Crowd behavior pedigree modeling method based on multi-modal identification and high-precision positioning

Abstract

The invention discloses a crowd behavior pedigree modeling method based on multi-modal identification and high-precision positioning. The method comprises the following steps: step 1, collecting sample data; 2, performing high-precision behavior data fitting; step 3, extracting a multi-modal behavior factor; 4, constructing a behavior type pedigree; and step 5, database integration and visual display. According to the method, accurate alignment of behaviors and spatial information is realized by fusing video images and positioning data and combining target detection, posture recognition and spatial reconstruction. And then multiple types of behavior factors including speed, direction, attitude, facility interaction and group interaction are extracted, a behavior graph structure is constructed, a graph neural network is adopted to identify behavior types, and a structured pedigree model is formed. The method can cope with application scenes such as public space behavior recognition, city design optimization and smart city management, and fine analysis and presentation of a target space crowd behavior mode are realized.

Description

Technical Field

[0001] This invention belongs to the field of urban behavior research, specifically involving a method for modeling the behavioral spectrum of crowds based on multimodal recognition and high-precision positioning. Background Technology

[0002] With the deepening of the new urbanization process, urban construction is gradually shifting from "incremental expansion" to "quality improvement of existing resources." However, many current urban design schemes are still at the rudimentary stage of "seeing things but not people," lacking a deep understanding of the behavioral characteristics of different groups, differences in spatial perception, and the actual needs of the population.

[0003] Against this backdrop, public space behavior spectrum modeling has become a crucial technological approach for achieving human-centered urban design. Building upon traditional urban design methods that rely on experience-based judgment and qualitative analysis, it can further refine the capture of complex crowd behaviors in cities, systematically identify the spatial needs of diverse populations in specific contexts, and thus enhance the responsiveness of urban spaces. Therefore, achieving high-precision, multi-level, and structured modeling of crowd behavior in urban public spaces possesses significant innovative and application value.

[0004] First, behavioral spectrum modeling in public spaces helps improve the adaptability and sophistication of urban public spaces. By deeply identifying and structurally analyzing the behaviors of different groups, the interaction between individuals and spaces can be more clearly depicted, thus providing a better basis for the morphological design and facility layout of public spaces. Second, behavioral spectrum modeling helps respond to diverse population needs. Users of urban public spaces exhibit significant differences in gender, age, physical condition, and usage habits. Spectrum modeling can accurately identify the activities and space usage preferences of specific groups, facilitating the development of personalized space design strategies and promoting the refined transformation of urban design. Behavioral spectrum modeling also helps provide decision support for the governance and management of public spaces. Urban public spaces not only serve as venues for daily activities but also as important locations for social interaction and the occurrence of emergencies. Extracting behavioral paths and group characteristics can assist in the prediction of potential risks, providing real-time perception capabilities and a foundation for forward-looking judgment in urban governance.

[0005] Current behavioral analysis techniques largely focus on identifying and studying single attributes of behavior, and have not yet formed a complete technical system for modeling population behavior spectrums. On the one hand, some techniques concentrate on behavioral trajectory identification, such as path extraction, speed calculation, and direction determination, and are relatively mature in simple indicator analysis. On the other hand, some research focuses on behavioral posture recognition and group social interaction, which can reflect the behavioral characteristics or communication patterns of individuals. However, current technologies do not address the coupling relationships between different behavioral information, and there are still gaps in behavioral structure modeling; a hierarchical spectrum covering complete behavioral information of a population has not yet been constructed.

[0006] Meanwhile, current technologies largely focus on crowd behavior information, neglecting the identification and coupled analysis of spatial information that carries such behavior. On the one hand, spatial identification primarily relies on 3D reconstruction technology for model rebuilding, lacking a spatial element classification and identification framework oriented towards behavior analysis, and lacking detailed characterization of the functions, characteristics, and behavioral adaptability of spatial facilities. On the other hand, technologies for identifying and analyzing spatial and behavioral interactions are mostly geared towards specific fields such as navigation, communication, and monitoring, lacking a modeling framework for spatial and behavioral collaboration in the large-scale context of urban planning and design. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a crowd behavior spectrum modeling method based on multimodal recognition and high-precision positioning. This method integrates video images and positioning data to extract behavioral factors including speed, direction, posture, facility interaction, and group interaction. It then uses graph structures and graph neural networks to identify structured spectrum models, thereby achieving a refined analysis and presentation of crowd behavior patterns.

[0008] The objective of this invention can be achieved through the following technical solutions: A method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning includes the following steps: (1) Sample data collection A public space sample was selected, and geographic information data of the sample area was obtained. Simultaneously, continuous video image data and spatial positioning data were acquired using a video capture device equipped with a high-precision positioning system. The spatial positioning data and geographic information data were then integrated into a geographic information platform, and the data was aligned and formatted using a unified geographic coordinate system.

[0009] (2) High-precision behavioral data fitting Based on video image data, continuous identification, trajectory tracking, and action recognition are performed on individuals within a crowd using target detection, tracking, and pose recognition algorithms, forming a set of behavioral data points. A 3D model of the spatial samples is obtained through 3D reconstruction methods, and spatial elements are identified using semantic recognition and segmentation algorithms. Semantic labels are assigned to these spatial elements, forming a set of spatial element coordinates. Based on the camera's intrinsic and extrinsic parameters and high-precision positioning information from the video acquisition equipment, a projection matrix is ​​constructed to map the behavioral data point set, 3D model, and spatial element coordinate set to the world coordinate system. This data is then integrated into a geographic information platform, forming a basic information database of the sample behaviors.

[0010] (3) Extraction of multimodal behavioral factors Based on a sample behavior database, multi-dimensional feature calculations are performed on individual behavioral processes to extract five core factors, including behavioral speed, trajectory direction, posture combination, facility interaction, and group interaction. Using the individual's behavioral features in each time series as nodes, the five behavioral factors as edges, and individual ID and behavioral point coordinates as attributes, a behavioral factor graph structure is constructed, forming a behavioral factor graph structure database.

[0011] (4) Behavioral type genealogy modeling Based on a factor graph structure database, multi-level filtering channels are constructed through attribute co-occurrence and temporal sequence associations to extract subgraph sets with specific combination features, forming a multi-channel path graph. Using a graph neural network algorithm, typical factor combination structures in different channel path graphs are identified, forming a structured behavior type genealogy, which is stored in a dynamic behavior type genealogy library. Simultaneously, based on the behavioral foundation information library of all stored public space samples, each space sample is clustered to form a dynamically expanding genealogy scene, resulting in a dynamic behavior type genealogy library under different sample types. The feature behavior factor weights under different sample types are calculated, and the behavior factor weights of the graph structure are dynamically adjusted in subsequently inserted samples.

[0012] (5) Database integration and visualization The extracted behavioral basic information database, behavioral factor graph structure database, and behavioral type dynamic genealogy database are stored in a unified data management system, which supports behavioral event query, type statistical analysis and graph structure display, and presents individual trajectory, behavioral pattern distribution and genealogy evolution process through a visual interface.

[0013] Furthermore, the behavioral data point set described in step two includes frame number, individual ID, pixel x-coordinate, pixel y-coordinate, and posture information. The posture information includes upper limb movement information, lower limb movement information, and body posture. Upper limb and lower limb movement information includes whether the limbs are moving, and body posture includes standing, bending, and sitting postures. The spatial element coordinate set includes the point cloud coordinates contained in each spatial element, which includes commercial facilities, entertainment facilities, landscape facilities, and transportation facilities. The final sample behavioral basic information database includes identity information, time series information, continuous positioning information, and spatial element information, stored in a GeoJSON structured format, preserving the geometric visualization information and attribute data of the data.

[0014] Furthermore, step three involves calculating multi-dimensional features of individual behavior processes, which means extracting factors including behavioral speed, trajectory direction, posture combination, facility interaction, and group interaction features. Behavior points are grouped based on individual IDs, and frame numbers are sorted in ascending order using a linear time sorting algorithm.

[0015] The calculation of behavioral speed is first done through Calculate the spatial distance between adjacent action points within a group, where Let i be the two-dimensional coordinates of the i-th row point. For the first The timestamps of each action point are used to calculate the instantaneous velocity. , For the first The timestamps of each action point are used to calculate the instantaneous velocity. The velocity is smoothed using a Kalman filter algorithm to obtain the individual behavior velocity; among which... The spatial distance between adjacent action points. The frame interval.

[0016] The trajectory direction features are calculated using the trajectory vector method, which generates vectors of adjacent points. Using the positive y-axis of the coordinate system as the reference, calculate the vector. Angle with north ,in , is the component vector in the x-direction. , is the component vector in the y direction, and this angle is recorded as the trajectory direction; The calculation of attitude combination information first involves adding combination labels to 12 combination methods, and then using these combination labels as attitude combination information. The combination label format is as follows: , The upper limb movements are set A, the lower limb movements are set B, the body postures are set C, and S is the group label.

[0017] The calculation of facility interaction information first generates buffer zones for different types of facilities based on the coordinate information of spatial elements, and establishes rules for the effective interaction area of ​​facilities. in p Let R be the trajectory point of the crowd, and R be the interaction radius. Based on individual trajectory coordinates, the system detects whether a crowd has entered the effective interaction area of ​​the facility. Simultaneously, it combines the effective dwell time to distinguish between "passing by" and "actual interaction" states, calculating the dwell time of the crowd within the facility area. T And set the interaction time. ,judge In order to generate interaction and record state, If you are just passing by, do not record the interaction state.

[0018] The calculation of group interaction features first obtains the feature vector of each individual based on continuous localization information and behavioral posture information. ,That Let be the coordinates of the individual at time t. For trajectory velocity and direction, Based on facial orientation, a graph convolutional network is used to identify group interaction states, including those of people in pairs, scattered groups, and large groups.

[0019] The five behavioral factors mentioned above are edges. When the dynamic genealogy library of behavioral types returns the feature behavioral factor weights under different sample types, the type of the current sample is first determined based on the type threshold of the spatial samples, and the weight vector of the five behavioral factors of that type is extracted. The five components of the vector represent the weight values ​​of behavioral speed, trajectory direction, posture combination, facility interaction, and group interaction features, respectively. Then, corresponding weight attributes are added to the five types of behavioral factors in the graph structure.

[0020] Furthermore, step four involves constructing multi-level filtering channels through attribute co-occurrence and temporal correlation, specifically including unit behavior channels, composite behavior channels, and collective behavior channels. The unit behavior channel extracts the basic action change path completed by an individual within a unit of time, containing factor combinations within that unit of time, suitable for instantaneous behavior recognition. The composite behavior channel consists of multiple unit behaviors, reflecting the combination of unit behaviors within a complete passage time range, suitable for continuous behavior structure extraction. The collective behavior channel constructs behavior path combinations based on the interaction relationships between individuals, used to capture the collaborative mechanisms between individuals in a group, suitable for group-level behavior pattern recognition and social relationship mining. The unit of time is set according to the temporal stability of human action changes, preferably every 3 seconds as a recognition unit cycle. In a video recognition environment with a frame rate of 25fps, this corresponds to a 75-frame image sequence, which can include the completion process of a single basic action.

[0021] The subgraph sets of the three types of behavioral channels are filtered based on the attribute co-occurrence weight function and the temporal correlation function. The core filtering function for channel construction is defined as follows: , Where G represents the original behavioral factor graph structure, For any subgraph in the graph, Subgraph The strength scoring function for attribute co-occurrence and time-series association. The preset filtering threshold, Indicates the channel type.

[0022] The structured behavior type hierarchy refers to the embedding modeling and hierarchy classification of subgraph structures in a multi-channel path graph using graph neural network algorithms. Specifically, it includes the following steps: First, for each subgraph in the behavior channel path graph Extract the behavioral factor features of all its nodes and construct a node feature matrix. ,in This indicates the number of nodes in the subgraph. For each node, feature dimensions are defined, including behavior speed, trajectory direction, posture combination, facility interaction, and group interaction features; simultaneously, an adjacency matrix is ​​constructed. This represents the structural connection relationship between any two nodes in the subgraph. Next, structural enhancement is performed on the adjacency matrix by calculating the normalized adjacency matrix with added self-connections: , in, It is the identity matrix. This represents the adjacency matrix after adding self-connections. for The degree matrix, It is a symmetric normalized adjacency matrix. and These represent the row and column indices of the node in the subgraph, respectively; further, multi-layer graph convolution operations are used to calculate and update the node features, the th... The feature update calculation formula for the layer is: , in, This represents the layer index of the graph convolutional layer. Indicates the first Layer node embedding matrix, For the first The trainable weight matrix of the layer, It is a non-linear activation function. The aforementioned normalized adjacency matrix; after L layers of convolution operations, the final node embedding is obtained. By embedding the final node matrix Perform average pooling to obtain the global representation vector of the subgraph. The data is then fed into a fully connected classifier, and the behavior type prediction result is output through the Softmax function. , where C is the number of predefined behavior categories. The result is a structured behavior type spectrum.

[0023] The calculation of feature behavior factor weights under different sample types refers to extracting behavior subgraph sets under different spatial sample types, statistically analyzing the frequencies of the five types of behavior factor edges, and calculating their relative importance weights according to the following normalization formula: , in, This represents the weight value of factor f in the j-th spatial sample. This represents the cumulative frequency of the edge corresponding to factor f in this type. Calculated This refers to the weights of the five behavioral factors for this sample type.

[0024] The beneficial effects of this invention are: 1. This patent integrates multimodal information recognition technology with high-precision positioning technology, which can reconstruct and extract information from the micro-scene of individual behavior based on fine sample data of visual images and location information. While maintaining the accuracy of behavioral and spatial information, it effectively preserves the continuity and integrity of behavioral data, laying a data foundation for modeling crowd behavior in complex scenarios.

[0025] 2. This patent fully considers the spatial elements upon which crowd behavior relies. It extracts structural information such as spatial facilities and site interfaces through methods such as semantic recognition and 3D reconstruction, and performs spatial mapping and precise fitting with behavioral trajectory data. This can effectively reveal the coupling relationship between the spatial environment and crowd behavior, improve the spatial interpretability of behavior recognition, and provide key support for behavior analysis and spatial optimization in complex scenarios.

[0026] 3. This patent constructs five categories of behavioral factors, including speed, direction, posture, facility interaction, and group interaction, and expresses them in a graph structure. This enables multi-dimensional and multi-level modeling of individual behavioral characteristics, further enhancing the ability to identify behaviors and analyze behavioral mechanisms, and meeting the research and application needs for refined expression of behavior in public spaces.

[0027] 4. This patent introduces a graph neural network model to embed and model behavioral subgraphs in multi-channel path graphs and classify them into hierarchical categories. It can automatically discover typical behavioral combinations and evolution patterns, effectively overcoming the problems of weak generalization ability and insufficient analysis depth of traditional subjective analysis methods when facing multimodal behavioral factors.

[0028] 5. This patent constructs a dynamic behavior type genealogy library and adaptively adjusts the weight configuration of behavior factors based on sample type, enabling the method to have good generalization and scalability. It can flexibly adapt to the analysis needs of different public space types, behavior scenarios and target groups, and improve the application breadth and intelligence level of the behavior modeling system.

[0029] 6. This patent integrates a behavioral basic information database, a factor graph structure database, and a genealogical evolution database to build a unified data management platform. It is equipped with a visual interface to realize the interactive presentation of behavioral trajectory backtracking, behavioral pattern recognition results display, and genealogical structure graph, providing an operable and quantifiable decision support tool for smart city management, crowd behavior prediction, and public space optimization. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1 As shown, the technical solution of the present invention will be described in detail below, taking the space of Laohuqiao Street in Nanjing City as an example.

[0033] A method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning includes the following steps: (1) Sample data collection. Geographic information data of Laohuqiao Street was acquired, and continuous video image data and spatial positioning data were acquired through video acquisition equipment with high-precision positioning devices. Spatial positioning data and geographic information data were placed into the geographic information platform, and the data were aligned and formatted through a unified geographic coordinate system.

[0034] (2) High-precision behavioral data fitting. Based on video image data, continuous identification, trajectory tracking, and action recognition of individuals in the crowd are performed using target detection, tracking, and posture recognition algorithms to form a behavioral data point set. A three-dimensional model of the Laohuqiao Street space is obtained through three-dimensional reconstruction methods. Spatial elements are identified through semantic recognition and segmentation algorithms, and semantic labels are assigned to the spatial elements to form a spatial element coordinate set. Based on the camera intrinsic and extrinsic parameters and high-precision positioning information of the video acquisition equipment, a projection matrix is ​​constructed to map the behavioral data point set, three-dimensional model, and spatial element coordinate set to the world coordinate system and place them into the geographic information platform to form a basic information database of spatial sample behavior in Laohuqiao Street.

[0035] The behavioral data point set includes frame number, individual ID, pixel x-coordinate, pixel y-coordinate, and posture information. The posture information includes upper limb movement information, lower limb movement information, and body posture. Upper limb and lower limb movement information includes whether the limbs are moving, and body posture includes standing, bending over, and sitting postures. The spatial element coordinate set includes the point cloud coordinates contained in each spatial element, which includes commercial facilities, entertainment facilities, landscape facilities, and transportation facilities. The final sample behavioral basic information database includes identity information, time series information, continuous location information, and spatial element information, stored in a GeoJSON structured format, preserving the geometric visualization information and attribute data of the data.

[0036] (3) Multimodal behavioral factor extraction. Based on the sample behavioral information database, multi-dimensional feature calculations are performed on individual behavioral processes to extract five core factors including behavioral speed, trajectory direction, posture combination, facility interaction and group interaction features. Using the individual's behavioral features in each time series as nodes, the five behavioral factors as edges, and the individual ID and behavioral point coordinates as attributes, a behavioral factor graph structure is constructed to form a behavioral factor graph structure database.

[0037] First, the behavior points are grouped based on individual IDs, and then the frame numbers are sorted in ascending order using a linear time sorting algorithm.

[0038] The calculation of behavioral speed is first done through Calculate the spatial distance between adjacent action points within a group, where Let i be the two-dimensional coordinates of the i-th row point. For the first The timestamps of each action point are used to calculate the instantaneous velocity. , For the first The timestamps of each action point are used to calculate the instantaneous velocity. The velocity is smoothed using a Kalman filter algorithm to obtain the individual behavior velocity; among which... The spatial distance between adjacent action points. The frame interval.

[0039] The trajectory direction features are calculated using the trajectory vector method, which generates vectors of adjacent points. Using the positive y-axis of the coordinate system as the reference, calculate the vector. Angle with north ,in , is the component vector in the x-direction. , is the component vector in the y direction, and this angle is recorded as the trajectory direction; The calculation of attitude combination information first involves adding combination labels to 12 combination methods, and then using these combination labels as attitude combination information. The combination label format is as follows: , The upper limb movements are set A, the lower limb movements are set B, the body postures are set C, and S is the group label.

[0040] The calculation of facility interaction information first generates buffer zones for different types of facilities based on the coordinate information of spatial elements, and establishes rules for the effective interaction area of ​​facilities. in p Let R be the trajectory point of the crowd, and R be the interaction radius. Based on individual trajectory coordinates, the system detects whether a crowd has entered the effective interaction area of ​​the facility. Simultaneously, it combines the effective dwell time to distinguish between "passing by" and "actual interaction" states, calculating the dwell time of the crowd within the facility area. T And set the interaction time. ,judge In order to generate interaction and record state, If you are just passing by, do not record the interaction state.

[0041] The calculation of group interaction features first obtains the feature vector of each individual based on continuous localization information and behavioral posture information. ,That Let be the coordinates of the individual at time t. For trajectory velocity and direction, Based on facial orientation, a graph convolutional network is used to identify group interaction states such as buddy groups, scattered groups, and large groups. When the dynamic behavioral type genealogy library returns the feature behavioral factor weights for different sample types, the type of the current sample is first determined based on the type threshold of the spatial samples, and the weight vectors of the five behavioral factors for that type are extracted. The five components of the vector represent the weight values ​​of behavioral speed, trajectory direction, posture combination, facility interaction, and group interaction features, respectively. Then, corresponding weight attributes are added to the five types of behavioral factors in the graph structure.

[0042] (4) Behavioral type genealogy modeling. Based on the factor graph structure database, multi-level filtering channels are constructed through attribute co-occurrence and temporal correlation to extract subgraph sets with specific combination features, forming a multi-channel path graph. Through graph neural network algorithms, typical factor combination structures in different channel path graphs are identified to form a structured behavioral type genealogy, which is stored in the dynamic behavioral type genealogy library. At the same time, based on the behavioral basic information database of all stored Tiger Bridge Street spatial samples, each spatial sample is clustered to form a dynamically expanded genealogy scene, resulting in a dynamic behavioral type genealogy library under different sample types. The feature behavioral factor weights under different sample types are calculated, and the behavioral factor weights of the graph structure are dynamically adjusted in the subsequently inserted samples. The filtering channels specifically include unit behavior channels, composite behavior channels, and collective behavior channels. The unit behavior channel is used to extract the basic action change path completed by an individual within a unit of time, containing the combination of factors within a unit of time, and is suitable for the identification of instantaneous behavior. The composite behavior channel is composed of multiple unit behaviors, reflecting the combination of unit behaviors of an individual within a complete passage time range, and is suitable for the structural extraction of continuous behavior. The collective behavior channel constructs a combination of behavior paths based on the interaction relationship between individuals, used to capture the collaborative mechanism between individuals in a group, and is suitable for group-level behavior pattern recognition and social relationship mining.

[0043] The subgraph sets of the three types of behavioral channels are filtered based on the attribute co-occurrence weight function and the temporal correlation function. The core filtering function for channel construction is defined as follows: , Where G represents the original behavioral factor graph structure, For any subgraph in the graph, Subgraph The strength scoring function for attribute co-occurrence and time-series association. The preset filtering threshold, Indicates the channel type.

[0044] The structured behavior type genealogy refers to the embedding modeling and genealogy classification of subgraph structures in a multi-channel path graph using graph neural network algorithms, specifically including the following steps: First, for each subgraph in the behavior channel path graph Extract the behavioral factor features of all its nodes and construct a node feature matrix. ,in This indicates the number of nodes in the subgraph. For each node, feature dimensions are defined, including behavior speed, trajectory direction, posture combination, facility interaction, and group interaction features; simultaneously, an adjacency matrix is ​​constructed. This represents the structural connection relationship between any two nodes in the subgraph. Next, structural enhancement is performed on the adjacency matrix by calculating the normalized adjacency matrix with added self-connections: , in, It is the identity matrix. This represents the adjacency matrix after adding self-connections. for The degree matrix, It is a symmetric normalized adjacency matrix. and These represent the row and column indices of the node in the subgraph, respectively; further, multi-layer graph convolution operations are used to calculate and update the node features, the th... The feature update calculation formula for the layer is: , in, This represents the layer index of the graph convolutional layer. Indicates the first Layer node embedding matrix, For the first The trainable weight matrix of the layer, It is a non-linear activation function. The aforementioned normalized adjacency matrix; after L layers of convolution operations, the final node embedding is obtained. By embedding the final node matrix Perform average pooling to obtain the global representation vector of the subgraph. The data is then fed into a fully connected classifier, and the behavior type prediction result is output through the Softmax function. , where C is the number of predefined behavior categories. The result is a structured behavior type spectrum.

[0045] The calculation of feature behavior factor weights under different sample types refers to extracting behavior subgraph sets under different spatial sample types, statistically analyzing the frequencies of the five types of behavior factor edges, and calculating their relative importance weights according to the following normalization formula: , in, This represents the weight value of factor f in the j-th spatial sample. This represents the cumulative frequency of the edge corresponding to factor f in this type. Calculated This refers to the weights of the five behavioral factors for this sample type.

[0046] (5) Database integration and visualization. The extracted behavioral basic information database, behavioral factor graph structure database, and behavioral type dynamic genealogy database are stored in a unified data management system, which supports behavioral event query, type statistical analysis and graph structure display, and presents individual trajectory, behavioral pattern distribution and genealogical evolution process through a visualization interface.

Claims

1. A method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning, characterized in that, Includes the following steps: Step 1: Sample Data Collection Select public space samples, obtain geographic information data of the spatial sample area, and at the same time, acquire continuous video image data and spatial positioning data through video acquisition equipment with high-precision positioning devices; put the spatial positioning data and geographic information data into the geographic information platform, and align and format the data through a unified geographic coordinate system; Step 2: High-precision behavioral data fitting Based on video image data, continuous identification, trajectory tracking, and action recognition are performed on individuals in a crowd using target detection, tracking, and pose recognition algorithms, forming a set of behavioral data points. A 3D model of the spatial samples is obtained through 3D reconstruction methods, and spatial elements are identified using semantic recognition and segmentation algorithms. Semantic labels are assigned to these spatial elements, forming a set of spatial element coordinates. Based on the camera intrinsic and extrinsic parameters and high-precision positioning information of the video acquisition equipment, a projection matrix is ​​constructed to map the behavioral data point set, 3D model, and spatial element coordinate set to the world coordinate system. This data is then placed into a geographic information platform, forming a basic information database of sample behavior. Step 3: Extraction of Multimodal Behavioral Factors Based on the sample behavior information database in step two, multi-dimensional feature calculations are performed on individual behavior processes to extract five core factors including behavior speed, trajectory direction, posture combination, facility interaction and group interaction features. Using the individual's behavioral characteristics in each time series as nodes, five types of behavioral factors as edges, and individual ID and behavioral point coordinates as attributes, a behavioral factor graph structure is constructed to form a behavioral factor graph structure database. Step 4: Behavioral Type Spectrum Modeling Based on the behavioral factor graph structure database in step three, multi-level filtering channels are constructed through attribute co-occurrence and temporal sequence associations. Subgraph sets with specific combination features are extracted to form a multi-channel path graph. Through graph neural network algorithms, typical factor combination structures in different channel path graphs are identified to form a structured behavioral type genealogy, which is stored in a dynamic behavioral type genealogy library. At the same time, based on the behavioral basic information library of all stored public space samples, each space sample is clustered to form a dynamically expanded genealogy scene, resulting in a dynamic behavioral type genealogy library under different sample types. The feature behavioral factor weights under different sample types are calculated, and the behavioral factor weights of the graph structure in step three are dynamically adjusted in subsequently inserted samples. Step 5: Database Integration and Visualization The extracted behavioral basic information database, behavioral factor graph structure database, and behavioral type dynamic genealogy database are stored in a unified data management system, which supports behavioral event query, type statistical analysis and graph structure display, and presents individual trajectory, behavioral pattern distribution and genealogy evolution process through a visual interface.

2. The method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning according to claim 1, characterized in that, In step two, the behavioral data point set includes frame number, individual ID, pixel x-coordinate, pixel y-coordinate, and posture information. The posture information includes upper limb movement information, lower limb movement information, and body posture. The upper limb and lower limb movement information includes whether the limbs are moving, and the body posture includes standing, bending over, and sitting postures. The spatial element coordinate set includes the point cloud coordinates contained in each spatial element. The spatial elements include commercial facilities, entertainment facilities, landscape facilities, and transportation facilities. The final sample behavioral basic information database includes identity information, time series information, continuous positioning information, and spatial element information, stored in the GeoJSON structured format, preserving the geometric visualization information and attribute data of the data.

3. The method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning according to claim 2, characterized in that, In step three, multi-dimensional feature calculations are performed on the individual behavior process. First, the behavior points are grouped based on the individual ID, and the frame numbers are sorted in ascending order using a linear time sorting algorithm. The calculation of behavioral speed is first done through Calculate the spatial distance between adjacent action points within a group, where Let i be the two-dimensional coordinates of the i-th row point. For the first The timestamps of each action point are used to calculate the instantaneous velocity. The velocity is smoothed using a Kalman filter algorithm to obtain the individual behavior velocity; among which... The spatial distance between adjacent action points. For frame interval; The trajectory direction features are calculated using the trajectory vector method, which generates vectors of adjacent points. Using the positive y-axis of the coordinate system as the reference, calculate the vector. Angle with north ,in , is the component vector in the x-direction. , is the component vector in the y direction, and this angle is recorded as the trajectory direction; The calculation of attitude combination information first involves adding combination labels to 12 combination methods, and then using these combination labels as attitude combination information; the combination label format is as follows: ， The upper limb movements are set A, the lower limb movements are set B, the body postures are set C, and S is the group label; The calculation of facility interaction information first generates buffer zones for different types of facilities based on the coordinate information of spatial elements, and establishes rules for the effective interaction area of ​​facilities. in p Let R be the trajectory point of the crowd, and R be the interaction radius. Based on individual trajectory coordinates, the system detects whether a crowd has entered the facility's effective interaction area. Simultaneously, it combines the effective dwell time to distinguish between "passing by" and "real interaction" states, calculating the crowd's dwell time within the facility area. T And set the interaction time. ,judge In order to generate interaction and record state, If you are just passing by, do not record the interaction state; The calculation of group interaction features first obtains the feature vector of each individual based on continuous localization information and behavioral posture information. ,That Let be the coordinates of the individual at time t. For trajectory velocity and direction, Based on facial orientation, a graph convolutional network is used to identify group interaction states, including those of people in pairs, scattered groups, and large groups.

4. The method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning according to claim 3, characterized in that, In step three, the five behavioral factors are used as edges. When the dynamic genealogy library of behavioral types returns the feature behavioral factor weights for different sample types, the type of the current sample is first determined based on the type threshold of the spatial samples, and the weight vector of the five behavioral factors for that type is extracted. The five components of the vector represent the weight values ​​of behavioral speed, trajectory direction, posture combination, facility interaction, and group interaction features, respectively. Then, corresponding weight attributes are added to the five types of behavioral factors in the graph structure.

5. The method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning according to claim 4, characterized in that, In step four, multi-level filtering channels are constructed through attribute co-occurrence and temporal correlation relationships. Specifically, these include unit behavior channels, composite behavior channels, and collective behavior channels. The unit behavior channel extracts the basic action change path completed by an individual within a unit of time, containing factor combinations within that unit of time, and is suitable for recognizing instantaneous behavior. The composite behavior channel consists of multiple unit behavior combinations, reflecting the combination of unit behaviors of an individual within a complete passage time range, and is suitable for extracting the structure of continuous behavior. The collective behavior channel constructs behavior path combinations based on the interaction relationships between individuals, used to capture the collaborative mechanisms between individuals in a group, and is suitable for group-level behavior pattern recognition and social relationship mining. The unit of time is set according to the temporal stability of human action changes, with each 3 seconds constituting a recognition unit cycle. In a video recognition environment with a frame rate of 25fps, this corresponds to a 75-frame image sequence, containing the completion process of a single basic action. The subgraph sets of the three types of behavior channels are filtered based on attribute co-occurrence weighting functions and temporal correlation functions. The core filtering function for channel construction is defined as: ， Where G represents the original behavioral factor graph structure, For any subgraph in the graph, Subgraph The strength scoring function for attribute co-occurrence and time-series association. The preset filtering threshold, Indicates the channel type.

6. The method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning according to claim 5, characterized in that, The structured behavior type hierarchy in step four refers to the embedding modeling and hierarchy classification of subgraph structures in a multi-channel path graph using graph neural network algorithms, specifically including the following steps: First, for each subgraph in the behavior channel path graph Extract the behavioral factor features of all its nodes and construct a node feature matrix. ,in This indicates the number of nodes in the subgraph. For each node, feature dimensions are defined, including behavior speed, trajectory direction, posture combination, facility interaction, and group interaction features; simultaneously, an adjacency matrix is ​​constructed. This represents the structural connection between any two nodes in the subgraph; secondly, structural enhancement is performed on the adjacency matrix by calculating the normalized adjacency matrix with added self-connections: , in, It is the identity matrix. This represents the adjacency matrix after adding self-connections. for The degree matrix, It is a symmetric normalized adjacency matrix. and These represent the row and column indices of the node in the subgraph, respectively; further, multi-layer graph convolution operations are used to calculate and update the node features, the th... The feature update calculation formula for the layer is: ， in, This represents the layer index of the graph convolutional layer. Indicates the first Layer node embedding matrix, For the first The trainable weight matrix of the layer, It is a non-linear activation function. The aforementioned normalized adjacency matrix; after L layers of convolution operations, the final node embedding is obtained. ; By embedding the final node into the matrix Perform average pooling to obtain the global representation vector of the subgraph. The data is then fed into a fully connected classifier, and the behavior type prediction result is output through the Softmax function. , where C is the number of predefined behavior categories; the result is the structured behavior type spectrum.

7. The method for modeling crowd behavior spectrum based on multimodal recognition and high-precision positioning according to claim 6, characterized in that, Step four involves calculating the feature behavior factor weights for different sample types. This means extracting the behavior subgraph sets for different spatial sample types, statistically analyzing the frequencies of the five types of behavior factor edges, and calculating their relative importance weights using the following normalization formula: ， in, This represents the weight value of factor f in the j-th spatial sample. This represents the cumulative frequency of the edge corresponding to factor f in this type. ; Calculated This refers to the weights of the five behavioral factors for this sample type.

Images