A method and device for analyzing adaptability of a public place human-machine environment

Abstract

The application relates to a public place human-machine environment adaptability analysis method and device, and belongs to the technical field of barrier-free environment evaluation. The method comprises the following steps: acquiring human-machine environment multi-source data, and performing space-time alignment, target tracking and data standardization processing to obtain standardized scene data; identifying and segmenting user behaviors based on the standardized scene data, and cutting each behavior segment into multiple time slices; extracting human side features, machine side features, ring side features and coupling features of each time slice, and splicing the features into a feature vector; inputting the feature vector into a pre-trained collaborative adaptation analysis model to output slice collaboration indexes of each time slice and contribution degrees of the human side, the machine side and the ring side; performing step-by-step aggregation on the slice collaboration indexes to obtain behavior collaboration indexes, scene collaboration indexes and overall collaboration indexes of the public place; determining a main factor side affecting a current collaboration state according to the contribution degrees of the human side, the machine side and the ring side, and generating an intervention strategy according to the main factor side.

Description

Technical Field

[0001] This application belongs to the field of barrier-free environment assessment technology, and specifically relates to a method and apparatus for analyzing the adaptability of human-machine environment in public places. Background Technology

[0002] With the development of barrier-free environment construction and intelligent public places, barrier-free facilities and intelligent assistive equipment are gradually being installed in public places. The user's passage status in public places is affected by the user's own status, the operation status of assistive equipment and the status of the environment. Therefore, it is necessary to conduct a comprehensive analysis of the collaborative adaptation status between people, machines and environment.

[0003] Existing methods for adaptability analysis in public places typically collect user, auxiliary equipment, and environmental data, and analyze them based on a single indicator or a single evaluation model. This approach has the following problems: it is difficult to accurately divide behavioral stages based on changes in user behavior, resulting in data from different stages being mixed together; it mainly focuses on analyzing the state of one side of the user, machine, or environment, making it difficult to effectively characterize the coupling relationship between the three; and it is difficult to reflect the overall adaptability status from different levels such as time slices, behavioral segments, and scenarios. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide a method and apparatus for adaptability analysis of human-machine environment in public places.

[0005] The objective of this invention is achieved as follows: A first aspect of the present invention provides a method for adaptability analysis of human-machine environment in public places, comprising: Acquire multi-source human-machine environment data in public places, and perform spatiotemporal alignment, target tracking and data standardization processing to obtain standardized scene data; Based on the standardized scenario data, user behavior is identified and segmented to obtain multiple behavior segments, and each behavior segment is further divided into multiple time slices. Extract human-side features, machine-side features, loop-side features, and coupling features from each time slice, and concatenate them into a feature vector; The feature vector is input into a pre-trained collaborative adaptation analysis model, which outputs the slice collaboration index of each time slice and the contributions of the human side, machine side and ring side. Based on the hierarchical relationship between time slices, behavior segments, and scenes, the slice coordination index is aggregated level by level to obtain the behavior coordination index, scene coordination index, and overall public place coordination index, and corresponding level of barrier-free environment coordination and adaptation evaluation results are generated. Based on the contributions of the human, machine, and environmental sides, the main factors affecting the current collaborative state are determined, and intervention strategies are generated based on the main factors. The execution result data of the intervention strategy is fed back to the database to incrementally update the collaborative adaptation analysis model using the execution result data.

[0006] Furthermore, the acquisition of multi-source human-machine environment data in public places includes: collecting user movement trajectory and skeletal posture data, auxiliary equipment operation parameter data, spatial occupancy distribution and physical quantity data of the environment, and business event log data; the spatiotemporal alignment, target tracking, and data standardization processing to obtain standardized scene data includes: connecting each data acquisition device to a unified time base, adding a unified timestamp to each data stream, establishing a scene semantic raster map containing a static occupancy layer, a semantic layer, and a dynamic layer, and mapping the multi-source data to the scene semantic raster map; detecting and tracking users and auxiliary equipment in the acquired video streams, performing cross-field association of the same target in different fields of view based on spatiotemporal continuity, and establishing anonymous identifiers for associated targets; encoding the data after spatiotemporal alignment, spatial registration, and target tracking according to a unified data format, and processing abnormal and missing data to obtain standardized scene data with a unified time base, unified spatial coordinates, and a unified data format.

[0007] Furthermore, the identification and segmentation of user behavior based on the standardized scene data includes: constructing a trajectory kinematic sequence using instantaneous velocity, heading angle change rate, and trajectory curvature; locating abrupt change points in the trajectory kinematic sequence using cumulative sum and change point detection; dividing the user trajectory into multiple trajectory primitives based on the abrupt change points; constructing a hidden Markov model using behavior types in a preset behavior type set as hidden states; weighting and fusing functional area transition criteria, trajectory primitive criteria, and event triggering criteria to obtain the emission probability; solving for the hidden state sequence with the maximum joint probability in the hidden Markov model using the Viterbi algorithm; determining the trajectory intervals corresponding to consecutive identical behavior types in the hidden state sequence as behavior segments; and assigning corresponding behavior type labels to each behavior segment.

[0008] Furthermore, the extraction of human-side features, machine-side features, loop-side features, and coupling features for each time slice includes: extracting walking speed, stride variability, joint mobility, path tortuosity, number of stops, head orientation entropy, heart rate relative to baseline change, posture stability, and behavioral stage markers as human-side features from the time slice; extracting equipment response delay, number of times the user actively requests assistance, navigation and positioning error, battery percentage, ratio of motor current to rated current, and equipment task status as machine-side features from the time slice; extracting passenger flow density, dynamic obstacle occupancy rate, queue length, illumination, noise, and ground condition as loop-side features from the time slice; extracting human-machine collaboration features based on the speed sequence correlation between the user and auxiliary equipment and the stability of the following distance; extracting human-loop conflict features based on the minimum clearance between the user and dynamic obstacles and the detour rate of the user's actual path relative to the shortest passable path; and extracting machine-loop adaptation features based on the net channel margin and the matching degree of the positioning environment.

[0009] Further, the step of inputting the feature vector into a pre-trained collaborative adaptation analysis model and outputting the slice collaboration index and the contributions of the human, machine, and ring sides for each time slice includes: using the feature vector as the initial feature of the node to construct a heterogeneous graph for each time slice; inputting the heterogeneous graph into a graph attention network for spatial encoding; simultaneously inputting the temporal subsequences within the time slice into a temporal convolutional network for temporal encoding; fusing the results of spatial encoding and temporal encoding to obtain the slice collaboration index for each time slice; using the feature vector of the current time slice as the input to be attributed; using the average feature of historical samples corresponding to the current behavior type as the attribution baseline; calculating the contribution value of the slice collaboration index relative to each feature using the integral gradient method; and merging and normalizing the contribution values ​​according to the human, machine, and ring sides to which each feature belongs to, to obtain the human-side contribution, machine-side contribution, and ring-side contribution.

[0010] Furthermore, the graph attention network is a multi-layer multi-head graph attention network, and the temporal convolutional network is a dilated causal convolutional network, with the dilation factor increasing layer by layer along the network hierarchy; the results of the spatial encoding and the temporal encoding are concatenated, and the collaborative latent variables are obtained by mapping through a fully connected layer; the collaborative latent variables are mapped to the slice collaborative index through a Sigmoid function with a learnable steepness coefficient k and a center offset parameter.

[0011] Further, the step of aggregating the slice coordination index level by level according to the hierarchical relationship between time slices, behavior segments, and scenarios to obtain the behavior coordination index, scenario coordination index, and overall public place coordination index, and generating corresponding level accessibility environment coordination adaptation evaluation results, includes: for each behavior segment, determining the bottleneck interval based on the slice coordination index of each time slice within the behavior segment, and aggregating it according to the slice coordination index of each time slice and its temporal weight to obtain the behavior coordination index corresponding to the behavior segment; determining the behavior-level accessibility environment coordination adaptation evaluation result of the behavior segment based on the relationship between the behavior coordination index and the coordination evaluation benchmark of the corresponding behavior type; for each scenario, taking the concurrent behaviors within the scenario as behavior nodes, and aggregating them according to the concurrent behaviors... To determine the resource competition relationship between behavioral nodes based on the spatiotemporal occupancy relationship of shared resources, and to aggregate the behavioral coordination index corresponding to each behavioral node and the resource competition relationship, a scenario coordination index corresponding to the scenario is obtained. Based on the relationship between the scenario coordination index and the scenario coordination evaluation benchmark, the scenario-level barrier-free environment coordination adaptation evaluation result of the scenario is determined. For the public place, the association relationship between scenarios is determined based on the travel relationship between each scenario, and to aggregate the scenario coordination index corresponding to each scenario and the association relationship between the scenarios, an overall coordination index of the public place is obtained. Based on the relationship between the overall coordination index of the public place and the public place coordination evaluation benchmark, the overall barrier-free environment coordination adaptation evaluation result of the public place is determined.

[0012] Furthermore, the step of generating intervention strategies based on the main factor side includes: determining whether the slice coordination index is lower than the coordination threshold corresponding to the current behavior type for multiple consecutive time slices; if so, determining the corresponding intervention strategy based on the main factor side with the largest contribution; when the main factor side is the ring side, generating intervention strategies such as opening backup channels, adjusting passenger flow guidance, or dispatching staff to designated locations based on passenger flow density or facility occupancy status; when the main factor side is the machine side, generating intervention strategies such as replacing compatible equipment, adjusting equipment deployment points, or adjusting equipment task parameters based on equipment response delay or equipment positioning error; when the main factor side is the human side, generating personalized guidance information or manual assistance for reservations based on the difference between user ability characteristics and the current scenario's normal level.

[0013] Furthermore, the pre-training process of the collaborative adaptation analysis model includes: collecting multi-source data of multiple users completing predetermined trips within the target location, and storing it in the human-machine-loop database after spatiotemporal alignment, standardized encoding, and quality inspection; dividing the training set and validation set into units based on complete locations, and using cross-validation, reserving all data of one type of location as the validation set each time, and using the data of the remaining locations as the training set; using the behavioral-level result labels generated based on business event logs and the order labels of sample pairs as weak supervision signals, and using the weighted sum of the mean square error of the slice collaboration index after behavioral-level aggregation with the behavioral-level result labels, the sorting hinge loss with the order labels, and the parameter L2 regularization term as the total loss function, and training the graph attention network, temporal convolutional network, and fully connected layer parameters end-to-end; after training, calculating the model acceptance index using the validation set, and obtaining the collaborative adaptation analysis model after the index meets the standard.

[0014] A second aspect of the present invention provides a human-machine environment adaptability analysis device for public places, comprising: The data acquisition and preprocessing module is used to acquire multi-source human-machine environment data in public places, and perform spatiotemporal alignment, target tracking and data standardization to obtain standardized scene data; The behavior recognition and segmentation module is used to recognize and segment user behavior based on the standardized scene data, obtain multiple behavior segments, and divide each behavior segment into multiple time slices. The feature extraction module is used to extract human-side features, machine-side features, loop-side features, and coupling features for each time slice, and concatenate them into a feature vector; The collaborative adaptation analysis module is used to input the feature vector into the pre-trained collaborative adaptation analysis model and output the slice collaboration index of each time slice and the contribution of the human side, machine side and ring side. The collaborative adaptation evaluation module is used to aggregate the collaborative index of the slices according to the hierarchical relationship between time slices, behavior segments and scenes, to obtain the behavior collaborative index, scene collaborative index and public place overall collaborative index, and generate the corresponding level of barrier-free environment collaborative adaptation evaluation results. An intervention strategy generation module is used to determine the main factors affecting the current collaborative state based on the contributions of the human side, machine side, and environment side, and to generate an intervention strategy based on the main factors. The incremental update module is used to feed the execution result data of the intervention strategy back to the database, so as to use the execution result data to incrementally update the collaborative adaptation analysis model.

[0015] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the human-machine environment adaptation analysis method for public places as described in any embodiment.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the human-machine environment adaptation analysis method for public places as described in any embodiment.

[0017] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: By performing spatiotemporal alignment, target tracking, and standardization on multi-source human-machine environment data, unified standardized scene data is obtained, improving the correlation and consistency between different data sources. By identifying, segmenting, and dividing user behavior into time slices, collaborative adaptation analysis can correspond to specific behavioral and time stages, reducing the impact of mixed data from different stages on the analysis results. By extracting human-side, machine-side, and environment-side features and their coupling characteristics, a slice collaboration index is obtained using a collaborative adaptation analysis model, enabling a comprehensive characterization of users, auxiliary equipment, environment, and their interactions, thus improving the completeness of human-machine environment collaborative adaptation state analysis. By aggregating the slice collaboration index hierarchically according to the time slice, behavioral segment, and scene hierarchy, a layered collaborative adaptation evaluation is achieved from local time state to behavior, scene, and the overall public place. By determining the main factor side based on the contribution of the human-side, machine-side, and environment side, and generating intervention strategies based on the main factor side, the analysis results can be further used for targeted adaptation interventions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 A flowchart of the human-machine environment adaptability analysis method for public places provided in Embodiment 1 of the present invention; Figure 2 A schematic diagram of the human-machine environment adaptability analysis device for public places provided in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the electronic device architecture provided in Embodiment 3 of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. It should be noted that, unless otherwise specified, the implementation methods and features in the implementation methods in this disclosure can be combined, separated, interchanged, and / or rearranged. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Example 1 A specific embodiment of the present invention, such as Figure 1 As shown, a method for adaptability analysis of human-machine environment in public places is disclosed, including the following steps: S1. Acquire multi-source data on the human-machine environment in public places, and perform spatiotemporal alignment, target tracking and data standardization processing to obtain standardized scene data.

[0022] In this embodiment, step S1 specifically includes: S101. Collect user movement trajectory and skeletal posture data within the venue, auxiliary equipment operation parameter data, spatial occupancy distribution and physical quantity data of the venue environment, and business event log data.

[0023] For example, fixed sensing devices, environmental sensors, telemetry terminals for auxiliary equipment, and business event data interfaces are deployed in public places; the fixed sensing devices include panoramic cameras, depth cameras, and lidar or millimeter-wave radar; wherein, the panoramic camera is used to collect video data of users and auxiliary equipment in the public place, the depth camera is used to obtain the depth information of the target, and the lidar or millimeter-wave radar is used to assist in extracting the target trajectory in the public place.

[0024] In one specific implementation, at least two panoramic cameras and one depth camera are deployed for every 500 square meters of scene area, and lidar or millimeter-wave radar is deployed in permitted locations; illuminance, noise, and temperature and humidity sensors are deployed at key nodes such as passages and gates; auxiliary equipment reads operating parameters such as positioning coordinates, three-axis acceleration, angular velocity, power consumption, motor current, navigation positioning error, obstacle recognition results, voice interaction timestamps, and task status codes through reserved data interfaces; human-side data includes users' motion trajectories in public places, skeletal posture data, and physiological data collected during the database construction phase; for volunteers recruited during the database construction phase, motion and heart rate data can be collected through lumbar inertial measurement units and heart rate belts; for users in real operating scenarios, users are not required to wear additional equipment, but kinematic features are extracted through anonymous video data and radar data; business event data is accessed through message queues from gate opening records, elevator call and arrival records, manual service dispatch records, and customer service and complaint records, and is uniformly converted into event records containing "event type, timestamp, location, and associated objects".

[0025] S102. Connect each data acquisition device to a unified time base, add a unified timestamp to each data stream, establish a scene semantic grid map including a static occupancy layer, a semantic layer and a dynamic layer, and map the multi-source data to the scene semantic grid map. The system detects and tracks users and auxiliary equipment in the acquired video stream, performs cross-field association of the same target in different fields of view based on spatiotemporal continuity, and establishes anonymous identifiers for associated targets. Data that has undergone spatiotemporal alignment, spatial registration, and target tracking is encoded in a unified data format, and abnormal and missing data is processed to obtain standardized scene data with a unified time base, unified spatial coordinates, and unified data format.

[0026] Specifically, after acquiring the aforementioned multi-source data, time alignment is performed on different data sources: all data acquisition devices are connected to a unified time base, and a unified timestamp is added to each data stream; using the camera frame time as a reference time, time alignment is performed on other data streams, and the nearest neighbor interpolation method is used to map equipment telemetry data, environmental sensor data, and business event data to the camera frame time axis. When the time alignment residual of a certain data segment exceeds the preset allowable range, the data segment is marked as invalid data and removed, so that the data from video, radar, auxiliary equipment telemetry terminals, environmental sensors, and business systems are under a unified time base; Based on temporal alignment, spatial alignment is performed on multi-source data: First, a scene semantic raster map of public places is established, which includes a static occupancy layer, a semantic layer, and a dynamic layer. The static occupancy layer describes walls, fixed facilities, and other static obstacles. The semantic layer identifies functional areas and facility types within the public place, such as ticket check areas and accessible elevator lobbies. The dynamic layer describes the real-time changing occupancy of dynamic obstacles during operation. For camera data, camera extrinsic parameter calibration is completed using a ground checkerboard pattern or coded markers within the scene, and the pixel coordinates in the image are mapped to the scene semantic raster map coordinate system. For LiDAR data, the point cloud is registered to the static occupancy layer, thereby unifying the data collected by different sensing devices into the same scene coordinate system. This yields user, auxiliary equipment, facility, and dynamic obstacle data under unified spatial coordinates.

[0027] Furthermore, after spatial registration, target detection and tracking are performed on users and assistive devices in the video stream: target detection is performed frame by frame to identify targets such as users, wheelchairs, and intelligent assistive devices; target numbers within the scene are maintained based on target motion states and appearance characteristics, and trajectory extrapolation is performed using motion states when a target is temporarily occluded; when the same target moves from one camera's field of view to another, cross-field association is performed based on the target's temporal and spatial continuity and appearance feature similarity. When the association confidence is insufficient, a new number is created for the target; for targets that have been associated, an anonymous target file is created, storing only skeletal keypoint sequences and trajectory coordinates, and other data related to collaborative adaptation analysis, without storing face images; the original video is automatically deleted according to a preset retention period; for skeletal pose data, human pose estimation algorithms are used to obtain human keypoint coordinates and corresponding confidence levels, and for keypoints with confidence levels lower than the set requirements, interpolation is performed using adjacent frames to fill in the gaps.

[0028] Furthermore, after completing time alignment, spatial registration, and target tracking, various types of data are converted into a unified data encoding format, and data with different physical meanings are uniformly mapped to the [0,1] interval. For data of different properties, proportional, inverse proportional, ordered hierarchical, Boolean and existence, and distance decay mapping functions are used for standardization. For example, proportional mapping is used for continuous quantities where larger values ​​are more advantageous; inverse proportional mapping is used for parameters such as navigation positioning errors and voice response delays where smaller values ​​are more advantageous; ordered hierarchical mapping is used for ordered level data such as muscle strength levels; Boolean mapping is used for status data such as facility existence; and distance decay mapping is used for facility accessibility parameters. Human-side data is further mapped using the ICF system for main and sub-coding, and the encoding number, original value, mapping function, version number, and mapping value are saved. After standardization, data quality control is performed. For example, jump points in the trajectory that clearly exceed the normal range of motion are removed; the sample level of samples whose effective trajectory length is less than a certain proportion of the complete behavioral process is reduced; for time slices where equipment telemetry data has been interrupted for a long time, the corresponding equipment features are removed and missing markers are added; and data segments with excessive time alignment residuals are removed.

[0029] After the above processing, standardized scene data with a unified time base, unified spatial coordinates, and unified data encoding format are obtained.

[0030] S2. Based on the standardized scenario data, user behavior is identified and segmented to obtain multiple behavior segments, and each behavior segment is divided into multiple time slices.

[0031] In this embodiment, user behavior is identified and segmented based on standardized scenario data, specifically including: A trajectory kinematic sequence is constructed using instantaneous velocity, rate of change of heading angle, and trajectory curvature. Accumulation and change point detection are used to locate abrupt changes in the trajectory kinematic sequence, and the user trajectory is divided into multiple trajectory primitives based on the abrupt changes. A hidden Markov model is constructed by using the behavior types in the preset behavior type set as hidden states and the emission probability by weighted fusion of the functional area transfer criterion, trajectory primitive criterion and event triggering criterion. The Viterbi algorithm is used to solve the hidden state sequence with the maximum joint probability in the hidden Markov model. The trajectory intervals corresponding to consecutive identical behavior types in the hidden state sequence are determined as behavior segments, and corresponding behavior type labels are assigned to each behavior segment.

[0032] In one implementation, trajectory primitives are classified into types such as straight-line movement, turning, stopping, queuing, and obstacle avoidance; wherein, when the user's speed is lower than the corresponding speed threshold and remains below it for a certain period of time, it is identified as stopping; when the user's speed is low and there are other targets in a certain range ahead, it is identified as queuing; when the trajectory curvature increases significantly and lateral deviation occurs, it is identified as obstacle avoidance.

[0033] Furthermore, a hidden Markov model is used to segment the behavior: the behavior types in the preset behavior type set are used as hidden states, such as straight-line passage, queuing for inspection, passing through the gate, taking the elevator to transfer, passing through the ramp, and waiting for help; functional area transfer information, trajectory primitive information, and business event information are used as observation information. Specifically, the functional area transition criterion is determined based on the semantic functional area where the trajectory centroid is located; the trajectory primitive criterion is determined based on the matching degree between the current trajectory primitive and the candidate behavior type; the event triggering criterion is determined based on whether business events such as gate opening and elevator door opening occur; the matching scores of the above three criteria for the candidate behavior type are calculated respectively, and the three criteria are weighted and fused using the fusion coefficients obtained from training to obtain the emission probability corresponding to the candidate behavior type; at the same time, the state transition probability is determined based on the transition relationship between adjacent behavior types, thereby constructing a hidden Markov model; the Viterbi algorithm is used to solve the hidden state sequence with the maximum joint probability, and the continuous trajectory intervals with the same behavior type in the hidden state sequence are determined as a behavior segment, and the corresponding behavior type label is assigned to the behavior segment.

[0034] For example, a sliding time window is used to generate time slices for behavior segments. The default time window length is 5 seconds, with a step size of 2.5 seconds, so that there is a 50% overlap between adjacent time slices. The time window length is adjusted according to the current behavior stage. For quasi-static behavior stages with slow changes such as queuing and staying, the time window length is extended to 10 seconds to reduce the amount of computation. For key transition stages such as passing through gates and getting on and off auxiliary equipment, the time window length is shortened to 2 seconds to improve time resolution. For each generated time slice, a slice record is created, including slice number, behavior number, functional area, start and end timestamps, associated user number, associated equipment number, and feature vector. After the model inference is completed, the corresponding slice coordination index is further written.

[0035] S3. Extract the human-side features, machine-side features, ring-side features, and coupling features of each time slice, and concatenate them into a feature vector.

[0036] In this embodiment, step S3 specifically includes: Step speed, stride variability, joint range of motion, path tortuosity, number of stops, head orientation entropy, heart rate change relative to baseline, posture stability, and behavioral stage markers were extracted from time slices as human-side features. The equipment response delay, number of times the user actively requests assistance, navigation and positioning error, battery percentage, ratio of motor current to rated current, and equipment mission status are extracted from the time slice as machine-side features. Passenger flow density, dynamic obstacle occupancy rate, queue length, illumination, noise, and ground condition are extracted from time slices as ring-side features; Human-machine collaboration features are extracted based on the correlation between the speed sequence of the user and the auxiliary equipment and the stability of the following distance. Human-loop conflict features are extracted based on the minimum clearance between the user and dynamic obstacles and the detour rate of the user's actual path relative to the shortest passable path. Machine-loop adaptation features are extracted based on the net width of the channel and the matching degree of the positioning environment.

[0037] Specifically, the extraction process for each individual's features is as follows: Walking speed is calculated based on the trajectory displacement and time length within a time slice, and the coefficient of variation of walking speed is calculated from the standard deviation and average value of walking speed within the slice; stride length is determined based on the distance sequence between the left and right ankle key points, and its coefficient of variation is calculated; joint mobility is formed by creating a joint angle sequence based on the vector angles between human key points, and the difference between the maximum and minimum values ​​is taken; path tortuosity is determined based on the ratio of the actual movement path length to the straight-line distance from the starting point to the ending point; the number of pauses is statistically analyzed based on situations where the user's speed is below a certain threshold and persists for a certain period of time; head orientation entropy is calculated by dividing the head yaw angle into multiple directional sectors, and is used to describe the user's hesitation in finding their way in the current scene; heart rate relative to baseline change is determined based on the difference between the current heart rate and the user's resting baseline heart rate; posture stability is determined based on the lateral swing amplitude of the human body's center of mass key point; and behavior stage markers are encoded based on the relative position of the current time slice within the corresponding behavior segment. The extraction process for each machine-side feature is as follows: the equipment response delay is determined based on the time difference between the timestamp of the user's voice or key press request and the timestamp of the auxiliary equipment starting to act; the number of times the user actively seeks help is obtained based on the statistics of help events within the time slice; the navigation and positioning error is based on the positioning error reported by the auxiliary equipment; the equipment mission status can be encoded as idle, guiding, charging, or fault status, etc. The extraction process for each ring side feature is as follows: passenger flow density is calculated based on the number of users and passable area in the dynamic layer grid; dynamic obstacle occupancy rate is calculated based on the proportion of grids occupied by obstacles within a certain range centered on the user; queue length is determined based on the number of targets corresponding to the queuing behavior primitives in front; illuminance and noise are collected by environmental sensors; ground status is determined based on environmental sensor data and operation and maintenance records. The coupling feature is used to describe the relationships between humans, machines, and the environment; The human-machine collaboration features include the zero-delay cross-correlation coefficient between the speed sequences of the user and the auxiliary equipment, and the stability of the following distance between the user and the auxiliary equipment. The human-environment conflict features include the minimum clearance between the user and dynamic obstacles, and the detour rate of the user's actual movement path relative to the shortest passable path in a static environment. The shortest passable path can be obtained using the A* algorithm on a static occupancy layer. The machine-environment adaptation features include channel net margin and positioning environment matching degree. The channel net margin is determined based on the difference between the actual net width of the channel and the net width required by the auxiliary equipment. The positioning environment matching degree is determined based on the correlation between the positioning error reported by the auxiliary equipment and the environmental attributes of its grid.

[0038] After extracting the human-side features, machine-side features, ring-side features, and coupling features, the features are concatenated in a preset order to form the feature vector of the current time slice.

[0039] S4. Input the feature vector into the pre-trained collaborative adaptation analysis model and output the slice collaboration index of each time slice and the contribution of the human side, machine side and ring side.

[0040] In this embodiment, step S4 specifically includes: S401. Using the feature vector as the initial feature of the node, construct a heterogeneous graph for each time slice. Input the heterogeneous graph into a graph attention network for spatial encoding. At the same time, input the temporal subsequences inside the time slice into a temporal convolutional network for temporal encoding. Combine the results of spatial encoding and temporal encoding to obtain the slice synergy index of each time slice.

[0041] Specifically, the graph attention network is a multi-layer multi-head graph attention network, and the temporal convolutional network is a dilated causal convolutional network, with the dilation factor increasing layer by layer along the network hierarchy; The results of the spatial encoding and the temporal encoding are concatenated and mapped through a fully connected layer to obtain collaborative latent variables. The collaborative latent variables are then mapped to the slice collaborative indexes using a Sigmoid function with a learnable steepness coefficient k and a center offset parameter.

[0042] Specifically, the heterogeneous graph includes user nodes, auxiliary equipment nodes, and functional area or facility nodes related to the user's current functional area. Usage relationships, guidance relationships, occlusion relationships, and competition relationships are established between nodes. Usage relationships describe the relationship between the user and auxiliary equipment; guidance relationships describe the relationship between equipment and facilities; occlusion relationships describe the spatial occlusion relationship between the user and facilities; and competition relationships describe the competition between users or equipment for shared resources. For different relationships, corresponding edge features can also be set. For example, the strength of a usage relationship can be represented by the following distance between the user and equipment, and the strength of a competition relationship can be represented by the percentage of time the resources overlap.

[0043] In this embodiment, the collaborative adaptation analysis model consists of three parts connected in series: (1) Spatial encoder: A two-layer graph attention network is adopted, with 4 attention heads in each layer and a hidden layer dimension of 64. The feature representation of the current node is updated by performing attention weighted aggregation on the neighboring nodes. For node i, the attention coefficient is calculated according to the degree of correlation between node i and its neighboring node j, and the features of the neighboring nodes are weighted and aggregated according to the attention coefficient to obtain the spatial encoding result. (2) Time encoder: A dilated causal convolutional network is used with 3 convolution kernels and dilation coefficients of 1, 2 and 4 respectively. The temporal pattern is extracted from the kinematic subsequences (step speed, joint angle, human-machine distance, etc.) with a granularity of 0.5 seconds inside the slice, and the time-aligned encoding vector is output. (3) Output Header: The spatial coding vector and the temporal coding vector are concatenated and mapped through a fully connected layer to obtain the collaborative latent variables; the collaborative latent variables are mapped to slice collaborative exponents through a sigmoid function with learnable steepness coefficients and center offset parameters, expressed as: Where z represents the fused collaborative latent variable, and k represents the learnable steepness coefficient. This represents the learnable center offset parameter; through end-to-end training, the steepness coefficient and center offset parameter can learn the corresponding output mapping relationship according to different scenarios, which improves the model's sensitivity to changes in cooperative state in critical transition scenarios and reduces exponential jitter in ordinary traffic scenarios.

[0044] S402. Using the feature vector of the current time slice as the input to be attributed, and the average feature of the historical samples corresponding to the current behavior type as the attribution baseline, the contribution value of the slice synergy index relative to each feature is calculated using the integral gradient method. The contribution values ​​are then merged and normalized according to the human side, machine side, and ring side to which each feature belongs, to obtain the human side contribution, machine side contribution, and ring side contribution.

[0045] Specifically, after obtaining the slice synergy index, the contribution of the current time slice is attributed to the human side, machine side, and ring side. The feature vector of the current time slice is used as the actual input, and the average feature vector of the historical samples corresponding to the current behavior type is used as the attribution baseline. The integral gradient method is used to construct a straight path between the attribution baseline and the actual input, and the gradient integral is performed on the model output along the straight path to calculate the contribution value of each input feature to the slice synergy index. For the i-th feature, its integral gradient contribution can be expressed as: Where x is the actual feature vector of the current time slice. As the attribution baseline, represents the path interpolation coefficients between 0 and 1, and F is the output function corresponding to the co-fit analysis model.

[0046] In one specific implementation, the Riemann sum is used to numerically approximate the above integral; after obtaining the integral gradient contribution value of each feature, it is merged according to the human side, machine side, and ring side to which each feature belongs; for coupled features that are simultaneously associated with both sides, their contribution values ​​are respectively included in the associated two sides according to the corresponding proportions; for example, half of the coupled features that are simultaneously associated with both sides are included; the merged results are normalized to obtain the human side contribution, machine side contribution, and ring side contribution respectively.

[0047] Specifically, the pre-training process of the collaborative adaptation analysis model includes: Multi-source data on multiple users completing their scheduled trips at the target location is collected, and after spatiotemporal alignment, standardized coding, and quality inspection, it is stored in the human-machine-environment database. The training set and validation set are divided into complete venues. Cross-validation is used, with all data of one type of venue reserved as the validation set and the data of the remaining venues used as the training set. Using behavioral-level result labels and order labels of sample pairs generated from business event logs as weak supervision signals, and the weighted sum of the mean square error of the slice synergy index after behavioral-level aggregation with the behavioral-level result labels, the sorting hinge loss with the order labels, and the parameter L2 regularization term as the total loss function, the graph attention network, temporal convolutional network, and fully connected layer parameters are trained end-to-end. After training, the validation set is used to calculate the model acceptance index. Once the index is met, the collaborative adaptation analysis model is obtained.

[0048] Specifically, the model training uses a weakly supervised approach: First, behavioral-level result labels are generated based on business event logs. These labels are obtained by encoding and mapping result information such as task completion time ratio, task results, number of requests for help and intervention, and physiological load increment. Secondly, sample pairs are constructed from sampled similar behavioral fragments, and experts judge which of the two behavioral fragments has a better collaborative state, thus forming the sample pair order label.

[0049] During model training, the slice synergy index of time slices is aggregated at the behavior level to obtain the behavior synergy index, and a fitting loss is constructed based on the difference between the behavior synergy index and the behavior-level result label. Simultaneously, a pairwise ranking hinge loss is constructed based on the order labels provided by experts, and an L2 regularization term is added to the model parameters to form the total loss function. ,in, This represents the mean squared error between the behavioral-level target value and the model's aggregated prediction value. This represents the pairwise ordering hinge loss between sample pairs. The L2 regularized loss represents the model parameters. and These represent the weight coefficients of the corresponding loss terms. Through the total loss function, the graph attention network, temporal convolutional network, and fully connected output layer are trained end-to-end, enabling the model to simultaneously learn the spatial and temporal relationships between humans and the environment within a time slice, as well as the mapping relationship between time slices and behavioral collaborative states. After training, leave-one-place cross-validation is used to validate the model. When the model's sample pair ranking consistency rate, correlation coefficient with expert order judgment, and low-cooperative slice recognition ability all meet the preset acceptance requirements, the current model is determined as the pre-trained cooperative adaptation analysis model.

[0050] In some embodiments, the method further includes: generating simulation data in a scene digital twin environment for abnormal working conditions with a small number of real samples, such as fall hazards, equipment failures, and extreme congestion, and training the simulation data by mixing it with real data according to a preset ratio. At the same time, the method uses domain discrimination loss to constrain the difference in feature distribution between simulation data and real data, so as to reduce the overfitting of the model to simulation features.

[0051] S5. According to the hierarchical relationship between time slices, behavior segments, and scenes, the slice coordination index is aggregated level by level to obtain the behavior coordination index, scene coordination index, and overall public place coordination index, and the corresponding level of barrier-free environment coordination and adaptation evaluation results are generated.

[0052] In this embodiment, step S5 specifically includes: For each behavior segment, a bottleneck interval is determined based on the slice coordination index of each time slice within the behavior segment, and the slice coordination index and its temporal weight of each time slice are aggregated to obtain the behavior coordination index corresponding to the behavior segment; the behavior-level accessibility environment coordination adaptation evaluation result of the behavior segment is determined based on the relationship between the behavior coordination index and the coordination evaluation benchmark of the corresponding behavior type. For each scenario, concurrent behaviors within the scenario are taken as behavior nodes. The resource competition relationship between behavior nodes is determined based on the spatiotemporal occupancy relationship of each concurrent behavior on shared resources. The scenario collaboration index corresponding to each behavior node and the resource competition relationship are aggregated to obtain the scenario collaboration index corresponding to the scenario. The scenario-level barrier-free environment collaboration adaptation evaluation result of the scenario is determined based on the relationship between the scenario collaboration index and the scenario collaboration evaluation benchmark. For the public place, the relationship between the scenes is determined based on the travel relationship between each scene, and the overall coordination index of the public place is obtained by aggregating the scene coordination index corresponding to each scene and the relationship between the scenes. The overall accessibility environment coordination and adaptation evaluation result of the public place is determined based on the relationship between the overall coordination index of the public place and the coordination evaluation benchmark of the public place.

[0053] Specifically, hierarchical aggregation includes the following processes: (1) Collaborative aggregation from time slices to behavior level: Select abnormal results samples from the training set, extract all time slices of the corresponding behavior type from the abnormal results samples, and calculate the median of the slice coordination index distribution according to the behavior type. Use this median as the coordination evaluation threshold for the corresponding behavior type and update it with model iteration. For multiple consecutive time slices below the corresponding threshold within the same behavior segment, identify them as bottleneck intervals. For each bottleneck interval, the bottleneck severity is calculated based on the degree to which the coordination index of each time slice is below the threshold and the duration, in order to reflect the combined impact of the degree and duration of low coordination. A behavior-level temporal attention layer is set up, using the encoding vector corresponding to each time slice as the key and the behavior-level learnable query vector as the query, to calculate the attention weight corresponding to each time slice; the slice coordination index of each time slice is weighted and aggregated according to the attention weight, and then corrected by combining the severity corresponding to the bottleneck interval to obtain the behavior coordination index B; based on the relationship between the behavior coordination index B and the coordination evaluation benchmark of the corresponding behavior type, the behavior-level accessibility environment coordination adaptation evaluation result of the corresponding behavior segment is determined.

[0054] (2) Collaborative aggregation from behavior level to scenario level: Concurrent behaviors within the scene are used as behavioral nodes. Based on the spatiotemporal occupancy of shared resources by different concurrent behaviors, a resource competition relationship is constructed. Shared resources include passage grid segments, barrier-free elevators, service desks, and on-site shared equipment. When two actions overlap in spatiotemporal occupation of the same shared resource within the same time window, a resource competition edge is established between the corresponding action nodes, and the edge weight is determined based on the amount of shared resource occupied by the two actions and the resource capacity. Perform critical resource load analysis on shared resources. For each type of shared resource, statistically analyze the resource arrival demand and service capacity based on business logs, and use a queuing model to calculate resource utilization and blocking probability. At the same time, predict the movement position of the subject in the future based on the movement trajectory, and count the number of target pairs whose minimum distance between predicted trajectories is less than the preset safety distance, as the count of predicted conflict events. A scenario-level graph network is constructed using a resource contention graph as input. The behavior coordination index B and behavior type embedding of each behavior node are used as the behavior node features. Resource utilization, blocking probability, and predicted conflict event count are used as global node features. Scenario-level encoding is performed through a graph attention network to obtain the scenario coordination index S. Based on the relationship between the scenario coordination index S and the scenario coordination evaluation benchmark, the scenario-level barrier-free environment coordination adaptation evaluation result of the corresponding scenario is determined.

[0055] For example, the scenario-level supervision signal consists of the traffic efficiency per unit time, the occurrence rate of complaints and requests for help, and the records of dangerous events. These are encoded and mapped to form scenario-level target values, which are used to train the scenario-level aggregation network.

[0056] (3) Collaborative aggregation from the scene level to the overall level of public places: Public places are represented as directed graphs, where nodes represent different scenarios and edges represent connecting paths between scenarios. Based on historical user trajectory data, users are clustered according to the starting and ending functional areas of their trips to obtain the occurrence ratio of different typical trips, thereby constructing an OD demand matrix. For each scenario, the success probability of the trip corresponding to the scenario is calculated based on the scenario coordination index S; the mapping relationship between the scenario coordination index and the success of the trip within the scenario is established through logistic regression, and the scenario coordination index S is converted into the success probability p of the scenario; for a typical trip, the success probabilities of the trip through each scenario are multiplied together, and a correction coefficient is set according to the historical correlation between adjacent scenarios to obtain the completion probability of the trip. Based on the occurrence rate of different typical trips, the completion probability of each typical trip is weighted and summed to obtain the overall public place coordination index P. The overall public place coordination index P comprehensively considers the coordination status of different scenarios and the usage needs of different scenarios in the user's actual trip, and can characterize the overall barrier-free environment coordination and adaptation status of the entire public place. Based on the relationship between P and the public place coordination evaluation benchmark, the evaluation result of the overall barrier-free environment coordination and adaptation of the public place is determined.

[0057] Furthermore, each scenario in the public place can be sequentially downgraded to a failure level, the overall coordination index P of the public place can be recalculated, and the scenarios can be sorted according to the change in the overall coordination index before and after downgrading. This will identify key scenarios that have a significant impact on the overall coordination status of the public place and output corresponding suggestions for transformation priorities.

[0058] S6. Determine the main factors affecting the current collaborative state based on the contributions of the human side, machine side, and ring side, and generate an intervention strategy based on the main factors.

[0059] In this embodiment, generating an intervention strategy based on the main factors includes: Determine whether the slice synergy index is lower than the synergy threshold corresponding to the current behavior type for multiple consecutive time slices. If so, determine the corresponding intervention strategy based on the main factor side with the greatest contribution. When the main factor side is the ring side, intervention strategies such as opening backup channels, adjusting passenger flow guidance, or dispatching staff to designated locations are generated based on passenger flow density or facility occupancy status. When the main factor is the machine side, intervention strategies such as replacing the adapted equipment, adjusting the equipment deployment point, or adjusting the equipment task parameters are generated based on the equipment response delay or equipment positioning error. When the main factor is human, personalized guidance information or intervention strategies for manual assistance in making appointments are generated based on the difference between the user's ability characteristics and the current scenario's normal level.

[0060] S7. The execution result data of the intervention strategy is fed back to the database to incrementally update the collaborative adaptation analysis model using the execution result data.

[0061] Specifically, the actions, timestamps, and 10 slices before and after each intervention will be included. The sequence was written into the intervention record table to form effect feedback data; after the intervention... Records with an average improvement of less than 0.05 are marked as invalid interventions and fed back for rule base revision. All the above-mentioned operational data, including slice data, synergy index results, intervention records, and effect feedback data, are continuously fed back into the "human-machine-environment" database. Distribution drift detection is performed, and the maximum mean difference between the new operational data and the training data in the feature space is calculated. If the maximum mean difference exceeds the limit set when the database was built, incremental fine-tuning is triggered. The new data is the main component, mixed with 10% historical replay samples. Only the parameters of the output head and the last graph attention layer are fine-tuned, and the bottom layer is frozen to prevent catastrophic amnesia. After the new model version meets the offline verification indicators, it is gray-scale distributed to a single edge node for a one-week trial run. If there are no abnormalities, it is then fully released. All model and coding library versions are traceable and rollbackable, forming a complete closed loop of "perception, analysis, decision-making, intervention, feedback, and evolution".

[0062] Compared with existing technologies, the adaptability analysis method for human-machine environments in public places provided in this embodiment can achieve the following technical effects: By performing spatiotemporal alignment, target tracking, and standardization on multi-source human-machine environment data, unified standardized scene data is obtained, improving the correlation and consistency between different data sources. By identifying, segmenting, and dividing user behavior into time slices, collaborative adaptation analysis can correspond to specific behavioral and time stages, reducing the impact of mixed data from different stages on the analysis results. By extracting human-side, machine-side, and environment-side features and their coupling characteristics, a slice collaboration index is obtained using a collaborative adaptation analysis model, enabling a comprehensive characterization of users, auxiliary equipment, environment, and their interactions, thus improving the completeness of human-machine environment collaborative adaptation state analysis. By aggregating the slice collaboration index hierarchically according to the time slice, behavioral segment, and scene hierarchy, a layered collaborative adaptation evaluation is achieved from local time state to behavior, scene, and the overall public place. By determining the main factor side based on the contribution of the human-side, machine-side, and environment side, and generating intervention strategies based on the main factor side, the analysis results can be further used for targeted adaptation interventions.

[0063] Taking a high-speed rail station ticket checking scenario as an example, the method provided in this embodiment will be further explained: Database Construction Phase: Four panoramic cameras, two depth cameras, one top-mounted LiDAR, and four sets of environmental sensors were deployed in the ticket hall (approximately 2000 square meters) of a high-speed railway station, and the logs of the turnstile and elevator systems were connected. Forty volunteers, including elderly people, wheelchair users, visually impaired individuals, and a healthy control group, were recruited. Each volunteer wore a lumbar inertial measurement unit and a heart rate monitor, and completed at least 20 rounds of the "entry-security check-ticket check-waiting" process, simultaneously collecting telemetry data from mobility aids and guide devices. Two weeks of real passenger flow data were also collected. A total of over 6200 behavioral samples and approximately 110,000 time slices were obtained, which were aligned, coded, and quality-checked before being entered into the database. The model was trained using the model training method provided in this embodiment, and after validation in one location, the Kendall τ = 0.78, Spearman correlation coefficient 0.83, and low collaborative recognition AUC = 0.87 were obtained, meeting the deployment standards.

[0064] Online Phase: An elderly user using mobility aids enters the ticket hall. An anonymous profile is created and linked to their device data stream. Behavior recognition determines the current behavior as "passing through the ticket gate," entering a critical transition phase. The slice window length automatically switches to 2 seconds. A certain slice feature is: passenger density 0.9 people / square meter, walking speed 0.5 m / s, stride variation coefficient 0.31, minimum net distance 0.4 m, detour rate 0.35, human-machine speed correlation coefficient 0.52, and following distance stability 0.44. The model output C(t) = 0.42, which is lower than the behavior threshold θ = 0.55. The attribution results are 0.58 for the loop side, 0.27 for the machine side, and 0.15 for the human side. The bottleneck side is determined to be the loop side. After three consecutive low values ​​in the slices, the strategy engine triggers intervention, issuing the instruction "open the barrier-free ticket gate + dispatch nearby volunteers" to the operation system. After the intervention, C(t) rebounds from the fourth slice and reaches 0.71 in the sixth slice.

[0065] Aggregation Demonstration: The user's complete "security check - ticket check" behavior contains 38 segments, identifying 2 bottleneck sections (queues in front of the gates, cumulative severity 1.84). After temporal attention aggregation, the behavior coordination index B = 0.68. At that time, there were 17 concurrent behaviors in the ticket hall, with a gate channel resource utilization rate of 0.83, a congestion probability of 0.21, and 3 predicted conflict events. The scene aggregation network output scene coordination index S = 0.62. This scene participated in the site-level aggregation as an in-station travel path node, and the calibration mapping yielded p = 0.81, with an overall coordination index P = 0.74 for the entire station. Key scene identification showed that after downgrading the ticket hall, P decreased by 0.11, and after downgrading the accessible elevator hall, P decreased by 0.08. These two were listed as the current bottleneck scenes, and based on this, the priority recommendation for "adding an accessible security check channel" was output.

[0066] Example 2 This embodiment provides a human-machine environment adaptability analysis device for public places, such as... Figure 2 As shown, it includes: The data acquisition and preprocessing module is used to acquire multi-source human-machine environment data in public places, and perform spatiotemporal alignment, target tracking and data standardization to obtain standardized scene data; The behavior recognition and segmentation module is used to recognize and segment user behavior based on the standardized scene data, obtain multiple behavior segments, and divide each behavior segment into multiple time slices. The feature extraction module is used to extract human-side features, machine-side features, loop-side features, and coupling features for each time slice, and concatenate them into a feature vector; The collaborative adaptation analysis module is used to input the feature vector into the pre-trained collaborative adaptation analysis model and output the slice collaboration index of each time slice and the contribution of the human side, machine side and ring side. The collaborative adaptation evaluation module is used to aggregate the collaborative index of the slices according to the hierarchical relationship between time slices, behavior segments and scenes, to obtain the behavior collaborative index, scene collaborative index and public place overall collaborative index, and generate the corresponding level of barrier-free environment collaborative adaptation evaluation results. An intervention strategy generation module is used to determine the main factors affecting the current collaborative state based on the contributions of the human side, machine side, and environment side, and to generate an intervention strategy based on the main factors. The incremental update module is used to feed the execution result data of the intervention strategy back to the database, so as to use the execution result data to incrementally update the collaborative adaptation analysis model.

[0067] Example 3 This embodiment provides an electronic device, such as... Figure 3 As shown, it includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the human-machine environment adaptability analysis method for public places as described in any of the above embodiments.

[0068] Example 4 This embodiment provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the human-machine environment adaptability analysis method for public places as described in any of the above embodiments.

[0069] Computer-readable storage media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0070] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0071] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0072] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptability analysis of human-machine environment in public places, characterized in that, include: Acquire multi-source human-machine environment data in public places, and perform spatiotemporal alignment, target tracking and data standardization processing to obtain standardized scene data; Based on the standardized scenario data, user behavior is identified and segmented to obtain multiple behavior segments, and each behavior segment is further divided into multiple time slices. Extract human-side features, machine-side features, loop-side features, and coupling features from each time slice, and concatenate them into a feature vector; The feature vector is input into a pre-trained collaborative adaptation analysis model, which outputs the slice collaboration index of each time slice and the contributions of the human side, machine side and ring side. Based on the hierarchical relationship between time slices, behavior segments, and scenes, the slice coordination index is aggregated level by level to obtain the behavior coordination index, scene coordination index, and overall public place coordination index, and corresponding level of barrier-free environment coordination and adaptation evaluation results are generated. Based on the contributions of the human, machine, and environmental sides, the main factors affecting the current collaborative state are determined, and intervention strategies are generated based on the main factors. The execution result data of the intervention strategy is fed back to the database to incrementally update the collaborative adaptation analysis model using the execution result data.

2. The method for adaptability analysis of human-machine environment in public places according to claim 1, characterized in that, The acquisition of multi-source human-machine environment data in public places includes: collecting user movement trajectory and skeletal posture data in the place, operating parameter data of auxiliary equipment, spatial occupancy distribution and physical quantity data of the place environment, and business event log data; The process of performing spatiotemporal alignment, target tracking, and data standardization yields standardized scene data, including: Connect all data acquisition devices to a unified time base, add a unified timestamp to each data stream, establish a scene semantic raster map containing a static occupancy layer, a semantic layer, and a dynamic layer, and map the multi-source data to the scene semantic raster map. The system detects and tracks users and auxiliary equipment in the acquired video stream, performs cross-field association of the same target in different fields of view based on spatiotemporal continuity, and establishes anonymous identifiers for associated targets. Data that has undergone spatiotemporal alignment, spatial registration, and target tracking is encoded in a unified data format, and abnormal and missing data is processed to obtain standardized scene data with a unified time base, unified spatial coordinates, and unified data format.

3. The method for adaptability analysis of human-machine environment in public places according to claim 2, characterized in that, The process of identifying and segmenting user behavior based on the standardized scenario data includes: A trajectory kinematic sequence is constructed using instantaneous velocity, rate of change of heading angle, and trajectory curvature. Accumulation and change point detection are used to locate abrupt changes in the trajectory kinematic sequence, and the user trajectory is divided into multiple trajectory primitives based on the abrupt changes. A hidden Markov model is constructed by using the behavior types in the preset behavior type set as hidden states and the emission probability by weighted fusion of the functional area transfer criterion, trajectory primitive criterion and event triggering criterion. The Viterbi algorithm is used to solve the hidden state sequence with the maximum joint probability in the hidden Markov model. The trajectory intervals corresponding to consecutive identical behavior types in the hidden state sequence are determined as behavior segments, and corresponding behavior type labels are assigned to each behavior segment.

4. The method for adaptability analysis of human-machine environment in public places according to claim 1, characterized in that, The extraction of human-side features, machine-side features, loop-side features, and coupling features for each time slice includes: Step speed, stride variability, joint range of motion, path tortuosity, number of stops, head orientation entropy, heart rate change relative to baseline, posture stability, and behavioral stage markers were extracted from time slices as human-side features. The equipment response delay, number of times the user actively requests assistance, navigation and positioning error, battery percentage, ratio of motor current to rated current, and equipment mission status are extracted from the time slice as machine-side features. Passenger flow density, dynamic obstacle occupancy rate, queue length, illumination, noise, and ground condition are extracted from time slices as ring-side features; Human-machine collaboration features are extracted based on the correlation between the speed sequence of the user and the auxiliary equipment and the stability of the following distance. Human-loop conflict features are extracted based on the minimum clearance between the user and dynamic obstacles and the detour rate of the user's actual path relative to the shortest passable path. Machine-loop adaptation features are extracted based on the net width of the channel and the matching degree of the positioning environment.

5. The method for adaptability analysis of human-machine environment in public places according to claim 1, characterized in that, The step of inputting the feature vector into a pre-trained collaborative adaptation analysis model and outputting the slice collaboration index and the contributions of the human side, machine side, and ring side for each time slice includes: The feature vector is used as the initial feature of the node to construct a heterogeneous graph for each time slice. The heterogeneous graph is input into a graph attention network for spatial encoding, and the temporal subsequences inside the time slice are input into a temporal convolutional network for temporal encoding. The results of spatial encoding and temporal encoding are fused to obtain the slice synergy index of each time slice. Using the feature vector of the current time slice as the input to be attributed, and the average feature of historical samples corresponding to the current behavior type as the attribution baseline, the contribution value of the slice synergy index relative to each feature is calculated using the integral gradient method. The contribution values ​​are then merged and normalized according to the human side, machine side, and ring side to which each feature belongs, to obtain the human side contribution, machine side contribution, and ring side contribution.

6. The method for adaptability analysis of human-machine environment in public places according to claim 5, characterized in that, The graph attention network is a multi-layer multi-head graph attention network, and the temporal convolutional network is a dilated causal convolutional network, with the dilation factor increasing layer by layer along the network hierarchy; The results of the spatial encoding and the temporal encoding are concatenated and mapped through a fully connected layer to obtain collaborative latent variables. The collaborative latent variables are then mapped to the slice collaborative indexes using a Sigmoid function with a learnable steepness coefficient k and a center offset parameter.

7. The method for adaptability analysis of human-machine environment in public places according to claim 1, characterized in that, The process involves aggregating the slice coordination indices hierarchically according to the relationship between time slices, behavior segments, and scenes to obtain behavior coordination indices, scene coordination indices, and overall public space coordination indices. Corresponding levels of accessibility environment coordination adaptation evaluation results are then generated, including: For each behavior segment, a bottleneck interval is determined based on the slice coordination index of each time slice within the behavior segment, and the slice coordination index and its temporal weight of each time slice are aggregated to obtain the behavior coordination index corresponding to the behavior segment; the behavior-level accessibility environment coordination adaptation evaluation result of the behavior segment is determined based on the relationship between the behavior coordination index and the coordination evaluation benchmark of the corresponding behavior type. For each scenario, concurrent behaviors within the scenario are taken as behavior nodes. The resource competition relationship between behavior nodes is determined based on the spatiotemporal occupancy relationship of each concurrent behavior on shared resources. The scenario collaboration index corresponding to each behavior node and the resource competition relationship are aggregated to obtain the scenario collaboration index corresponding to the scenario. The scenario-level barrier-free environment collaboration adaptation evaluation result of the scenario is determined based on the relationship between the scenario collaboration index and the scenario collaboration evaluation benchmark. For the public place, the relationship between the scenes is determined based on the travel relationship between each scene, and the overall coordination index of the public place is obtained by aggregating the scene coordination index corresponding to each scene and the relationship between the scenes. The overall accessibility environment coordination and adaptation evaluation result of the public place is determined based on the relationship between the overall coordination index of the public place and the coordination evaluation benchmark of the public place.

8. The method for adaptability analysis of human-machine environment in public places according to claim 1, characterized in that, The generation of intervention strategies based on the main factors includes: Determine whether the slice synergy index is lower than the synergy threshold corresponding to the current behavior type for multiple consecutive time slices. If so, determine the corresponding intervention strategy based on the main factor side with the greatest contribution. When the main factor side is the ring side, intervention strategies such as opening backup channels, adjusting passenger flow guidance, or dispatching staff to designated locations are generated based on passenger flow density or facility occupancy status. When the main factor is the machine side, intervention strategies such as replacing the adapted equipment, adjusting the equipment deployment point, or adjusting the equipment task parameters are generated based on the equipment response delay or equipment positioning error. When the main factor is human, personalized guidance information or intervention strategies for manual assistance in making appointments are generated based on the difference between the user's ability characteristics and the current scenario's normal level.

9. The method for adaptability analysis of human-machine environment in public places according to claim 5, characterized in that, The pre-training process of the collaborative adaptation analysis model includes: Multi-source data on multiple users completing their scheduled trips at the target location is collected, and after spatiotemporal alignment, standardized coding, and quality inspection, it is stored in the human-machine-environment database. The training set and validation set are divided into complete venues. Cross-validation is used, with all data of one type of venue reserved as the validation set and the data of the remaining venues used as the training set. Using behavioral-level result labels and order labels of sample pairs generated from business event logs as weak supervision signals, and the weighted sum of the mean square error of the slice synergy index after behavioral-level aggregation with the behavioral-level result labels, the sorting hinge loss with the order labels, and the parameter L2 regularization term as the total loss function, the graph attention network, temporal convolutional network, and fully connected layer parameters are trained end-to-end. After training, the validation set is used to calculate the model acceptance index. Once the index is met, the collaborative adaptation analysis model is obtained.

10. A human-machine environment adaptability analysis device for public places, characterized in that, The device includes: The data acquisition and preprocessing module is used to acquire multi-source human-machine environment data in public places, and perform spatiotemporal alignment, target tracking and data standardization to obtain standardized scene data; The behavior recognition and segmentation module is used to recognize and segment user behavior based on the standardized scene data, obtain multiple behavior segments, and divide each behavior segment into multiple time slices. The feature extraction module is used to extract human-side features, machine-side features, loop-side features, and coupling features for each time slice, and concatenate them into a feature vector; The collaborative adaptation analysis module is used to input the feature vector into the pre-trained collaborative adaptation analysis model and output the slice collaboration index of each time slice and the contribution of the human side, machine side and ring side. The collaborative adaptation evaluation module is used to aggregate the collaborative index of the slices according to the hierarchical relationship between time slices, behavior segments and scenarios, to obtain the behavior collaborative index, scenario collaborative index and public place overall collaborative index, and generate the corresponding level of barrier-free environment collaborative adaptation evaluation results. An intervention strategy generation module is used to determine the main factors affecting the current collaborative state based on the contributions of the human side, machine side, and environment side, and to generate an intervention strategy based on the main factors. The incremental update module is used to feed the execution result data of the intervention strategy back to the database, so as to use the execution result data to incrementally update the collaborative adaptation analysis model.

Images