Vehicle track motion feature recognition method based on high-altitude view angle

Abstract

The invention relates to the technical field of intelligent traffic monitoring, and discloses a vehicle track motion feature recognition method based on a high-altitude view angle. The method comprises the following steps: performing noise filtering and coordinate unification processing on original trajectory data acquired at high altitude to form standardized data; and then mapping the data to hierarchical space grids, counting track point density and speed distribution of each grid under different time granularities, and generating multi-time-space-dimension statistical characteristics. Based on features, a significant motion pattern grid is preliminarily identified through pattern matching, and a primary feature value is calculated. And then motion continuity analysis is carried out, and iterative optimization is carried out on the primary characteristic value to generate an optimization value. And finally, integrating all optimized values to construct a regional motion characteristic spectrum, carrying out characteristic propagation and consistency verification based on spectrum topology, and outputting a final result. According to the method, the problems of inaccuracy and result fragmentation of single-scale analysis are solved, and the recognition accuracy and the space-time continuity are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent traffic monitoring, in particular to a vehicle trajectory motion feature recognition method based on high-altitude perspective. BACKGROUND

[0002] The existing vehicle trajectory analysis technology based on high-altitude perspective usually divides the monitoring area with fixed-size grids. This method has limitations when dealing with complex and variable traffic scenes. The fixed grid size is difficult to adapt to the spatial scale difference brought by different road grades and vehicle speeds. For example, larger grids are needed to capture the overall flow direction on highways, while fine grids are needed to distinguish turning behaviors at urban intersections. A single scale grid cannot simultaneously consider macro motion trends and micro motion details, resulting in rough motion feature granularity that cannot accurately reflect the multi-scale motion patterns of vehicles in space.

[0003] In the time dimension, conventional technologies rely on a single time window for statistics, such as aggregating trajectory data in fixed minutes or hours. Traffic flow itself is time-varying, with significant differences in motion patterns between peak and off-peak periods. Single time granularity analysis cannot capture this dynamic change, which may misjudge the instantaneous congestion or smooth state as a long-term regularity, making the time representativeness of the motion feature insufficient and affecting the understanding of the traffic state evolution process.

[0004] After generating preliminary motion features, existing methods often directly make judgments based on the results of isolated grid cells, lacking consideration of the continuity of trajectories in the time-space dimension. Vehicle motion is essentially continuous, and the current position and speed are closely related to the previous state. Ignoring this continuity will result in fragmented motion features, with feature values in adjacent grid cells possibly jumping or being inconsistent, making it difficult to form a smooth and coherent regional-level motion map, reducing the credibility and practicality of the recognition results. SUMMARY

[0005] The purpose of the present application is to provide a vehicle trajectory motion feature recognition method based on high-altitude perspective to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides a vehicle trajectory motion feature recognition method based on high-altitude perspective, which comprises: Obtain the original trajectory sequence covering the target area from the high-altitude monitoring platform, and perform noise filtering and coordinate unification processing on the original trajectory sequence to form standardized trajectory data; Map the processed standardized trajectory data to a spatial division grid with a hierarchical structure, and statistically analyze the trajectory point density and velocity distribution of each grid cell under different time granularities to generate multi-time-space dimensional trajectory statistical features; Based on the multi-dimensional trajectory statistical features, a pattern matching rule is used to preliminarily identify a set of grid cells with significant motion patterns, and a primary motion feature value of each grid cell in the set is calculated; The set of preliminarily identified grid cells is subjected to motion continuity analysis, and the primary motion feature value is iteratively optimized according to the analysis result, to generate an optimized motion feature value; The optimized motion feature values of all grid cells are integrated to construct a regional motion feature map, and feature propagation and consistency checking are performed based on the topological structure of the map, to output a final vehicle motion feature recognition result.

[0007] Preferably, the construction of the space division grid and data mapping comprises the following steps: According to the main road intersections and traffic bottleneck positions in the target area, a plurality of grid division reference points are set; Taking each level of reference point as the center, a plurality of grid cells with nested relationship are generated by outward expansion according to a preset edge length increasing rule; Each trajectory point in the standardized trajectory data is assigned to the finest grid cell at the bottom layer according to its latitude and longitude coordinates; The trajectory data of the bottom layer grid cells is summarized level by level to form a spatial data index structure supporting multi-resolution analysis.

[0008] Preferably, the generation process of the multi-dimensional trajectory statistical features comprises: Three analysis time scales of short time window, medium time window and long time window are defined for each grid cell; At each analysis time scale, the average speed, speed variance, direction angle concentration and number of trajectory points of all trajectory points in the corresponding grid cell are calculated respectively; The calculation results of the same grid cell at different analysis time scales are compared longitudinally to extract the speed stability index and the direction consistency index; The calculation results of all time scales and the comparison indexes are integrated to form a multi-dimensional trajectory statistical feature vector describing the motion characteristics of the grid cell.

[0009] Preferably, the specific implementation of the preliminary identification of the set of grid cells with significant motion patterns by the pattern matching rule comprises: A reference threshold range is set for each dimension in the multi-dimensional trajectory statistical feature vector; The numerical value of each dimension of the feature vector of each grid cell is matched with the corresponding reference threshold range, and the number of dimensions exceeding the threshold range is counted; If the number of dimensions exceeding the threshold range reaches a set number, it is determined that the grid cell has a significant motion pattern, and is included in a set of grid cells with significant motion patterns; For each grid cell included in the set, a primary motion feature value of the grid cell is calculated by weighting the deviation of the number of dimensions exceeding the threshold range from the corresponding threshold.

[0010] Preferably, the specific step of motion continuity analysis includes: For each target grid cell in the set of grid cells with significant motion patterns, retrieve the primary motion feature values of all grid cells in the spatial neighborhood of the target grid cell over consecutive time slices; Analyze the variation trend of the primary motion feature values of the target grid cell and its neighborhood grid cells over the time sequence, and calculate a spatial propagation coefficient and a temporal decay coefficient of motion intensity; Determine whether the motion pattern of the target grid cell has continuity in the space-time dimension according to the spatial propagation coefficient and the temporal decay coefficient; Mark the motion pattern without continuity as an isolated abnormal pattern; The specific step of calculating the spatial propagation coefficient of motion intensity includes: Define the spatial propagation coefficient as the degree of influence of the variation of the motion feature value of the target grid cell on the neighborhood grid cells; For each neighborhood grid cell, calculate the lag time of the variation of its motion feature value relative to the variation of the motion feature value of the target grid cell; According to the lag time and the difference in motion feature values, calculate the propagation rate, i.e. the propagation distance of the variation of the motion feature value per unit time; Integrate the propagation rates of all neighborhood grid cells, and obtain the spatial propagation coefficient by weighted averaging, with the weight being set based on the distance between the neighborhood grid cell and the target grid cell; The specific step of calculating the temporal decay coefficient of motion intensity includes: Define the temporal decay coefficient as the rate of attenuation of the motion feature value over time; Extract the sequence of motion feature values of the target grid cell over consecutive time slices, and fit the curve of the variation of the motion feature value over time; Calculate the first derivative of the curve to obtain the instantaneous change rate of the motion feature value; Calculate the decay coefficient based on the sequence of change rates, fit the motion feature value attenuation process using an exponential decay model, and solve the decay constant; Determine the temporal decay coefficient according to the decay constant and the goodness of fit.

[0011] Preferably, the specific step of analyzing the variation trend of the primary motion feature values of the target grid cell and its neighborhood grid cells over the time sequence includes: obtaining a sequence of primary motion characteristic values of the target grid cell in consecutive time slices, and obtaining a sequence of primary motion characteristic values of all grid cells in the spatial neighborhood of the target grid cell in the same time slices; For each time slice, calculating the difference between the primary motion characteristic value of the target grid cell and each neighboring grid cell, and calculating the average and variance of the difference; Based on the sequence of differences, calculating the correlation coefficient of the motion characteristic values of the target grid cell and the neighboring grid cells to measure the consistency of the motion trend; According to the correlation coefficient and the variance of the difference, judging the stability and propagation direction of the motion trend.

[0012] Preferably, the process of iteratively optimizing the primary motion characteristic value includes: For the target grid cell not marked as an isolated abnormal pattern, enhancing the primary motion characteristic value according to the spatial propagation coefficient and the temporal decay coefficient of the motion intensity; For the target grid cell marked as an isolated abnormal pattern, suppressing the primary motion characteristic value according to the difference between the motion pattern of the target grid cell and the motion pattern of the neighboring grid cells; Taking the corrected characteristic value as the new primary motion characteristic value, repeating the motion continuity analysis and correction operation until the variation of the characteristic value generated by two consecutive iterations is less than the convergence threshold, and the obtained characteristic value is determined as the optimized motion characteristic value.

[0013] Preferably, the method for constructing the regional motion characteristic map includes: Abstracting each grid cell as a node in the map, and taking the optimized motion characteristic value as the core attribute of the node; If two grid cells are adjacent in space and their motion patterns are recognized in consecutive time slices, a edge is established between the corresponding nodes, and the edge is given a weight according to the spatial propagation coefficient of the motion intensity; Combining all nodes and weighted edges to form a regional motion characteristic map with grid cells as nodes and spatial correlation as edges.

[0014] Preferably, the feature propagation based on the topology of the map includes: Randomly selecting a node in the map as a starting node, and taking the optimized motion characteristic value of the starting node as an initial propagation value; According to the weight of the edge, the initial propagation value is proportionally distributed to all neighbor nodes directly connected to the starting node; Each neighbor node receives the propagation values from all connected nodes and accumulates them to obtain the cumulative propagation value of the neighbor node; Fusing the cumulative propagation value with the optimized motion characteristic value of the node itself to update the motion characteristic value of the node; The multi-round propagation and updating is completed for all nodes of the graph until the motion characteristic values of all nodes in the graph tend to be stable.

[0015] Preferably, the specific execution steps of the consistency check include: After the feature propagation is completed, the difference between the updated motion characteristic value of each node and the average of the updated motion characteristic values of all direct neighbor nodes of the node is calculated; If the difference exceeds the allowable deviation, the motion characteristic value of the node is adjusted smoothly to approach the average of the neighbor nodes; After the smooth adjustment of all nodes is completed, the difference is calculated again, and the smooth adjustment process is repeated until the difference of all nodes is lower than the allowable deviation, and the finally adjusted motion characteristic value is taken as the final vehicle motion characteristic identification result of the node.

[0016] Compared with the prior art, the present application has the following beneficial effects: By mapping the trajectory data to the hierarchical spatial grid and performing statistics at different time granularities, the multi-scale characteristics of vehicle motion are captured. The hierarchical grid can express both the macro motion trend of a wide area and the fine motion mode of a local road network, avoiding the problem of information loss or confusion under a fixed grid size. Combined with multi-time granularity analysis, short-term fluctuations and long-term regularities of traffic flow can be distinguished, so that the extracted trajectory statistical features can reflect real-time dynamics and periodic patterns. This multi-time and space dimensional feature representation provides a richer and more accurate data basis for subsequent pattern recognition, improving the resolution of preliminary motion pattern recognition.

[0017] Based on the iterative optimization of the primary motion characteristic value based on motion continuity analysis, the fragmentation and noise interference problems of the preliminary identification result are effectively solved. Motion continuity analysis can identify and correct abnormal characteristic values that do not fit the surrounding motion trend by checking the correlation of adjacent grid units in space and time. The iterative optimization process ensures smooth transition of the characteristic value in the space-time domain, so that the finally generated regional motion characteristic graph has strong internal consistency. The feature propagation based on the graph topology further utilizes the spatial correlation to transfer reliable feature information to areas with weak evidence, filling in the data gaps and enhancing the robustness of the entire system to local data loss or noise, outputting a coherent and consistent vehicle motion characteristic identification result. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A working principle diagram of the vehicle trajectory motion feature identification method based on high-altitude perspective according to the present application; Figure 2 A flowchart of spatial grid division and data mapping; Figure 3 Flow chart for preliminary identification of significant motion patterns by pattern matching rules; Figure 4 Analysis graph for optimizing the association of motion feature values and edge weights for source nodes; Figure 5 Heat map for motion feature distribution of different regional grid cells. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0020] Please refer to Figure 1 The present application provides a vehicle trajectory motion feature recognition method based on high-altitude perspective, which comprises the following steps: obtaining original trajectory sequences covering a target area from a high-altitude monitoring platform, the high-altitude monitoring platform comprising a satellite or a drone system, which can collect latitude and longitude coordinates and timestamp data generated by vehicle movement; performing noise filtering and coordinate unification processing on the original trajectory sequences, the noise filtering using a filtering algorithm to remove abnormal drift points, and the coordinate unification processing converting trajectory data of different sources to the same geographic coordinate system to form standardized trajectory data; mapping the processed standardized trajectory data to a spatial division grid with a hierarchical structure, the grid being divided into multiple levels based on the geographic features of the target area, and the trajectory point density and speed distribution of each grid cell being counted at different time granularities, the time granularities being divided into short-time, medium-time and long-time scales, and the trajectory statistical features of multiple time and space dimensions being generated by calculating average speed, speed variance and other indicators; based on the trajectory statistical features of multiple time and space dimensions, a set of grid cells with significant motion patterns is preliminarily identified by using pattern matching rules, the pattern matching rules setting a reference threshold range of feature dimensions, comparing the feature values of the grid cells to determine the significance, and calculating the primary motion feature values of each grid cell; performing motion continuity analysis on the set of grid cells preliminarily identified, the analysis involving retrieving the feature value changes of the neighbor grid cells in the time sequence, calculating the spatial propagation coefficient and the time attenuation coefficient, iteratively optimizing the primary motion feature values according to the analysis results, and generating optimized motion feature values; integrating the optimized motion feature values of all grid cells to construct a regional motion feature map, the map abstracting the grid cells as nodes and the spatial association as edges, and performing feature propagation and consistency checking based on the topological structure of the map, the feature propagation being realized by value transmission between nodes to achieve smoothing, and the consistency checking being adjusted by comparing the differences between the values of the nodes and the neighbors, and finally outputting the vehicle motion feature recognition result, which can be used for traffic monitoring or behavior analysis applications.

[0021] Embodiment 1: refer to Figure 2 In specific implementation, the construction of the spatial division grid and the data mapping process set reference points for multi-level grid division according to major road intersections and traffic bottleneck locations within the target area. These reference points are automatically determined by geographic information system analysis of road network density, ensuring that the spatial division grid cells cover critical traffic flow areas. With each reference point as the center, the multi-level spatial division grid cells are generated by outward expansion following a pre-set edge length increment rule. The edge length increment rule adopts geometric progression, for example, the edge length of the bottom-level spatial division grid cell is 10 meters, which is gradually expanded to 100 meters or higher, forming a hierarchical structure from fine to coarse. Each trajectory point in the standardized trajectory data is assigned to the finest bottom-level spatial division grid cell according to its latitude and longitude coordinates. The assignment process uses spatial query algorithms such as point positioning retrieval to ensure that each trajectory point is accurately mapped to the corresponding spatial division grid cell. The trajectory data of the bottom-level spatial division grid cells is summarized level by level to form a spatial data index structure that supports multi-resolution analysis. The summarization operation is implemented through aggregation functions such as cumulative trajectory point count or average speed calculation. The spatial data index structure uses a quadtree or similar hierarchical model for efficient access to multi-scale data. In some embodiments, the reference point selection for multi-level grid division can also dynamically adjust the reference point positions based on historical traffic flow data to optimize the distribution of spatial division grid cells. The reference point setting process introduces machine learning algorithms to automatically identify key nodes in the road network, ensuring that the spatial division grid cells can effectively capture hotspots of vehicle movement. The specific parameters of the edge length increment rule can be customized according to the size and shape of the target area, for example, smaller edge length increments are used in urban center areas and larger edge length increments are used in suburban areas to maintain the adaptability of the spatial division grid cells. Optionally, when assigning trajectory points to spatial division grid cells, efficient spatial partitioning algorithms such as grid indexing or R-tree structure can be used to speed up the mapping process. The assignment algorithm ensures that each trajectory point belongs to only one spatial division grid cell, avoiding duplicate counting. When summarizing the bottom-level data, the aggregation functions include summation, average calculation, or statistical distribution analysis to form a multi-level spatial data index structure that supports fast querying and updating. It can be understood that the construction process of the spatial division grid aims to create a hierarchical data structure to facilitate subsequent multi-scale analysis. The design of the spatial data index structure considers computational efficiency and storage optimization to ensure the feasibility of large-scale trajectory data processing.

[0022] The generation process of the multi-time-space dimension trajectory statistical feature divides each space into grid units to define three analysis time scales of short time window, medium time window and long time window. The short time window can be minute level, the medium time window is hour level, and the long time window is day level, so as to capture the movement mode of different time granularity. In each analysis time scale, the average speed, speed variance, direction angle concentration and the number of trajectory points in the grid unit are calculated. The direction angle concentration is quantified by calculating the variance or entropy value of the direction angle. The calculation results of the same grid unit in different analysis time scales are compared longitudinally to extract the speed stability index and the direction consistency index. The speed stability index is calculated based on the cross-time scale change of the speed variance, and the direction consistency index is obtained by comparing the difference of the direction angle concentration. The calculation results of all time scales and the comparison indexes are combined to form a multi-time-space dimension trajectory statistical feature vector describing the movement characteristics of the grid unit. The multi-time-space dimension trajectory statistical feature vector is used as the basic data for subsequent pattern matching. In some embodiments, the specific definition of the analysis time scale can be adjusted according to the application scenario. For example, the short time window is set to 5 minutes, the medium time window is set to 1 hour, and the long time window is set to 24 hours to adapt to different traffic monitoring needs. The calculation of the direction angle concentration uses the standard deviation or circular statistics of the direction angle to ensure accurate measurement of the direction consistency. The longitudinal comparison process introduces time series analysis methods such as sliding window calculation to dynamically evaluate the trend of the speed stability index and the direction consistency index. Optionally, the number of trajectory points can include a de-duplication process to avoid bias caused by counting the same vehicle multiple times. The calculation of the average speed and the speed variance is based on the time interval weighting of the trajectory points to improve the accuracy of the index. The formation process of the multi-time-space dimension trajectory statistical feature vector can also include a normalization step to scale the values of different dimensions to the same range for subsequent comparison. It can be understood that the generation of the multi-time-space dimension trajectory statistical feature aims to extract the space-time movement mode of the grid unit. The speed stability index reflects the volatility of vehicle movement, and the direction consistency index reveals the concentration degree of the movement direction. The multi-time-space dimension trajectory statistical feature vector integrates multiple indexes to provide a comprehensive description of the movement characteristics.

[0023] In specific implementation, the data mapping process of the spatial partition grid cell ensures the accurate correspondence of trajectory points and the grid, the mapping algorithm handles coordinate conversion and boundary conditions, such as trajectory points located at the edge of the spatial partition grid cell, the nearest neighbor principle is adopted for assignment, the construction of the spatial data index structure supports efficient range query and aggregation operation, the index update mechanism allows real-time processing of new inflow trajectory data. The calculation process of multi-spatiotemporal dimension trajectory statistical features realizes parallel processing, each spatial partition grid cell independently calculates the feature vector, reduces the calculation time, the feature vector storage uses a compact data structure, optimizes the memory usage. In specific implementation, the definition of short time window, medium time window and long time window is based on timestamp alignment, ensuring the comparability of data at different time scales, the calculation of trajectory point direction angle is based on the vector difference of consecutive trajectory points, the direction angle concentration uses ring variance measurement, improving the robustness of direction consistency. The speed stability index is obtained by calculating the ratio of speed variance at different time scales, the direction consistency index is based on the overlap degree evaluation of direction angle distribution, the longitudinal comparison process uses difference analysis or correlation calculation, revealing the time evolution of movement pattern. In specific implementation, the dimensions of the multi-spatiotemporal dimension trajectory statistical feature vector include numerical and categorical indicators, the vector normalization method uses minimum-maximum scaling or Z-score standardization, ensuring the consistency of feature scale. In specific implementation, the hierarchical structure of the spatial partition grid cell supports top-down or bottom-up analysis path, multi-resolution analysis allows users to select appropriate granularity as needed, the query interface of the spatial data index structure provides flexible data access method. In specific implementation, the trajectory point density statistics considers the time decay factor, recent data is given higher weight, the speed distribution calculation introduces kernel density estimation, smoothing the distribution curve, improving the representativeness of statistical features.

[0024] Example 2: see Figure 3In a specific implementation, the specific implementation of the preliminary identification of the set of grid cells with significant motion patterns using the pattern matching rule includes setting a reference threshold range for each dimension of the multi-dimensional trajectory statistical feature vector, the reference threshold range being determined based on historical data or statistical distribution, for example, the lower limit of the reference threshold range of the average speed dimension is set to 5 km / h, and the upper limit is set to 80 km / h, and matching each dimension value of the multi-dimensional trajectory statistical feature vector of each grid cell with the corresponding reference threshold range, the matching process checks whether each dimension value is higher than the upper limit of the reference threshold range or lower than the lower limit of the reference threshold range, counts the number of dimensions that exceed the reference threshold range, and if the number of dimensions that exceed the reference threshold range reaches a set number, for example, the set number is half of the total number of dimensions of the multi-dimensional trajectory statistical feature vector, it is determined that the grid cell has a significant motion pattern, and it is included in the set of grid cells with significant motion patterns. For each grid cell in the set of grid cells with significant motion patterns, the deviation degree of the dimension value that exceeds the reference threshold range from the corresponding reference threshold range is weighted to calculate the primary motion feature value of the grid cell, and the weighting coefficient is allocated based on the dimension importance. The greater the deviation, the higher the weight. In some embodiments, the setting of the reference threshold range can use a dynamic adjustment method to update the upper and lower limits of the reference threshold range according to the real-time traffic flow state, and the set number can also be set differently according to the grid level, and the set number requirement of high-level grid cells can be appropriately reduced. Optionally, the calculation of the deviation degree can use absolute value difference or relative proportion difference, and the deviation of different dimensions can be normalized when calculating the primary motion feature value to ensure the comparability of the primary motion feature value. It can be understood that the core of the pattern matching rule is to filter the abnormal motion pattern through multi-dimensional threshold comparison, and the primary motion feature value quantifies the significance of the motion pattern of the grid cell.

[0025] The specific steps of the motion continuity analysis include, for each target grid cell in the set of grid cells of a significant motion pattern, retrieving the primary motion eigenvalues of all grid cells within its spatial neighborhood, which is defined as the eight neighboring grid cells around the target grid cell, on consecutive time slices, analyzing the trend of the primary motion eigenvalues of the target grid cell and its neighboring grid cells over the time sequence, calculating the spatial propagation coefficient and the temporal decay coefficient of the motion intensity, determining whether the motion pattern of the target grid cell has continuity in the spatiotemporal dimension based on the spatial propagation coefficient and the temporal decay coefficient of the motion intensity, and marking the motion pattern without continuity as an isolated abnormal pattern. The specific steps of calculating the spatial propagation coefficient of the motion intensity include defining the spatial propagation coefficient as the degree of influence of the motion eigenvalue change of the target grid cell on the neighboring grid cells, for each neighboring grid cell, calculating the lag time of the motion eigenvalue change of the neighboring grid cell relative to the motion eigenvalue change of the target grid cell, calculating the propagation rate based on the lag time and the difference in motion eigenvalues, defining the propagation rate as the propagation distance of the motion eigenvalue change per unit time, and obtaining the spatial propagation coefficient by weighted averaging of the propagation rates of all neighboring grid cells, with the weights being set based on the distances of the neighboring grid cells from the target grid cell. The specific steps of calculating the temporal decay coefficient of the motion intensity include defining the temporal decay coefficient as the rate of attenuation of the motion eigenvalue over time, extracting the sequence of motion eigenvalues of the target grid cell on consecutive time slices, fitting the curve of the motion eigenvalue change over time, calculating the instantaneous change rate of the motion eigenvalue by taking the first derivative of the curve, calculating the decay coefficient based on the sequence of change rates, fitting the attenuation process of the motion eigenvalue using an exponential decay model and solving the decay constant, and determining the temporal decay coefficient based on the decay constant and the goodness of fit. In some embodiments, the definition of the spatial neighborhood can be extended to more distant grid cells, for example, considering second-order neighborhoods, the calculation of the lag time can use cross-correlation analysis methods, and the calculation of the propagation rate can consider the amplitude and persistence of the motion eigenvalue change. Optionally, the form of the exponential decay model can be expressed as: wherein: represents the motion eigenvalue at time , represents the initial motion eigenvalue, represents the decay constant to be solved, and the goodness of fit is evaluated by calculating the coefficient of determination . The decay constant is the temporal decay coefficient.

[0026] In a specific implementation, the specific steps of analyzing the change trend of the primary motion characteristic value of the target grid cell and its neighborhood grid cells in the time sequence include obtaining the primary motion characteristic value sequence of the target grid cell in consecutive time slices, while obtaining the primary motion characteristic value sequence of all grid cells in the spatial neighborhood in the same time slices, calculating the difference value of the primary motion characteristic value of the target grid cell and each neighborhood grid cell for each time slice, and calculating the average value and variance of the difference value, calculating the motion characteristic value correlation coefficient of the target grid cell and the neighborhood grid cell based on the difference value sequence, measuring the consistency of the motion trend, and judging the stability and propagation direction of the motion trend according to the correlation coefficient and the difference value variance. The average value of the primary motion characteristic value difference reflects the overall deviation level of the motion characteristics of the target grid cell and the neighborhood grid cell, the variance of the primary motion characteristic value difference reflects the fluctuation degree of the deviation, the motion characteristic value correlation coefficient is calculated by using the Pearson product-moment correlation coefficient, and the consistency of the motion trend is judged by the correlation coefficient being higher than a threshold value and the difference value variance being lower than a threshold value. The propagation direction of the motion trend is inferred by analyzing the relative size of the lag time of different neighborhood grid cells, and the direction of the neighborhood grid cell with short lag time is regarded as the main direction of the motion propagation. In the calculation process of the spatial propagation coefficient of the motion intensity, the distance weight used in the weighted average can adopt the inverse distance weight method, that is, the weight is inversely proportional to the spatial distance between the neighborhood grid cell and the target grid cell.

[0027] In a specific implementation, the specific steps of analyzing the change trend of the primary motion feature value of the target grid cell and its neighborhood grid cells over a time sequence include obtaining the primary motion feature value sequence of the target grid cell over consecutive time slices, while obtaining the primary motion feature value sequence of all grid cells in the spatial neighborhood over the same time slices, the time slice interval is consistent with the analysis time scale, for example, the time slice can be set to one slice per minute, and the consecutive time slices cover a complete analysis window such as 30 minutes, the difference value of the primary motion feature value of the target grid cell and each neighborhood grid cell is calculated for each time slice, and the average value and variance of the difference value are calculated, the average value of the difference value reflects the overall deviation level of the motion feature of the target grid cell and the neighborhood grid cell, and the variance of the difference value indicates the fluctuation degree of the deviation, the correlation coefficient of the motion feature value of the target grid cell and the neighborhood grid cell is calculated based on the difference value sequence, which measures the consistency of the motion trend, the correlation coefficient is calculated by using the Pearson product-moment correlation coefficient, and the stability and propagation direction of the motion trend are judged according to the correlation coefficient and the variance of the difference value, the stability is high, which is manifested as the correlation coefficient is close to 1 and the variance of the difference value is lower than a preset threshold, and the propagation direction is inferred by analyzing the distribution of the correlation coefficient in different neighborhood directions. The time slice setting can be adjusted according to the actual traffic flow density, for example, a finer time granularity is used during the peak period, the difference value statistics can use the exponential weighted moving average method to emphasize recent data, and the calculation of the correlation coefficient can introduce time lag analysis to capture the delay effect of motion propagation. Optionally, the stability judgment threshold of the motion trend can be learned from historical data, and the propagation direction inference can be combined with the spatial topological relationship of the grid cell to improve the accuracy.

[0028] The process of iteratively optimizing the primary motion feature value includes, for the target grid cell not marked as an isolated abnormal mode, enhancing and correcting the primary motion feature value according to the spatial propagation coefficient and the time decay coefficient of the motion intensity, the enhancement and correction is realized by applying a correction multiplier, the correction multiplier is calculated based on the spatial propagation coefficient and the time decay coefficient, and the formula is: Wherein: represents the corrected primary motion feature value, represents the primary motion feature value before correction, represents the spatial propagation coefficient of the motion intensity, represents the time decay coefficient of the motion intensity, and is a positive weight coefficient used to balance the influence of spatial propagation and temporal decay, for the target grid cell labeled as isolated anomaly mode, the primary motion eigenvalue is suppressed and corrected according to the difference degree of motion pattern with the surrounding grid cells, the suppression and correction adopts a neighborhood weighted average method, for example, the primary motion eigenvalue is adjusted to the neighborhood average value to reduce the influence of abnormal value, the corrected eigenvalue is taken as the new primary motion eigenvalue, the motion continuity analysis and correction operation are repeated until the eigenvalue variation generated by two consecutive iterations is less than a convergence threshold, the convergence threshold is set to a small positive number such as 0.001, at this time the obtained eigenvalue is determined as the optimized motion eigenvalue.

[0029] In specific implementation, the construction method of the regional motion feature graph includes abstracting each grid cell as a node in the graph, taking the optimized motion eigenvalue as the core attribute of the node, the node attribute can also include the spatial coordinate identifier of the grid cell for subsequent spatial query, if two grid cells are adjacent in space and are identified as motion patterns in consecutive time slices, a edge is established between the corresponding nodes, the definition of spatial adjacency is based on the adjacency relationship of the grid cells, for example, whether the grid cells share a boundary or a vertex is determined by using the four-connected or eight-connected rule, the determination of consecutive time slices requires that the two grid cells have non-zero optimized motion eigenvalues in a preset number of consecutive time slices, and the edge is assigned a weight according to the spatial propagation coefficient of the motion intensity, the weight assignment follows the principle that the higher the spatial propagation coefficient of the motion intensity, the greater the edge weight, all nodes and weighted edges are combined to form a regional motion feature graph with grid cells as nodes and spatial correlation as edges, the data structure of the graph can be stored in an adjacency list or an adjacency matrix to realize efficient traversal. In some embodiments, the core attribute of the node can be extended to include timestamp information to record the latest update time of the motion pattern, the establishment condition of the edge can increase the consistency constraint of the motion direction, which requires that the motion direction of the connected grid cells has an included angle less than a threshold to ensure the continuity of the motion pattern, and the weight assignment can use a linear mapping function to convert the spatial propagation coefficient of the motion intensity to a fixed interval. Optionally, the judgment of spatial adjacency relationship can introduce a distance threshold to allow the establishment of edges between grid cells that are not directly adjacent but close in space, the preset number of consecutive time slices can be adjusted according to the motion persistence requirement of the actual application scene, and the graph combination process can filter redundant edges to remove edges with a weight lower than a minimum threshold to simplify the graph structure. It can be understood that the construction of the regional motion feature graph converts the discrete grid cell motion information into a network representation, and the spatial propagation path of the motion pattern is revealed through the relationship between nodes and edges.

[0030] In specific implementation, the weight assignment process of the edge can use a normalization method to ensure that the weight values are within a consistent range, and the weight calculation formula can be expressed as: wherein: denotes the edge weight connecting node i and node j, denotes the spatial propagation coefficient of motion intensity between node i and node j, and are the minimum and maximum values of the spatial propagation coefficient of motion intensity for all edge pairs, respectively, after the edge weight assignment is completed, the topology of the graph can be analyzed by graph traversal algorithms, for example, using breadth-first search to identify connected components to discover large-scale motion pattern regions.

[0031] Table 1: Data table of nodes and edges of regional motion feature graph

[0032] In a specific implementation, referring to Table 1, Table 1 shows the data of some nodes and edges in the regional motion feature graph, the node identifier uses a unique code to represent each grid cell, the grid cell center coordinates use the latitude and longitude format to describe the spatial position of the grid cell, the optimized motion feature value comes from the result of the previous iteration optimization process, the adjacent node identifier lists the node codes connected to the current node by edges, the edge weight is calculated based on the spatial propagation coefficient of motion intensity, and the value range is [0, 1], the data in Table 1 reflects the propagation path of the motion feature along the direction of grid cells N001 to N004, and the change of the edge weight embodies the attenuation trend of the motion intensity propagation. In a specific implementation, such table data can be stored in a relational database or a graph database for subsequent feature propagation module query, and the connection relationship in the table can be used to verify the correctness of the graph construction, for example, checking the consistency of the edge weight and the spatial propagation coefficient of motion intensity.

[0033] Referring to Figure 4 , the correlation between the source node optimized motion feature value and the edge weight (based on the spatial propagation coefficient), the target node feature value difference is presented, and the trend line of the two is fitted ( ). In the figure, the horizontal axis is the edge weight (reflecting the spatial propagation coefficient of the motion intensity between grid cells), the vertical axis is the difference value of the target node eigenvalue, and the color scale of the node color (from dark blue to yellow) corresponds to the size of the source node optimized motion eigenvalue (the value range is about 0.55-0.90). From the trend of the figure, the edge weight and the difference value of the target node eigenvalue present a weak negative correlation: as the edge weight (spatial propagation coefficient) increases, the difference value of the eigenvalue between the target node and the source node presents a slight downward trend as a whole, which is consistent with the logic that "the stronger the spatial correlation (the higher the edge weight), the smaller the feature difference between nodes" in the process of motion feature propagation. At the same time, the distribution of nodes of different colors shows that the target node eigenvalue corresponding to the source node optimized motion eigenvalue with a higher value (such as yellow nodes) is at a medium level, while the nodes with a lower eigenvalue (such as dark blue) are distributed in a smaller difference value area, reflecting the intensity attenuation characteristics of motion feature propagation. The figure directly reflects the influence law of the edge weight (spatial propagation coefficient) on the feature difference between nodes in the regional motion feature map, which can provide a visual basis for the subsequent weight configuration of feature propagation.

[0034] In a specific implementation, the feature propagation based on the topological structure of the map includes randomly selecting a node in the regional motion feature map as a starting node, taking the optimized motion eigenvalue of the starting node as an initial propagation value, distributing the initial propagation value to all neighbor nodes directly connected to the starting node according to the edge weight, and the proportion calculation is based on the normalization processing of the edge weight, specifically, the proportion allocated to each neighbor node is equal to the proportion of the edge weight connecting the neighbor node in the total sum of the out-edge weight of the starting node, each neighbor node receives the propagation value from all connected nodes and accumulates to obtain the cumulative propagation value of the neighbor node, and the cumulative propagation value and the optimized motion eigenvalue of the neighbor node are fused to update the motion eigenvalue of the neighbor node, and the fusion process adopts a weighted average method, and the formula is: wherein: denotes the updated motion eigenvalue of the neighbor node, denotes the cumulative propagation value received by the neighbor node from all connected nodes, denotes the original optimized motion eigenvalue of the neighbor node, is a propagation weight coefficient for controlling the influence degree of the cumulative propagation value on the update result, and all nodes in the regional motion feature map are traversed to complete multiple rounds of propagation and update until the motion eigenvalues of all nodes in the regional motion feature map tend to be stable, and the stability condition is defined as the average absolute value of the change of the motion eigenvalue of all nodes in the continuous two rounds of iteration being less than a preset threshold. The selection of the starting node can use a polling strategy instead of random selection to ensure the uniformity of the propagation coverage, the normalization processing of the edge weight can use a soft maximum value function to enhance the significance of the weight difference, and the propagation weight coefficient The propagation weight coefficient of a high-degree node can be appropriately reduced to avoid over-smoothing according to the dynamic adjustment of node degree. Optionally, the calculation of the cumulative propagation value can introduce a decay factor that reduces the propagation strength as the propagation hop number increases. The setting of the stability threshold can consider the size of the regional motion feature map, and a larger-scale map can use a more stringent threshold requirement.

[0035] The specific execution steps of the consistency check include calculating the difference between the updated motion feature value of each node and the average of the updated motion feature values of all its direct neighbor nodes after feature propagation is completed. The difference calculation uses absolute difference or relative difference measurement. If the difference exceeds the allowable deviation, the motion feature value of the node is adjusted to make the node motion feature value closer to the average of the neighbor node motion feature values. The smoothing adjustment uses an iterative relaxation method, which specifically adjusts the node motion feature value to a weighted combination of the current value and the average of the neighbor values. After completing the smoothing adjustment of all nodes, the difference is calculated again, and the smoothing adjustment process is repeated until the difference of all nodes is below the allowable deviation. The final adjusted motion feature value is taken as the final vehicle motion feature recognition result of the node. The difference calculation can use the root mean square error instead of the absolute difference to improve the robustness to abnormal values. The setting of the allowable deviation can be hierarchical, with different deviation thresholds for different regions to adapt to the local motion feature variation. The smoothing adjustment weight can be adjusted according to the node degree, and the adjustment amplitude of a high-degree node can be appropriately reduced. Optionally, the consistency check process can be alternated with the feature propagation process, with a check performed after each round of propagation to improve convergence efficiency. The output format of the final vehicle motion feature recognition result includes the node identifier and the corresponding motion feature value vector.

[0036] In specific implementation, the traversal order of feature propagation can adopt a breadth-first search strategy, starting from the starting node and expanding layer by layer outward, ensuring uniform advancement of the propagation wave front. The accumulation operation of the accumulated propagation value needs to handle the case where multiple propagation paths reach the same node. The accumulation process performs numerical clipping to prevent overflow. The motion feature value update operation needs to ensure atomicity to avoid data inconsistency caused by concurrent propagation. In specific implementation, the difference degree calculation needs to access the direct neighbor set of the node. The neighbor relationship is obtained by querying the edge table of the regional motion feature graph. In the iterative relaxation method of smoothing adjustment, the weight of weighted combination is usually set to a fixed value such as 0.5, indicating that the current node value and the neighbor average value each account for half of the weight. The typical value of the allowable deviation can be set to 5% of the range of the motion feature value. The round limit of feature propagation can be set to a fixed value such as 100 rounds to prevent infinite loops. The repetition number of consistency check is also set to an upper limit. After exceeding the upper limit, the current result is forced to terminate and output. The storage of the final vehicle motion feature recognition result contains timestamp information, supporting time series analysis. In specific implementation, two versions of node motion feature values need to be maintained during the feature propagation process. The current version is used for propagation calculation, and the next version is used for storage of updated results. After each round of propagation is completed, version switching is performed. The consistency check process needs to record the difference degree history of each node for detecting convergence trend.

[0037] In specific implementation, the processing of large-scale regional motion feature graph can adopt a distributed computing framework. After the graph is divided into blocks, feature propagation and consistency check are performed in parallel on different computing nodes. Nodes at the boundary of sub-graphs need special processing to ensure the correctness of propagation across sub-graph boundaries. The stability of feature propagation can be evaluated by monitoring the change trajectory of node motion feature values. The effectiveness of consistency check can be verified by comparing the distribution uniformity of node motion feature values before and after check. The output interface of the final vehicle motion feature recognition result provides multiple access methods, including spatial range query and feature value threshold filtering, which is convenient for downstream application systems to use.

[0038] Referring to Figure 5In the regional motion feature analysis of stage 3-4, the number distribution of grid cells of different functional areas (main road, secondary road, intersection, residential area, commercial area) under different motion modes (stopping state, deceleration, acceleration, constant speed, turning) is presented in the form of heat map. The color depth in the figure corresponds to the number of grid cells (dark color represents more, light color represents less): for example, the number of grid cells of area B (secondary road) under the "deceleration" mode reaches 14 (the maximum value in the figure), which reflects the high frequency of deceleration behavior of vehicles on secondary roads; the number of grid cells of area D (residential area) under the "acceleration" and "turning" modes is 13, which reflects the concentration characteristics of the motion mode of vehicles in residential areas; the number of grid cells of area E (commercial area) under the "deceleration" mode is only 3, which is significantly different from the secondary road, and is consistent with the traffic characteristics of commercial areas. The quantitative distribution of the figure can provide direct statistical basis for the node attribute assignment of the regional motion feature graph, and support the local parameter adjustment of subsequent feature propagation and consistency check.

[0039] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A method for recognizing vehicle trajectory motion features based on a high-altitude perspective, characterized in that, The method specifically includes the following processing stages: The original trajectory sequence covering the target area is obtained from the high-altitude monitoring platform, and the original trajectory sequence is subjected to noise filtering and coordinate unification processing to form standardized trajectory data; The processed standardized trajectory data is mapped to a hierarchical spatial grid, and the trajectory point density and velocity distribution of each grid cell are statistically analyzed at different time granularities to generate multi-temporal and spatial statistical features of the trajectory. Based on the multi-temporal and spatial trajectory statistical features, a set of grid cells with significant motion patterns is initially identified using pattern matching rules, and the primary motion feature value of each grid cell in the set is calculated. Motion continuity analysis is performed on the initially identified set of grid cells. Based on the analysis results, the initial motion feature values ​​are iteratively optimized to generate optimized motion feature values. The optimized motion feature values ​​of all grid cells are integrated to construct a regional motion feature map. Based on the map topology, feature propagation and consistency verification are performed to output the final vehicle motion feature recognition result.

2. The method for recognizing vehicle trajectory motion features based on a high-altitude perspective according to claim 1, characterized in that, The construction and data mapping of the spatial grid include the following steps: Based on the main road intersections and traffic bottlenecks within the target area, benchmark points for multi-level grid division are set. Centered on each level of reference point, multi-level grid cells with nested relationships are generated by expanding outward according to the preset side length increasing rule; Each trajectory point in the standardized trajectory data is assigned to the lowest-level fine grid cell based on its latitude and longitude coordinates; The trajectory data of the lower-level grid cells are aggregated level by level to form a spatial data index structure that supports multi-resolution analysis.

3. The method for recognizing vehicle trajectory motion features based on a high-altitude perspective according to claim 2, characterized in that, The process of generating the statistical features of the multi-spatiotemporal dimensional trajectory includes: Three analysis time scales—short time window, medium time window, and long time window—are defined for each grid cell. At each analysis time scale, calculate the average velocity, velocity variance, orientation angle concentration, and number of trajectory points for all trajectory points within the corresponding grid cell; By comparing the calculation results of the same grid cell at different analysis time scales longitudinally, the velocity stability index and the directional consistency index are extracted. By combining the calculation results and comparison indicators across all time scales, a multi-temporal and spatial dimension trajectory statistical feature vector describing the motion characteristics of the grid cell is formed.

4. The method for recognizing vehicle trajectory motion features based on a high-altitude perspective according to claim 3, characterized in that, The specific implementation of using pattern matching rules to initially identify a set of grid cells with significant motion patterns includes: Set a reference threshold range for each dimension in the multi-temporal and spatiotemporal trajectory statistical feature vector; Match the feature vector values ​​of each dimension of each grid cell with the corresponding reference threshold range, and count the number of dimensions that exceed the threshold range; If the number of dimensions exceeding the threshold reaches a set number, the grid cell is determined to have a significant motion pattern and is included in the set of grid cells with significant motion patterns. For each grid cell included in the set, the primary motion characteristic value of the grid cell is calculated by weighting the deviation of its dimension values ​​that exceed the threshold range from the corresponding threshold.

5. The method for recognizing vehicle trajectory motion features based on a high-altitude perspective according to claim 4, characterized in that, The specific steps of the motion continuity analysis include: For each target grid cell in the set of grid cells with significant motion patterns, retrieve the primary motion feature values ​​of all grid cells in its spatial neighborhood on consecutive time slices; Analyze the changing trends of the primary motion eigenvalues ​​of the target grid cell and its neighboring grid cells in the time series, and calculate the spatial propagation coefficient and time decay coefficient of the motion intensity. Based on the spatial propagation coefficient and the time decay coefficient, determine whether the motion pattern of the target grid cell is continuous in the spatiotemporal dimension; Motion patterns that do not have continuity are marked as isolated anomalous patterns; The specific steps for calculating the spatial propagation coefficient of motion intensity include: The spatial propagation coefficient is defined as the degree of influence of changes in the motion characteristic value of the target grid cell on neighboring grid cells. For each neighboring grid cell, calculate the lag time of its motion characteristic value change relative to the motion characteristic value change of the target grid cell; The propagation rate is calculated based on the difference between the lag time and the motion characteristic value, which is the propagation distance of the change in motion characteristic value per unit time. The spatial propagation coefficient is obtained by combining the propagation rates of all neighboring grid cells and averaging them. The weights are set based on the distance between the neighboring grid cells and the target grid cell. The specific steps for calculating the time decay coefficient of exercise intensity include: The time decay coefficient is defined as the rate at which the motion characteristic value decreases over time; Extract the motion feature value sequence of the target mesh cell in continuous time slices, and fit the curve of motion feature value change with time; Calculate the first derivative of the curve to obtain the instantaneous rate of change of the motion characteristic values; The decay coefficient is calculated based on the rate of change sequence, the decay process of motion characteristic values ​​is fitted using an exponential decay model, and the decay constant is solved. The time decay coefficient is determined based on the decay constant and the goodness of fit.

6. The vehicle trajectory motion feature recognition method based on a high-altitude perspective as described in claim 5, characterized in that, The specific steps for analyzing the change trend of the primary motion characteristic values ​​of the target grid cell and its neighboring grid cells in the time series include: Obtain the primary motion feature value sequence of the target grid cell in consecutive time slices, and at the same time obtain the primary motion feature value sequence of all grid cells in its spatial neighborhood in the same time slice; For each time slice, calculate the difference in primary motion eigenvalues ​​between the target grid cell and each neighboring grid cell, and calculate the mean and variance of the difference. The correlation coefficient of motion characteristic values ​​between the target grid cell and its neighboring grid cells is calculated based on the difference sequence to measure the consistency of motion trends; Based on the correlation coefficient and the variance of the difference, the stability and propagation direction of the movement trend can be determined.

7. The method for recognizing vehicle trajectory motion features based on a high-altitude perspective according to claim 6, characterized in that, The iterative optimization process of the primary motion feature values ​​includes: For target mesh cells not marked as isolated anomaly modes, the primary motion characteristic values ​​are enhanced and corrected based on the spatial propagation coefficient and time decay coefficient of their motion intensity. For target mesh cells marked as isolated anomaly patterns, the primary motion feature values ​​are suppressed and corrected according to the degree of difference between their motion patterns and those of surrounding mesh cells. Using the corrected eigenvalues ​​as new primary motion eigenvalues, repeat the motion continuity analysis and correction operation until the change in eigenvalues ​​generated by two consecutive iterations is less than the convergence threshold. At this point, the obtained eigenvalues ​​are determined as optimized motion eigenvalues.

8. The method for recognizing vehicle trajectory motion features based on a high-altitude perspective according to claim 7, characterized in that, The method for constructing the regional motion feature map includes: Each grid cell is abstracted as a node in the graph, and the optimized motion feature value is taken as the core attribute of the node. If two grid cells are spatially adjacent and both are identified as having motion patterns in consecutive time slices, then an edge is established between the corresponding nodes, and the edge is weighted according to the spatial propagation coefficient of the motion intensity. By combining all nodes and weighted edges, a regional motion feature map is formed with grid cells as nodes and spatial relationships as edges.

9. A method for recognizing vehicle trajectory motion features based on a high-altitude perspective according to claim 8, characterized in that, The feature propagation based on graph topology includes: Randomly select a node in the graph as the starting node, and use its optimized motion feature value as the initial propagation value; Based on the edge weights, the initial propagation value is proportionally distributed to all neighboring nodes directly connected to the starting node; Each neighboring node receives the propagation values ​​from all connected nodes and sums them up to obtain the cumulative propagation value of that neighboring node; The cumulative propagation value is fused with the node's own optimized motion feature value to update the node's motion feature value; The process involves traversing all nodes in the graph to complete multiple rounds of propagation and updates until the motion characteristic values ​​of all nodes in the graph tend to stabilize.

10. A method for recognizing vehicle trajectory motion features based on a high-altitude perspective according to claim 8, characterized in that, The specific execution steps of the consistency check include: After feature propagation is complete, the difference between the updated motion feature value of each node and the average of the updated motion feature values ​​of all its direct neighbors is calculated. If the difference exceeds the allowable deviation, the motion characteristic value of the node is smoothed to make it closer to the average value of the neighboring nodes. After traversing all nodes and completing the smoothing adjustment, the difference is calculated again, and the smoothing adjustment process is repeated until the difference of all nodes is lower than the allowable deviation. The final adjusted motion feature value is then used as the final vehicle motion feature recognition result for that node.

Images