestrian interaction behavior recognition system, enabling the system to make preliminary judgments on some common pedestrian interaction behaviors at an early stage. The rules to be optimized from the initial coarse-based recognition provide direction for further exploration of pedestrian interaction behaviors; through subsequent optimization, more accurate recognition rules can be discovered.

[0089] S520, based on the initial coarse recognition rules, constructs initial fine recognition rules for pedestrian interaction behavior; the initial fine recognition rules include initial fine recognition preset rules and initial fine recognition rules to be optimized.

[0090] The initial refined recognition preset rules, by adding more judgment conditions, can more accurately identify pedestrian interaction behavior and reduce misjudgments and omissions. The initial refined recognition optimization rules have optimized and refined the initial coarse recognition optimization rules, making the entire rule system more complete. With continuous optimization, these rules can better adapt to different scenarios and complex pedestrian interaction behaviors.

[0091] By first constructing initial coarse recognition rules and then building initial refined recognition rules based on these rules, a progressively refined and optimized rule construction process is formed. This process can fully utilize existing knowledge and data, improving the efficiency of rule construction. The entire solution has strong flexibility and scalability. In different target scenarios, only appropriate adjustments and optimizations to the initial coarse recognition rules are needed to construct pedestrian interaction behavior recognition rules suitable for that scenario. Dividing the rules into preset rules and rules to be optimized facilitates continuous rule optimization. With the continuous accumulation of data and the continuous improvement of algorithms, the initial refined recognition rules to be optimized can be continuously optimized, thereby continuously improving the accuracy and reliability of pedestrian interaction behavior recognition.

[0092] Reference Figure 3 The method of S600, which "determines the dynamic optimization strategy of the initial rules to be optimized based on the target scenario, optimizes the initial rules to be optimized based on the dynamic optimization strategy, and uses the initial preset rules and the optimized initial rules to be optimized as the target dynamic identification rules," specifically includes the following:

[0093] S610, determine the initial coarse recognition dynamic optimization strategy for the rules to be optimized based on the target scenario. The coarse recognition dynamic optimization strategy includes the first preset rule parameter space of the initial coarse recognition rules to be optimized.

[0094] Determining the parameter space based on the target scenario makes the optimization strategy more targeted and allows the rules to better adapt to the needs of the specific scenario. The first preset rule parameter space provides a clear direction for subsequent optimization, limits the scope of optimization, and avoids blindly adjusting rule parameters.

[0095] Among them, the initial coarse recognition rules to be optimized include two categories. The first category is when the target interaction behavior is a transcending behavior, the corresponding initial coarse recognition preset rules include: the number of common frames in which two pedestrians appear in the video at the same time is not less than the preset minimum number of frames, and the proportion of the first type of horizontal approach frames is not less than the first preset proportion of frames.

[0096] The minimum number of common frames in a video when two pedestrians appear simultaneously means that the two pedestrians have a common frame that is no less than a certain threshold. A common frame is the overlapping part of the trajectories of the two pedestrians in the time dimension, that is, the number of frames in the video when the two pedestrians appear simultaneously.

[0097] Furthermore, the minimum frame rate is preset to be no less than 4 seconds, because the typical process of overtaking behavior (approaching, changing lanes, accelerating, taking the lead) usually takes 3-5 seconds.

[0098] The first preset frame percentage is preferably 10%. This 10% threshold effectively filters out brief, accidental approaches (such as path intersections), ensuring the capture of continuous overtaking behavior while adapting to dynamic scenes and avoiding false detections or missed detections. In this embodiment, lateral proximity requires that in overlapping frames, the number of frames in which the lateral distance between two pedestrians is less than a certain threshold must meet a certain proportion, reflecting the degree of path proximity.

[0099] In terms of direction of movement, overtaking behavior only occurs between pedestrians moving in the same direction, so it is necessary to satisfy that the pedestrians in front and behind have a consistent movement trend in the longitudinal direction.

[0100] The velocity vector of pedestrian i at a certain moment can be approximated by the difference between adjacent coordinates:

[0101] Where Δt is the time interval, and the video frame rate f is selected. ps The reciprocal of .

[0102] The angle θ between the two velocity vectors of pedestrians i and j ij Calculated using the vector dot product formula:

[0103]

[0104] A threshold is used to determine whether the pedestrian flow is "in the same direction". When there is multi-directional pedestrian flow, cosθ must be satisfied. ij ∈(0,1]; when there is only one-way or opposite pedestrian flow, cosθ must be satisfied. ij ∈[0.7,1].

[0105] Overtaking behavior is the behavior of pedestrians moving in the same direction who, in order to shorten travel time or optimize the path, actively adjust their walking speed, path and other spatiotemporal parameters to avoid speed conflicts with slow individuals or obstacles ahead. Specifically, it is a dynamic spatial adjustment process in which pedestrians accelerate to increase relative speed and deviate from the route to overtake based on environmental parameters, and then return to the original trajectory or continue to move.

[0106] When the target interaction behavior is a transcendent behavior, the corresponding initial coarse recognition optimization rules include: the first type of horizontal spacing threshold, the first type of vertical spacing threshold, the first type of vertical approach frame ratio, and the first type of vertical approach continuous frame ratio.

[0107] When the target interaction behavior is an overtaking behavior, the corresponding first preset rule parameter space includes: the parameter space of the first type of horizontal spacing threshold, the parameter space of the first type of vertical spacing threshold, the parameter space of the first type of vertical approach frame ratio, and the parameter space of the first type of vertical approach continuous frame ratio.

[0108] The parameter space for the first type of lateral spacing threshold includes: d1≤D xmax≤H, where d1 is the minimum pedestrian shoulder width, and H is the one-way width of the pedestrian walkway. Preferably, 0.4≤D xmax The lateral distance threshold of ≤1.5 is based on the typical interaction distance between pedestrians in a walking scenario. 0.4m is the minimum shoulder width of a pedestrian, ensuring sufficient safe distance between the person passing and the person being passed; 1.5m is the common separation distance in pedestrian passages (such as the spacing between flow lines in subway passages), and exceeding this distance usually does not necessitate passing.

[0109] The parameter space for the first type of longitudinal spacing threshold includes: Z1≤D ymax ≤Z2, where Z1 is the minimum perceived distance for changes in pedestrian stride, and Z2 is the safe forward and backward distance for pedestrians walking; preferably, 0.1≤D ymax ≤1.0, the longitudinal distance range is based on the longitudinal interaction characteristics of pedestrian walking. 0.1m is the minimum perceived distance for changes in pedestrian stride, which may lead to collisions; 1.0m is the typical value of the safe distance between pedestrians (about 1-2 steps) when walking, which usually does not have the urgency to be overtaken.

[0110] The first type of parameter space for vertical approximation of frame rate percentage includes: 0.1 ≤ P y1 ≤0.2; The proportion of vertically approaching frames reflects the persistence of the approach between two pedestrians. 10% is the minimum requirement to ensure that the approach is not an instantaneous phenomenon; 20% is the upper limit to avoid overly strict conditions that result in too few candidate pairs.

[0111] The first type of parameter space for the percentage of vertically adjacent consecutive frames includes: 0.05 ≤ P y2 A proximity of ≤0.15% in 5% of shared frames indicates that the approach between the two pedestrians is not instantaneous noise, but a meaningful interactive behavior sufficient to trigger the underlying motivation for overtaking. In complex scenarios (such as subway tunnels or pedestrian streets), the initial approach phase of overtaking behavior typically lasts 0.5-1.0 seconds. During this time, the pedestrian behind will gradually approach the pedestrian in front by adjusting their speed or path, creating conditions for subsequent lane changes and acceleration. A proximity of 15% in 15% of shared frames is consistent with the dynamic characteristics of the initial phase of overtaking behavior.

[0112] The second category is when the target interaction behavior is an avoidance behavior. The corresponding initial coarse recognition preset rules include: the second category time overlap threshold is not less than the preset minimum number of frames, the second category lateral approach frame ratio is not less than the second preset frame ratio, and the second preset frame ratio is greater than the first preset frame ratio.

[0113] Furthermore, for the avoidance behavior in the direction of movement, since the avoidance behavior occurs between pedestrians moving in opposite directions, the condition of opposite longitudinal displacements must be satisfied. Using a threshold θ... thr Determine if it is a "opposite" flow. When there is multi-directional pedestrian flow, cosθ must be satisfied.thr ∈[-1,0]; when there is only opposing pedestrian flow, cosθ needs to be satisfied. thr ∈[-1,-0.7].

[0114] In this embodiment, the second type of time overlap threshold is no less than 4 seconds. This is because the typical process of avoidance behavior usually requires a certain amount of time to fully unfold. Avoidance behavior typically includes stages such as approaching each other, sensing and making decisions, executing trajectory deviation, completing avoidance, and returning to the normal trajectory. From the moment both parties become aware of the potential collision risk, to one or both parties adjusting their speed and direction to achieve lateral offset, until finally eliminating the risk of conflict and gradually returning to the desired travel state, this series of dynamic adjustments usually takes several seconds to complete. Therefore, setting a time overlap threshold of no less than 4 seconds aims to ensure that the analyzed interaction trajectory segments can fully contain the core dynamic features of avoidance behavior, thereby improving the accuracy and reliability of identification. Fixing this parameter also helps simplify the subsequent parameter optimization problem and ensures that there is sufficient observation data to capture key changes during the avoidance process.

[0115] Furthermore, the second preset frame percentage is preferably 20%, meaning that the two pedestrians must meet at least 20% lateral proximity in shared frames. To ensure the detection of avoidance behavior with significant interactive characteristics, the 20% threshold can effectively distinguish between transient random behavior and genuine avoidance behavior. Compared to overtaking behavior, avoidance behavior involves more path conflicts laterally. The 20% threshold reflects the persistence requirement of this behavior, which helps improve detection accuracy while avoiding misjudgments caused by excessively low thresholds, thus adapting to behavior pattern analysis in dynamic scenes.

[0116] For the selection of candidate pairs for the first stage of avoidance behavior, the time dimension mainly considers the degree of overlap of pedestrian time to ensure that the time of the two pedestrians in the video sequence is sufficient to capture the dynamic features of the avoidance process; the spatial dimension focuses on the direction of movement and the degree of lateral proximity to capture possible avoidance behaviors in opposite or lateral movements.

[0117] The initial rough identification of rules to be optimized includes: the second type of horizontal spacing threshold.

[0118] The first preset rule parameter space includes the parameter space of the second type of horizontal spacing threshold D2, and the parameter space of the second type of horizontal spacing threshold D2 includes: d1≤D2≤H.

[0119] S620 uses a Bayesian optimization algorithm and a coarse recognition dynamic optimization strategy to optimize the initial coarse recognition rules to be optimized, and uses the initial coarse recognition preset rules and the optimized initial coarse recognition rules to be optimized as the target coarse recognition rules.

[0120] Bayesian optimization algorithms can efficiently search for the optimal combination of parameters in the parameter space by utilizing historical data and prior knowledge, reducing the number of attempts in the optimization process and improving optimization efficiency. By optimizing the initial coarse recognition rules, the accuracy and reliability of the target coarse recognition rules can be improved, enabling them to better identify pedestrian interaction behaviors.

[0121] In this embodiment, the preset rule parameter space can effectively limit the search range of the Bayesian optimization algorithm. This avoids the algorithm from searching on meaningless or unreasonable parameter values, improving search efficiency and reducing the waste of computing resources. At the same time, it provides the algorithm with a reasonable search direction, enabling the algorithm to optimize within a meaningful parameter space, thus making it easier to find the optimal parameter values ​​that match the target scenario. In addition, it can also ensure that the optimization results do not exceed the feasible range of practical applications, improving the practicality of the optimization results.

[0122] S630, determine the dynamic optimization strategy for the initial refined rules to be optimized based on the target scenario. The dynamic optimization strategy includes the second preset rule parameter space of the initial refined rules to be optimized.

[0123] Determining the parameter space for the initial refined identification rules allows the rule depth to better adapt to the characteristics of the target scenario, making the refined identification rules more detailed and complex. Through accurate parameter space settings, key information in the scenario can be better extracted. The second preset rule parameter space provides a foundation for subsequent optimization of the refined identification rules, ensuring that the optimization process is carried out within a reasonable range.

[0124] The initial refined identification rules fall into two categories: the first is for when the target interaction behavior is an overtaking behavior, and the second is for when the target interaction behavior is an avoidance behavior. The second-stage identification rules for overtaking behavior aim to conduct a detailed analysis of the candidate pairs selected in the first stage to confirm genuine overtaking behavior. In the time dimension, speed and lane-change duration are considered to reflect the dynamic adjustment process during overtaking; in the spatial dimension, lateral spacing and longitudinal position are considered to verify the overtaking effect. This stage ensures the accuracy and reliability of the identification results through the collaborative verification of multi-dimensional indicators. The second-stage identification rules for avoidance behavior aim to conduct a detailed analysis of the candidate pairs selected in the first stage to confirm genuine avoidance behavior. In the time dimension, dynamic adjustment characteristics during avoidance are captured by analyzing speed changes to reflect the continuity of the pedestrian avoidance process; in the spatial dimension, the trend of lateral spacing changes is used to assess whether there is effective spatial transfer between pedestrians.

[0125] The first category is when the target interaction behavior is an overtaking behavior. The corresponding initial refined rules to be optimized include the longitudinal speed lead time threshold, the lateral spacing range threshold, the longitudinal position lead time and proportion threshold, and the lane change duration.

[0126] The second preset rule parameter space includes: the parameter space of the velocity window n, and the first minimum lateral spacing X. 1min The parameter space, the first maximum lateral spacing X 1max The parameter space of the vertical leading window k, the parameter space of the vertical leading percentage P y3 Parameter space, lane change duration T lane Parameter space;

[0127] Wherein, 25f≤n≤50f, the speed window size range is set based on the typical duration of speed changes during overtaking. 25 frames (based on a frame rate of 25) is the minimum observation window for speed changes during lane changes; 50 frames is the typical duration of overtaking from lane change to leading, covering the acceleration and stabilization phases.

[0128] d1≤X 1min ≤H, preferably, 0.4≤X 1min ≤0.6, the minimum lateral spacing is based on the safe walking distance of pedestrians, and 0.4m is the minimum value of the pedestrian's shoulder width, ensuring the most basic anti-collision space.

[0129] Z2≤X 1max ≤Z3, where Z3 is the maximum lateral offset threshold after lane change during overtaking; preferably, 1.0≤X 1max ≤1.6, the maximum lateral spacing is based on the interaction distance in a walking scenario. In a relatively spacious environment, the separation distance between pedestrians is approximately 1.0m. 1.6m is the maximum possible lateral offset after changing lanes during overtaking; beyond this range, overtaking interaction is generally not possible.

[0130] Among them, 25f≤k≤75f, the vertical leading window size range is based on the continuous observation requirements after the overtaking behavior. 25f is the minimum observation time for leading, avoiding misjudgment of instantaneous overtaking; 75f is the typical duration of stable leading after overtaking behavior, ensuring the reliability of the results.

[0131] Where, 0.7≤P y3 ≤0.9; The percentage of vertical leading performance reflects the sustainability of the breakthrough. 70% is the minimum requirement to ensure that the leading performance is not accidental; 90% is the upper limit to avoid being too strict and causing missed detections.

[0132] Where, 0.5≤T lane≤3 seconds; The duration of a lane change must be within a certain range to ensure the authenticity and rationality of the lane change behavior, confirm that the potential overtaking party has performed a lateral avoidance action consistent with the overtaking intention, and exclude brief, meaningless lateral jerking. The lane change duration refers to the time elapsed from the start to the end of the overtaking party's lateral path adjustment (lane change). It is used to measure the continuity and rationality of the lane change behavior, reflecting the dynamic process of the overtaking party completing the path switch, and is an important spatiotemporal characteristic of the overtaking behavior. The lane change duration is based on the kinematic characteristics of pedestrian walking. 0.5 seconds is the minimum duration for a lane change to avoid instantaneous jerking; 3 seconds is the maximum reasonable duration for a lane change. Exceeding this range may indicate atypical behavior (such as prolonged deviation after path adjustment).

[0133] The second category is when the target interaction behavior is an avoidance behavior, the corresponding initial refined rules to be optimized include: horizontal spacing range threshold, horizontal speed change time and proportion threshold, where the horizontal speed change value is a preset threshold;

[0134] The second preset rule parameter space includes: the second minimum horizontal spacing X 2min The parameter space, the second maximum lateral spacing X 2max The parameter space of the lateral velocity change window m, and the percentage of lateral velocity change ρ. x The parameter space.

[0135] Second minimum lateral spacing X 2min The parameter space includes: d1≤X 2min ≤H, preferably, 0.4≤X 2min The minimum lateral spacing of ≤0.6m is based on the safety requirements of pedestrian avoidance at the moment of minimum longitudinal spacing. 0.4m is the minimum value of pedestrian shoulder width, which ensures that oncoming pedestrians can avoid physical contact by adjusting laterally when they are closest, reflecting the most basic spatial avoidance requirements of avoidance behavior. The upper limit of 0.6m takes into account the typical safe buffer distance between pedestrians in dense walking scenarios. Based on behavioral observation data, this range is sufficient to cope with collision risks, while avoiding excessive restriction of natural interaction.

[0136] Second maximum lateral spacing X 2max The parameter space includes: Z2≤X 2max ≤Z4, where Z4 is the typical interval of the dividing lines in the pedestrian passage of the target scene; preferably, 1.0≤X 2maxThe maximum lateral spacing is ≤1.5, which is based on the necessity of pedestrian avoidance behavior at the moment of minimum longitudinal spacing. 1.0m is a common personal space boundary in pedestrian environments; the upper limit of 1.5m corresponds to the typical spacing of dividing lines in pedestrian passages (such as subway passages). Beyond this distance, oncoming pedestrians usually do not need to make significant avoidance, and the risk of collision is reduced. This range ensures that the lateral spacing of avoidance behavior is limited within a reasonable interaction range, reflecting the dynamic characteristics of pedestrians completing avoidance through lateral movement in high-density scenarios.

[0137] The parameter space for the lateral velocity variation window *m* includes: 15f ≤ *m* ≤ 35f. The size range of the lateral velocity variation window is set based on the time scale of the avoidance behavior. 15f is the minimum observation time to avoid misjudgment due to instantaneous deviation; 35f is the time to ensure the stability of the avoidance behavior and the reliability of the results.

[0138] The percentage of lateral velocity change ρ x The parameter space includes: 0.7≤ρ x ≤0.95, the percentage of lateral velocity change reflects the dynamic process and intensity of avoidance behavior. 70% is the minimum requirement, indicating the continuity of the pedestrian's active avoidance intention and avoiding misjudgment due to noise or random fluctuations. The upper limit of 95% takes into account the diversity of avoidance behaviors in real-world scenarios, allowing some frames to not reach the threshold due to trajectory noise or atypical adjustments (such as short pauses), while avoiding overly strict judgments that lead to missed detections.

[0139] In this type of behavior, the threshold for lateral velocity change per frame during avoidance is set to a fixed value of 0.05 m / s. Avoidance behavior typically occurs when there is a potential path conflict between pedestrians, who need to adjust their lateral position to avoid collision. The 0.05 m / s threshold reflects the reasonable reaction range for pedestrians at normal walking speeds. Based on the typical speed of pedestrians (approximately 1-1.5 m / s) and the fine-tuning characteristics during avoidance, this value effectively captures the lateral movement changes caused by pedestrians during avoidance, without mistaking random swaying or noise for avoidance behavior.

[0140] S640 uses a Bayesian optimization algorithm and a dynamic optimization strategy to optimize the initial refined rules, and uses the initial refined preset rules and the optimized initial refined rules as the target refined rules.

[0141] In the optimization of precise identification rules, the Bayesian optimization algorithm can find the optimal parameter combination more accurately, further improving the recognition accuracy of the rules. Precise identification rules are inherently complex, and this optimization method can fully leverage the potential of the rules. Combining the optimized precise identification rules with the preset rules forms a complete target precise identification rule, making the entire rule system more complete and enabling more accurate identification of pedestrian interaction behavior.

[0142] By dynamically optimizing the coarse and fine recognition rules respectively, the performance of the rules can be comprehensively improved. From the initial coarse recognition to the fine recognition, the rules can better adapt to the needs of the target scenario. The optimized coarse and fine recognition rules can more accurately identify pedestrian interaction behavior, reduce false positives and false negatives, and improve the reliability of the system. The solution has strong adaptability and scalability. For different target scenarios, it is only necessary to redefine the parameter space and optimize it to obtain the target dynamic recognition rules applicable to the scenario.

[0143] Specific reference Figure 4 Bayesian optimization is a sequential, model-based global optimization strategy suitable for optimizing "black box" functions. Its core idea is to construct a surrogate model of the objective function (such as a Gaussian process) and intelligently select the next parameter point to be evaluated using a sampling function, thus approximating the global optimum with a finite number of evaluations.

[0144] In pedestrian interaction behavior recognition, the basic principle of using Bayesian optimization to find the optimal parameter combination of the rule system is as follows: based on a positive sample dataset of known scenarios, combined with a two-stage pedestrian recognition rule system, the positive sample recognition rate is maximized through an iterative process. The core steps include:

[0145] ① Constructing a surrogate model: Based on the evaluated parameter combinations and their corresponding recognition performance, a probabilistic surrogate model is constructed to estimate the recognition performance of positive sample data across the entire parameter space. Commonly used surrogate models include Gaussian Process (GP) and Random Forest (RF). ② Defining a data collection function: The data collection function evaluates which parameter combinations are most likely to improve recognition performance under the current model based on the predicted mean and variance of the surrogate model, thus selecting the parameter combinations to be evaluated in the next round. The design of the data collection function needs to strike a balance between exploration and development. Commonly used data collection functions include Expected Improvement (EI), Probabilistic Improvement (PI), and Upper Confidence Bound (UCB). ③ Iterative optimization: In each iteration, a new parameter combination is selected by maximizing the data collection function. Recognition tests are performed on positive sample data to obtain recognition results and update the surrogate model. This process continuously approaches the optimal parameter set until the termination condition is met.

[0146] Optimization of behavior recognition parameters includes:

[0147] Where, θ * Let f(θ) be the optimal parameter combination, Θ be the parameter space, f(θ) be the objective function and also the recognition effect function, and θ be the parameter combination to be optimized.

[0148] For overtaking behavior, θ and θ* form an 8-tuple: θovertake = (lateral spacing threshold, vertical spacing threshold, percentage of vertical approach frames / continuous frames, vertical speed lead window, lateral spacing range threshold, vertical lead window, vertical lead percentage, lane change duration); for avoiding behavior, θ and θ* form a 4-tuple: θavoid = (lateral spacing threshold, lateral speed change window, lateral speed change percentage, lateral spacing range threshold).

[0149] This method combines two-stage behavior recognition rules with a Bayesian optimization algorithm. It utilizes a dataset D of manually labeled positive samples of overtaking or avoidance behaviors in the scene to be recognized. The optimal parameter combination is found iteratively through the Bayesian optimization algorithm to maximize the positive sample recognition rate while satisfying the behavior recognition rules. The main implementation steps are as follows:

[0150] ① Define the optimization problem: The goal of the optimization is to maximize the behavior recognition rate of the dataset. The objective function is defined as: Where TP is the number of correctly identified behaviors, and N is the total number of samples in the dataset (all of which are positive samples).

[0151] ② Define parameter space

[0152] The feasible region of the parameter vector is constructed based on the parameters to be optimized at each stage and their search range (boundary constraints). This mainly includes the parameter space of the aforementioned candidate pairs for overtaking and avoiding pedestrians, and the parameter space for behavior determination.

[0153] Specifically, the algorithm implementation process includes:

[0154] ① Initialization parameters and dataset

[0155] Step 0: Prepare a training dataset D of positive samples of overtaking / avoidance behavior in the target scenario, ensuring that the dataset contains only confirmed and real overtaking / avoidance behavior samples in the target scenario.

[0156] Step 1: Randomly select N within the parameter space Θ init Parameter combinations (N recommended) init =10), denoted as θ1,θ2,…, Random sampling uses a uniform distribution to ensure that the initial points cover different regions of the parameter space and avoid initial bias.

[0157] Step 2: For each parameter combination θ, run the behavior recognition algorithm to calculate the objective function value f(θ) based on the given dataset.

[0158] Step 3: Construct the initial dataset D0 = {(θ} based on the recognition results. i ,f(θ i ))∣i=1,2,...,N init}

[0159] ② Constructing a proxy model

[0160] Based on the currently evaluated combination of rule parameters and its corresponding dataset D0={(θ i ,f(θ i ))∣i=1,2,...,N init A proxy model is constructed to learn the complex mapping relationship between pedestrian interaction behavior rule parameters and positive sample detection, which is used to estimate the objective function f(θ). i The distribution of parameters throughout the parameter space.

[0161] ③ Calculate the acquisition function and select a new parameter combination

[0162] The acquisition function selects the next most promising combination of rule parameters to improve the detection rate of interactive behavior based on the predictions of the proxy model and evaluates them.

[0163] Select a new parameter combination by optimizing the acquisition function:

[0164] ④ Evaluate the new parameter combinations and update the dataset

[0165] Calculate the target value for identifying interactive behaviors using the new parameter combination, and update the dataset D0 = {(θ i ,f(θ i ))∣i=1,2,…,N init +1}.

[0166] ⑤ Determine the termination condition: 1) Reach the maximum number of iterations N max That is, the total number of evaluations is N. init +N max ;2) Recognition rate convergence, that is, the improvement of the recognition rate in multiple consecutive iterations is less than a certain threshold.

[0167] ⑥ Output the optimal parameter combination: Specifically, select the parameter combination with the highest recognition rate from all evaluated parameter points:

[0168]

[0169] Output the optimal parameter combination θ * The recognition rate f(θ) of positive samples of interaction behavior in the corresponding scenario. The final optimal parameter combination θ is obtained. * This represents a set of rule parameters highly optimized for the target scenario, maximizing the sensitivity of interactive behavior recognition within that scenario. These parameters will then be used to configure the recognition model for analyzing all data from that scenario.

[0170] For interactive behavior recognition, the specific steps include: applying this pedestrian interactive behavior recognition method in a complex scenario. First, data preprocessing and positive sample extraction are performed based on the data from the specific scenario. Then, a Bayesian optimization algorithm is used to train and obtain the optimal parameter combination suitable for the scenario. The optimized parameter combination for this specific scenario is configured into a two-stage recognition rule model to obtain a dynamic pedestrian behavior recognition strategy for that scenario. This strategy is then applied to the trajectory data of the target scenario to identify overtaking and avoidance behaviors. Finally, the relevant information of all identified overtaking and avoidance behaviors in the target scenario is output.

[0171] By analyzing the final output of overtaking and yielding behavior events, especially the spatial distribution, frequency, and accompanying characteristics (such as whether relative speed fluctuates significantly), managers can accurately locate congestion points and bottleneck areas where overtaking and yielding are difficult or frequent due to speed differences and path selection. Based on this precise micro-level analysis, more effective optimization measures can be developed and evaluated, such as optimizing signage and implementing dynamic pedestrian flow management strategies, thereby reducing unnecessary mutual interference and queuing, improving the overall traffic efficiency and utilization of the space, and enhancing the pedestrian experience.

[0172] Identified overtaking and yielding behaviors (especially those occurring in crowded conditions, with large lateral movements, or in confined spaces) are indicators of potential conflicts and risks. Analyzing areas with high rates of overtaking and yielding behaviors can directly pinpoint potential safety hazards. Real-time or near-real-time monitoring of surges in pedestrian yielding and interaction behaviors or trends in dangerous characteristics can serve as early warning signals of excessive crowd density, deteriorating mobility, or even the potential for stampedes and other dangerous events. This risk perception capability based on micro-behavior enables managers to take preventative intervention measures in advance, such as promptly deploying security or guidance personnel in identified high-risk areas, issuing safety warnings, or implementing temporary crowd control / diversion measures, thereby effectively reducing the probability of accidents and ensuring the safety of pedestrians' lives and property.

[0173] Reference Figure 5 The S700 method of "identifying pedestrian movement trajectory data in a target scene based on target dynamic recognition rules and determining pedestrian interaction behavior," specifically includes:

[0174] S710, perform coarse analysis on pedestrian movement trajectory data in the target scene according to the target coarse recognition rule, and screen out target candidate pairs with potential target behaviors; target candidate pairs with potential target behaviors include first-class candidate pairs with potential overtaking behavior and second-class candidate pairs with potential avoidance behavior.

[0175] In this embodiment, the target coarse identification rule is a relatively lenient rule. In the first stage, the relatively lenient rule can quickly filter out potential pedestrian candidate pairs. The goal of this stage is to identify those pedestrian pairs with sufficient spatiotemporal interaction potential from all pedestrian trajectories as the objects of subsequent fine analysis, thereby accelerating computational efficiency.

[0176] S720 analyzes target candidate pairs based on target recognition rules to determine pedestrian interaction behavior.

[0177] Among them, pedestrian interaction behavior information includes overtaking behavior information and / or avoidance behavior information.

[0178] The coarse identification strategy can quickly perform preliminary screening of large amounts of data, eliminating obviously unsuitable data and reducing the workload of subsequent fine identification. By using the initial range of the parameters to be optimized and preset rules in the optimization rules, coarse identification can filter out most invalid data in a short time, improving recognition efficiency. The fine identification strategy uses optimized dynamic rules and preset rules to further filter the results after coarse identification, enabling more accurate identification of target interaction behaviors. The dynamic rules are optimized according to the target scenario and can better adapt to the characteristics of the target scenario, thereby improving the accuracy and reliability of recognition. By combining the coarse and fine identification strategies, a dynamic recognition strategy is formed. This strategy can dynamically adjust the coarse and fine identification rules according to changes in the target scenario and data updates, ensuring that the recognition strategy always has high performance and adaptability.

[0179] This application designs a hierarchical pedestrian recognition strategy of "coarse recognition - fine recognition", which significantly improves the computational efficiency and real-time performance of the algorithm, enabling the pedestrian recognition system to handle larger-scale and more complex scenarios and making it feasible for practical applications.

[0180] By introducing a Bayesian optimization algorithm to overcome the limitations of traditional fixed thresholds, and adaptively searching for the optimal parameter combination based on scene trajectory data, dynamic threshold adjustment is achieved. This method significantly enhances the adaptability of the recognition system to different scenarios, ensuring optimal recognition results in complex parameter spaces and reducing false positives and false negatives. The two-stage screening rules and parameter threshold optimization are systematically integrated to construct a complete pedestrian interaction behavior recognition system from data collection and processing to final behavior determination. This technical solution can provide efficient and accurate data support for urban traffic management, public safety monitoring, and social behavior research, demonstrating broad practical value.

[0181] In a complex scenario, this pedestrian interaction behavior recognition method is applied. First, data preprocessing and positive sample extraction are performed based on the scenario-specific data. Then, a Bayesian optimization algorithm is used to train and obtain the optimal parameter combination suitable for the scenario. The optimized parameter combination for this specific scenario is configured into a two-stage recognition rule model to obtain a dynamic pedestrian behavior recognition strategy for that scenario. This strategy is then applied to the trajectory data of the target scenario to identify overtaking and avoidance behaviors. Finally, the relevant information of all identified overtaking and avoidance behaviors in the target scenario is output.

[0182] By analyzing the final output of overtaking and yielding behavior events, especially the spatial distribution, frequency, and accompanying characteristics (such as whether relative speed fluctuates significantly), managers can accurately locate congestion points and bottleneck areas where overtaking and yielding are difficult or frequent due to speed differences and path selection. Based on this precise micro-level analysis, more effective optimization measures can be developed and evaluated, such as optimizing signage and implementing dynamic pedestrian flow management strategies, thereby reducing unnecessary mutual interference and queuing, improving the overall traffic efficiency and utilization of the space, and enhancing the pedestrian experience.

[0183] Reference Figure 6 and Figure 7 To balance recognition efficiency and accuracy, this method employs a two-stage strategy to formulate recognition rules for pedestrian interaction behaviors. Pedestrian interaction behaviors are divided into overtaking and yielding behaviors, considered separately. This mainly includes screening candidate pairs for overtaking and yielding, and analyzing the overtaking and yielding behaviors of these candidate pairs. A rule parameter space is systematically constructed for the key judgment indicators in the two-stage recognition rules. This space includes the parameters to be optimized for each stage of overtaking and yielding behaviors, as well as the reasonable value ranges for these parameters. A Bayesian optimization algorithm is used to adaptively and data-drivenly optimize this rule parameter space. By learning and iteratively optimizing positive sample data of pedestrian trajectories in specific target scenarios, the method efficiently searches for and determines the parameter combinations that most accurately identify overtaking and yielding behaviors in the current target scenario. This overcomes the limitations of traditional fixed thresholds and significantly enhances adaptability, robustness, and recognition performance in various complex scenarios.

[0184] Overtaking behavior is the act of pedestrians moving in the same direction actively adjusting their walking speed, path, and other spatiotemporal parameters to shorten travel time or optimize their route, based on the need to avoid speed conflicts with slower individuals or obstacles ahead. Specifically, it manifests as a dynamic spatial adjustment process where, based on environmental parameters, the pedestrian accelerates to increase relative speed and deviates from their route to overtake, subsequently returning to their original trajectory or continuing their journey.

[0185] Avoidance behavior is the act of pedestrians actively adjusting their kinematic variables, such as walking speed, direction, and path, in situations of oncoming or lateral movement to avoid collisions with other pedestrians or obstacles, based on the need to ensure safe passage. Specifically, it manifests as a dynamic spatial adjustment process in which pedestrians, based on perceived environmental parameters such as relative distance and speed, change their route by decelerating or lateral displacement to bypass obstacles and then return to their original trajectory or continue moving.

[0186] Overtaking and avoidance behaviors, as two typical modes of pedestrian interaction, are constrained by both temporal continuity and spatial geometric and dynamic characteristics. Based on the motion characteristics of overtaking and avoidance behaviors, and taking time and space as core considerations, a systematic set of recognition rules is formulated. By deeply analyzing the inherent patterns of pedestrian movement in complex scenarios, and utilizing design rules to comprehensively capture pedestrian trajectory interaction behaviors from both temporal and spatial dimensions, efficient and accurate pedestrian behavior recognition is achieved.

[0187] The overtaking and yielding behaviors identified by the methods disclosed in this application (especially those occurring in crowded conditions, with large lateral movements, or in confined spaces) are indicators of potential conflicts and risks. Analyzing areas with high incidence of overtaking and yielding behaviors allows for the direct identification of potential safety hazards. Real-time or near-real-time monitoring of surges in pedestrian yielding and overtaking behaviors or changes in dangerous characteristics can serve as early warning signals of excessive crowd density, deteriorating mobility, or even the potential for stampedes and other dangerous events. This risk perception capability based on micro-behavior enables managers to take preventative intervention measures in advance, such as promptly deploying security or guidance personnel in identified high-risk areas, issuing safety warnings, or implementing temporary crowd control / diversion measures, thereby effectively reducing the probability of accidents and ensuring the safety of pedestrians' lives and property.

[0188] Reference Figure 8 The pedestrian avoidance and overtaking behavior recognition technology system disclosed in this application includes five key steps:

[0189] ① Video data acquisition: Specifically, high-definition cameras, surveillance networks and other equipment can be used to record video of the target scene from multiple angles for a long time to ensure coverage of the entire life cycle of pedestrian behavior.

[0190] ② Pedestrian recognition and trajectory tracking: Real-time detection and continuous trajectory extraction of pedestrian targets can be achieved by using computer vision algorithms (such as YOLO series detection models combined with DeepSORT tracking algorithm).

[0191] ③ Coordinate transformation: Convert the video pixel coordinate system to the real-world coordinate system to ensure that the trajectory data has a real spatial scale, so as to ensure that the thresholds obtained later are judgmental and operable.

[0192] ④ Trajectory data preprocessing: The original trajectory obtained from tracking is preprocessed to remove invalid data segments caused by occlusion and algorithm errors (such as trajectory breaks, drift points, and excessively short trajectories), and retain motion trajectories that meet the research requirements and are of reliable quality, so as to provide clean input for behavior recognition.

[0193] ⑤ Systematic Recognition of Interactive Behaviors: A two-stage overtaking / avoidance behavior screening rule is formulated using a "coarse recognition-refined recognition" strategy. The rule parameter space is determined, and a behavior recognition effect evaluation system is constructed. The Bayesian optimization algorithm is used to search for the optimal combination of rule parameter thresholds for the target scene in the rule parameter space and feeds it back to the two-stage recognition rule. An overtaking / avoidance behavior recognition strategy suitable for the target scene is constructed and applied to the preprocessed real trajectory data. The logic engine determines the behavior occurrence conditions frame by frame and finally outputs the recognition results of pedestrian overtaking, avoidance and other interactive behaviors, recording the interaction individual's ID, event type and the time of occurrence of the entire process.

[0194] The first three stages constitute the basic pedestrian trajectory data processing module, enabling a systematic transformation from raw video to structured trajectory data. The latter two stages form the core pedestrian behavior recognition module. Leveraging a data-driven model and rule engine, it learns behavioral characteristic thresholds for dynamic scenarios based on labeled data and combines this with real-time trajectory data to intelligently identify and output decisions regarding pedestrian overtaking and avoidance behaviors, thus constructing a complete behavior recognition chain. By combining data-driven threshold learning with rule-based logical judgment, the scenario adaptability and accuracy of behavior recognition are significantly improved.

[0195] In existing technologies, overtaking and avoidance behaviors overlap and collinearize with other behaviors in terms of spatiotemporal trajectory features, leading to decreased inter-class separability and high false positive rates in traditional machine learning framework behavior classification models. This application discloses a method for recognizing pedestrian interaction behaviors in complex scenarios, which dynamically acquires a dynamic recognition strategy for the target scene based on a Bayesian optimization algorithm. The Bayesian optimization algorithm can optimize parameters according to the target scene, finding recognition rules and parameters more suitable for that scene, improving feature discriminative power, enhancing inter-class separability, thereby reducing the false positive rate and more accurately recognizing pedestrian interaction behaviors.

[0196] Pedestrian interaction behavior exhibits multimodal evolutionary characteristics, which traditional methods struggle to address. This solution employs a dynamic recognition strategy, which adjusts recognition rules dynamically based on the characteristics and changes of different target scenarios. This better adapts to the multimodal evolution of pedestrian interaction behavior, improving the accuracy and stability of recognition.

[0197] Traditional microscopic simulation methods rely on idealized assumptions, making them difficult to adapt to real-world scene data and lacking in generalization ability. This approach collects pedestrian behavior video within a preset period in the target scene and converts the continuous trajectory data from the video pixel coordinate system to the real-world coordinate system, obtaining accurate original trajectory data. Subsequent analysis and recognition based on this real data can better adapt to different real-world scenarios, enhancing the method's generalization ability.

[0198] Trajectory feature engineering-based analysis methods belong to shallow representation learning, which cannot establish a quantitative description of the implicit dynamic mechanisms of pedestrian interactions and lacks sufficient behavior decoupling ability. This scheme, based on processing raw trajectory data, uses a Bayesian optimization algorithm to dynamically optimize the recognition strategy, potentially uncovering the implicit dynamic mechanisms behind pedestrian interactions, achieving effective decoupling of different behaviors, and improving the understanding and analysis capabilities of pedestrian interactions.

[0199] This solution forms a complete automated process, from collecting pedestrian behavior video, acquiring trajectory data, transforming coordinate systems, preprocessing data, to behavior recognition based on dynamic recognition strategies. This reduces manual intervention and improves recognition efficiency and consistency. Throughout the process, by processing video data, transforming coordinate systems, and optimizing dynamic recognition strategies, it can more accurately acquire pedestrian movement trajectories and behavioral information, thus improving the accuracy of pedestrian interaction behavior recognition.

[0200] The pedestrian interaction behavior recognition method disclosed in this application accurately identifies pedestrian interaction behaviors in complex scenarios, which helps to deeply understand pedestrian movement patterns and behavioral patterns, and provides a basis for the planning and design of complex scenarios such as urban pedestrian areas, transportation hubs, and shopping malls. Through the analysis and understanding of pedestrian interaction behavior, human activity spaces can be optimized, such as by rationally planning pedestrian passages and setting up traffic facilities, thereby improving space utilization efficiency and safety, and promoting the smooth flow of human activities.

[0201] Reference Figure 9 In one specific embodiment, the pedestrian interaction behavior recognition method includes: inputting pedestrian trajectory data, determining whether the data length meets a preset condition (preferably determining whether the data length is greater than 4s), if not, filtering out the trajectory; if so, determining whether the longitudinal displacement meets a preset condition (preferably determining whether the longitudinal displacement is greater than 2m), if not, filtering out the trajectory.

[0202] If the longitudinal displacement meets the preset conditions, the preprocessed data is output. The next step is to analyze the data according to the two-stage behavior screening rules, which includes: calculating the candidate pair characterization index and then judging the longitudinal direction of movement.

[0203] When the longitudinal forward movement is in the same direction, i.e., when the movement is in the same direction, it is determined whether the time overlap meets the preset condition (preferably, whether the time overlap is greater than 4f). ps If not, discard the candidate pair; if yes, determine if the lateral approach condition is met; if not, discard the candidate pair. If the lateral approach condition is met, obtain a set of overtaking pedestrian candidate pairs, and then calculate the overtaking behavior characterization indicators (including change point, average longitudinal speed, and lateral spacing). Next, determine if the lane change time constraint is met; if not, determine it as a non-overtaking behavior; if yes, determine the V within the time window. 超 Is it greater than V? 被超 If not, it is determined to be a non-overtaking behavior. If yes, it is determined to satisfy the vertical leading constraint. If not, it is determined to be a non-overtaking behavior. If yes, it is determined to satisfy the horizontal spacing constraint. If not, it is determined to be a non-overtaking behavior. If yes, it is output parameters and rule set.

[0204] When the longitudinal forward direction is opposite, i.e., when it is opposing motion, determine whether the time overlap meets the preset condition (preferably, determine whether the time overlap is greater than 4f). ps If not, discard the candidate pair; if yes, determine if the lateral approach condition is met; if not, discard the candidate pair. If the lateral approach condition is met, obtain a set of pedestrian avoidance candidate pairs, and then calculate the avoidance behavior characterization indicators (including lateral spacing and lateral velocity change). Next, determine if the lateral spacing constraint is met; if not, determine it as a non-avoidance behavior; if yes, determine if the lateral velocity change constraint is met; if not, determine it as a non-avoidance behavior; if yes, output the parameters and rule set.

[0205] The Bayesian optimization algorithm is used to optimize the output parameters and rule set to obtain optimized rule information. Then, based on the optimized rule information, the overtaking / avoidance behavior is accurately identified, and the overtaking / avoidance information of pedestrians is output.

[0206] Furthermore, the obtained information on overtaking / avoiding pedestrians can be used as a new information update sample space, thereby updating the positive samples of interactive behavior in the sample data.

[0207] In this application, the recognition performance includes precision, recall, and F1 score. Precision, in particular, measures the proportion of correctly identified interactions. It is defined as: TP represents correctly identified overtaking / avoidance behavior; FP represents non-interactive behavior that is misidentified as overtaking / avoidance behavior.

[0208] Among them, recall is the proportion of samples that were correctly predicted out of the total number of samples that were actually overtaking / avoiding behaviors. FN represents unidentified overtaking / avoidance behavior.

[0209] The F1 score is the harmonic mean of precision and recall, balancing the two.

[0210] In this application, a comparative analysis of Bayesian optimization algorithms is presented, taking pedestrian overtaking behavior recognition as an example. This analysis aims to evaluate the effectiveness of Bayesian optimization using different surrogate models (Gaussian process, random forest) and acquisition functions (expectation improvement, probability improvement, upper confidence bound) in pedestrian overtaking behavior recognition methods in complex scenarios, and to quantify their impact on recognition performance.

[0211] (1) Proxy Model

[0212] ① Gaussian Process (GP): The Gaussian process assumes that the objective function follows a Gaussian distribution. It models the joint distribution of function values ​​using a kernel function to predict the function value f(x) and uncertainty at point x. ② Random Forest (RF): The Random Forest uses multiple decision trees to model the objective function. Each tree predicts a value, and the average is used as an estimate of the function value, while the variance is used as an estimate of the uncertainty.

[0213] (2) Acquisition function

[0214] ① Expected Improvement (EI): EI balances exploration (high uncertainty region) and exploitation (high predicted value region), tending towards global search. ② Probabilistic Improvement (PI): Probabilistic Improvement favors exploitation, i.e., selecting points with higher predicted values, but its drawback is the potential to get trapped in local optima. ③ Upper Confidence Bound (UCB): The Upper Confidence Bound combines the predicted mean and uncertainty, selecting the point with the highest score.

[0215] (3) Comparative analysis: Refer to Figure 10 Table 1 shows a comparison of the recognition performance of different proxy models and acquisition functions.

[0216] Table 1. Recognition performance of different proxy models and acquisition functions

[0217]

[0218] In the comparison of surrogate models, Expected Improvement (EI) was chosen as the sole acquisition function primarily because EI achieves the optimal balance between exploration (in regions of high uncertainty) and exploitation (in regions of high predicted values), possessing strong global search capabilities and effectively avoiding getting trapped in local optima. This characteristic of Expected Improvement ensures that both Random Forest and Gaussian Process can fully explore the parameter space during optimization, reducing the impact of the acquisition function on the performance of the surrogate model, thus more fairly reflecting the differences in modeling capabilities between Gaussian Process and Random Forest.

[0219] As shown in Table 1, the Gaussian process's smooth modeling ability results in excellent recall, effectively identifying most transcendental behaviors. However, its precision is lower, possibly due to the discontinuity of the objective function, which limits the Gaussian process's modeling ability and leads to a higher false positive rate. Random forests significantly outperform Gaussian processes in precision, indicating a lower false positive rate while maintaining a high recall. The tree structure of random forests is better suited to high-dimensional parameter spaces and discontinuous objective functions, enabling them to more accurately capture transcendental behavior features in complex scenarios.

[0220] Under the Random Forest (RF) proxy model, the performance of different collection functions (EI, PI, and UCB) was analyzed. Expectation Improvement (EI) showed the best overall performance and is suitable for optimizing pedestrian interaction behavior parameters in complex scenarios; Probabilistic Improvement (PI) had high recall, suitable for behavior pattern analysis, but low precision; Upper Confidence Bound (UCB) had good stability and is suitable for scenarios requiring dynamic adjustment between exploration and development.

[0221] The following is a comparative analysis of the recognition performance between fixed parameters and Bayesian optimized parameters:

[0222] (1) Fixed parameter settings: The parameters for overtaking behavior recognition can be set to fixed values ​​based on experience, relevant literature, and field investigations, as follows:

[0223] Table 2 Candidate Pair Values ​​for Fixed Parameters

[0224]

[0225]

[0226] In Bayesian optimization, random forest is selected as the surrogate model and expected improvement is used as the acquisition function for parameter optimization.

[0227] (2) Comparative analysis: Refer to Figure 11 Table 3 provides a comparative analysis of the recognition performance of the two methods.

[0228] Table 3. Identification performance of fixed parameters and Bayesian optimized parameters

[0229]

[0230] from Figure 11As can be seen, compared to using a two-stage screening rule with fixed parameters for traversal behavior identification, the parameters obtained by the Bayesian optimization algorithm significantly outperform the fixed-parameter method. Bayesian optimization improves precision by approximately 51.44%, recall by approximately 45.45%, and F1 score by approximately 48.26% compared to the fixed-parameter method. When using fixed parameters for traversal behavior screening, the lower precision indicates a higher false positive rate; the lower recall indicates a higher false negative rate; and the lower F1 score indicates a difficulty in balancing precision and coverage. Bayesian optimization finds a better parameter combination through global optimization, significantly reducing the false positive and false negative rates and significantly improving the balance. The parameters obtained by Bayesian optimization show more stable performance in traversal behavior identification, indicating that its adaptive adjustment capability reduces performance fluctuations and is suitable for scenarios with different datasets.

[0231] In a second aspect, embodiments of this disclosure provide a pedestrian interaction behavior recognition system for complex scenarios, used to execute the pedestrian interaction behavior recognition method for complex scenarios disclosed in the first aspect of this application, including:

[0232] The acquisition module is used to acquire video of pedestrian behavior within a preset period in the target scene;

[0233] The continuous trajectory data acquisition module is used to acquire continuous trajectory data of each pedestrian from pedestrian behavior videos; the continuous trajectory is a pixel-level motion trajectory.

[0234] The raw trajectory data acquisition module is used to convert the video pixel coordinate system of continuous trajectory data into a real-world coordinate system to obtain the raw trajectory data.

[0235] The preprocessing module is used to preprocess the raw trajectory data to obtain the pedestrian movement trajectory data of the target scene to be identified;

[0236] The initial recognition rule construction module is used to construct the initial recognition rules for pedestrian interaction behavior. The initial recognition rules include initial preset rules and initial rules to be optimized.

[0237] The optimization module is used to determine the dynamic optimization strategy of the initial rule to be optimized based on the target scenario, and optimize the initial rule to be optimized based on the dynamic optimization strategy. The initial preset rule and the optimized initial rule to be optimized are used as the target dynamic recognition rule.

[0238] The recognition module is used to identify pedestrian movement trajectory data in the target scene according to the target dynamic recognition rules and determine pedestrian interaction behavior.

[0239] A computer device according to an embodiment of this disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0240] The processor may be a central processing unit (CPU) or other processing unit with data processing and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the pedestrian interaction behavior recognition method in complex scenarios described in the foregoing embodiments of this disclosure. Those skilled in the art should understand that, in order to solve the technical problem of how to obtain a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0241] like Figure 12 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 12 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0242] like Figure 12 As shown, a computer device may include a processor (e.g., a central processing unit, a graphics processing unit, etc.), which performs various appropriate actions and processes based on programs stored in read-only memory (ROM) or loaded from storage devices into random access memory (RAM). RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus. Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard disks; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 12 A computer apparatus with various devices is shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or included alternatively.

[0243] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processor, all or part of the steps of the pedestrian interaction behavior recognition method in complex scenarios according to embodiments of this disclosure are performed. Detailed descriptions of this embodiment can be found in the corresponding descriptions in the foregoing embodiments, and will not be repeated here. A computer-readable storage medium according to embodiments of this disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the pedestrian interaction behavior recognition method in complex scenarios according to the foregoing embodiments of this disclosure are performed.

[0244] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge). Detailed descriptions of this embodiment can be found in the corresponding descriptions of the foregoing embodiments, and will not be repeated here.

[0245] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0246] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for recognizing pedestrian interaction behavior in complex scenarios, characterized in that, include: Collect pedestrian behavior videos within a preset period in the target scene; Continuous trajectory data for each pedestrian is obtained from the pedestrian behavior video; The continuous trajectory is a pixel-level motion trajectory; The video pixel coordinate system of the continuous trajectory data is converted into a real-world coordinate system to obtain the original trajectory data; The original trajectory data is preprocessed to obtain the pedestrian movement trajectory data of the target scene to be identified; Construct initial recognition rules for pedestrian interaction behavior, wherein the initial recognition rules include initial preset rules and initial rules to be optimized; The dynamic optimization strategy of the initial rule to be optimized is determined according to the target scenario, and the initial rule to be optimized is optimized based on the dynamic optimization strategy. The initial preset rule and the optimized initial rule to be optimized are used as the target dynamic identification rule. Based on the target dynamic recognition rules, the pedestrian movement trajectory data in the target scene is identified to determine pedestrian interaction behavior; The initial recognition rules for constructing pedestrian interaction behavior include: Construct initial coarse recognition rules for pedestrian interaction behavior in the target scenario; the initial coarse recognition rules include initial coarse recognition preset rules and initial coarse recognition rules to be optimized; Based on the initial coarse recognition rules, initial fine recognition rules for pedestrian interaction behavior are constructed; the initial fine recognition rules include initial fine recognition preset rules and initial fine recognition rules to be optimized. The step of determining a dynamic optimization strategy for the initial rule to be optimized based on the target scenario, optimizing the initial rule to be optimized based on the dynamic optimization strategy, and using the initial preset rule and the optimized initial rule to be optimized as the target dynamic identification rule includes: Based on the target scenario, a coarse dynamic optimization strategy for the initial coarse recognition rule to be optimized is determined, and the coarse dynamic optimization strategy includes a first preset rule parameter space for the initial coarse recognition rule to be optimized. The initial coarse recognition rule to be optimized is optimized using the Bayesian optimization algorithm and the coarse recognition dynamic optimization strategy. The initial coarse recognition preset rule and the optimized initial coarse recognition rule to be optimized are used as the target coarse recognition rule. The dynamic optimization strategy for the initial refined rules to be optimized is determined based on the target scenario. The dynamic optimization strategy includes a second preset rule parameter space for the initial refined rules to be optimized. The initial refined knowledge rules to be optimized are optimized using the Bayesian optimization algorithm and the refined knowledge dynamic optimization strategy, and the initial refined knowledge preset rules and the optimized initial refined knowledge rules to be optimized are used as target refined knowledge rules.

2. The method for recognizing pedestrian interaction behavior in complex scenarios according to claim 1, characterized in that, When the target interaction behavior is a transcending behavior, the corresponding initial coarse recognition preset rules include: the number of common frames in which two pedestrians appear in the video at the same time is not less than the preset minimum number of frames, and the proportion of the first type of lateral approach frames is not less than the first preset proportion of frames. The initial coarse identification rules to be optimized include: the first type of horizontal spacing threshold, the first type of vertical spacing threshold, the first type of vertically close frame ratio, and the first type of vertically close consecutive frame ratio; The first preset rule parameter space includes: the parameter space of the first type of horizontal spacing threshold, the parameter space of the first type of vertical spacing threshold, the parameter space of the first type of vertically close frame ratio, and the parameter space of the first type of vertically close consecutive frame ratio; The parameter space for the first type of lateral spacing threshold includes: , This represents the minimum shoulder width for pedestrians. The width of the pedestrian walkway is for one-way traffic. The parameter space for the first type of longitudinal spacing threshold includes: , The minimum perceived distance for changes in pedestrian stride length. This refers to the safe distance between pedestrians and the front and rear pedestrians. The parameter space for the first type of vertical approximation of frame rate percentage includes: ; The parameter space for the first type of vertically approximate proportion of consecutive frames includes: .

3. The method for recognizing pedestrian interaction behavior in complex scenarios according to claim 2, characterized in that, When the target interaction behavior is an avoidance behavior, the corresponding initial coarse recognition preset rules include: the second type of time overlap threshold is not less than the preset minimum number of frames, the proportion of the second type of lateral approach frames is not less than the proportion of the second preset number of frames, and the proportion of the second preset number of frames is greater than the proportion of the first preset number of frames. The initial coarse identification rules to be optimized include: the second type of horizontal spacing threshold; The first preset rule parameter space includes a second type of horizontal spacing threshold. The parameter space, the second type of lateral spacing threshold The parameter space includes: .

4. The method for recognizing pedestrian interaction behavior in complex scenarios according to claim 2, characterized in that, When the target interaction behavior is an overtaking behavior, the corresponding initial precision optimization rules include: longitudinal speed lead time threshold, lateral spacing range threshold, longitudinal position lead time and proportion threshold, and lane change duration. The second preset rule parameter space includes: velocity window Parameter space, first minimum lateral spacing Parameter space, first maximum lateral spacing The parameter space, the parameter space of the vertical leading window k, and the vertical leading percentage. Parameter space, lane change duration Parameter space; Wherein, the speed window The parameter space includes: ; First minimum lateral spacing The parameter space includes: ; First maximum lateral spacing The parameter space includes: , This is the maximum lateral offset threshold after lane change during overtaking behavior; The parameter space of the vertical leading window k includes: ; The vertical leading percentage The parameter space includes: ; The lane change duration The parameter space includes: .

5. The method for recognizing pedestrian interaction behavior in complex scenarios according to claim 4, characterized in that, When the target interaction behavior is an avoidance behavior, the corresponding initial refined optimization rules include: a horizontal spacing range threshold, a horizontal speed change time and a percentage threshold, and a horizontal speed change value is a preset threshold. The second preset rule parameter space includes: the second minimum horizontal spacing. Parameter space, second maximum lateral spacing Parameter space, lateral velocity variation window Parameter space, proportion of lateral velocity change Parameter space; Second minimum lateral spacing The parameter space includes: ; Second maximum lateral spacing The parameter space includes: , This represents the typical spacing of dividing lines in a pedestrian walkway within the target scenario. The lateral velocity change window The parameter space includes: ; The proportion of lateral velocity change The parameter space; including: .

6. The method for recognizing pedestrian interaction behavior in complex scenarios according to claim 1, characterized in that, The step of identifying pedestrian movement trajectory data in the target scene according to the target dynamic recognition rule and determining pedestrian interaction behavior includes: Based on the target coarse recognition rules, the pedestrian movement trajectory data of the target scene is coarsely analyzed to screen out target candidate pairs with potential target behaviors; the target candidate pairs with potential target behaviors include first-class candidate pairs with potential overtaking behavior and second-class candidate pairs with potential avoidance behavior; The target candidate pairs are analyzed according to the target identification rules to determine pedestrian interaction behavior.

7. The method for recognizing pedestrian interaction behavior in complex scenarios according to claim 1, characterized in that, The step of converting the video pixel coordinate system of the continuous trajectory data into a real-world coordinate system to obtain the original trajectory data includes: Using Zhang Zhengyou's calibration method, camera intrinsic parameter information for the pedestrian behavior video was obtained through a checkerboard target. Based on the feature points of the known world coordinates and the camera's intrinsic parameter information, the camera's rotation matrix and translation vector are calculated to establish the transformation relationship between the world coordinate system and the camera coordinate system. Based on the transformation relationship, the video pixel coordinate system of the continuous trajectory data is converted into the real-world coordinate system to obtain the original trajectory data.

8. The method for recognizing pedestrian interaction behavior in complex scenarios according to claim 1, characterized in that, The preprocessing of the original trajectory data to obtain the pedestrian movement trajectory data of the target scene to be identified includes: Obtain the trajectory length of each of the original trajectory data, and remove the original trajectory data whose trajectory length is less than a preset length threshold to obtain the first target trajectory dataset; Obtain the longitudinal displacement of each original trajectory data in the first target trajectory dataset, and remove the original trajectory data whose longitudinal displacement is less than a preset longitudinal displacement threshold from the first target trajectory dataset to obtain the second target trajectory dataset; All the original trajectory data in the second target trajectory dataset are target motion trajectories.

Images