CN-121834375-B - Vehicle track similarity calculation method based on dynamic graph alignment and POI enhancement

CN121834375BCN 121834375 BCN121834375 BCN 121834375BCN-121834375-B

Abstract

The invention discloses a vehicle track similarity calculation method based on dynamic graph alignment and POI enhancement, which comprises the following steps of preprocessing tracks and map matching, generating multi-scale grid representation by adopting curvature-density dual-constraint self-adaptive grid division, combining POI distribution density with functional weight injection area semantics, guiding a graph attention network to learn road representation by using transition probability and co-occurrence probability counted by track data, constructing a dynamic graph taking GPS, grids and roads as nodes, dynamically calculating side weights based on multi-modal associated indexes, realizing cross-modal semantic alignment and fusion through the graph attention network, finally optimizing a model by fusion local structure maintenance, global contrast learning and modal regular joint loss function, outputting track similarity, improving geometric fidelity, semantic richness of track representation, robustness to noise and multi-modal fusion self-adaptation of track similarity calculation, and obviously improving accuracy of track similarity calculation.

Inventors

LIU LIYAN
SU YILIN
ZHOU XINMIN
QIU RUQIN
LI TIANLE
Wang Bingcan

Assignees

湖南工商大学

Dates

Publication Date: 20260508
Application Date: 20260311

Claims (10)

1. The vehicle track similarity calculation method based on dynamic graph alignment and POI enhancement is characterized by comprising the following steps of: S1, for original vehicle track point sequence Preprocessing, extracting space-time characteristics of a GPS point sequence, constructing a road network topological graph based on OpenStreetMap, and mapping the GPS point sequence into a structured road sequence through a map matching algorithm; S2, dynamically adjusting the grid resolution through a three-dimensional space-time octree structure according to the density constraint and the geometric curvature constraint of the vehicle track points in each candidate grid unit, extracting multi-level grid codes of leaf grid nodes and father nodes to which each vehicle track point belongs, and generating a geometric grid representation sequence of the vehicle track points through a multi-scale gating fusion mechanism; S3, acquiring POI data of different categories in each grid unit, counting POI data distribution density and functional weight, constructing semantic feature vectors of the grid units, dynamically fusing the semantic feature vectors with features of corresponding points in the grid characterization sequence, and generating a semantic enhanced grid characterization sequence; S4, counting transition probabilities between adjacent road segments and co-occurrence probabilities of the road segments in a sliding time window from the structured road sequence by adopting a double-probability-guided graph annotation semantic mechanism as learnable priori knowledge, and outputting final road topology semantic representation fused with static road attributes, topology connection and dynamic traffic rules through a multi-layer multi-head attention mechanism; s5, constructing a multi-mode semantic alignment mechanism of dynamic graph interaction, and generating a unified track multi-mode representation through a global aggregation node; S501, respectively using the GPS point sequence in the step S1, the enhanced Grid characterization sequence of the semantics of the step S3 and the final road topology semantic characterization in the step S4 to calculate cross-mode association weights among three modes of GPS mode characteristics, grid mode characteristics and road mode characteristics, converging the cross-mode association weights through a sequence reading module to obtain GPS mode track level representation, grid mode track level representation and road mode track level representation, and constructing inter-mode dynamic side weights according to the cross-mode association weights as initial characteristics of the three modes of a dynamic graph; The sequence reading module adopts average pooling and attention pooling; s502, constructing a fully-connected dynamic graph with three modes as nodes and associated weights as edges, carrying out iterative cross-mode information propagation and feature alignment through a multi-layer graph attention network, and generating unified track multi-mode characterization by aggregation; S6, constructing a joint loss function comprising local structure alignment loss, global contrast learning loss and modal fusion gating regularization loss, and performing supervision training and optimization on the unified track multi-modal characterization generated in the step S5; and S7, extracting unified multi-mode characterization of the two tracks to be compared by adopting the optimized model, and calculating the similarity, wherein the similarity is output as a final similarity score of the two tracks.
2. The method for calculating the similarity of the vehicle track based on the dynamic image alignment and the POI enhancement according to claim 1, wherein the preprocessing and the map matching specifically comprise the following steps: s101, cleaning original vehicle track point sequence Extracting GPS modal characteristics of the vehicle track from tracks with the number of medium track points smaller than 20 and larger than 200; S102, extracting a road network topological graph from OpenStreetMap data Wherein the node sets Aggregate representing road intersection and end points, edge set The method comprises the steps of representing directed road sections of connection nodes, stacking all the road section initial vectors into a matrix, wherein each road section initial vector comprises a road section length, a road type, the number of lanes and a speed limit; S103, the original vehicle track point sequence is processed Projecting the map matching to a road network topological graph, adopting a hidden Markov model to carry out map matching, determining the state transition probability by the topological connectivity and the direction consistency among road sections, and determining the emission probability by the distance between a GPS point and a candidate road section.
3. The method for calculating the similarity of the vehicle track enhanced based on the dynamic image alignment and the POI as claimed in claim 1, wherein the step of generating the final grid characterization sequence of the vehicle track points specifically comprises the following steps: s201, mapping the vehicle track points into a normalized three-dimensional space-time coordinate system to construct a normalized track point sequence ; S202, calculating density constraint and geometric curvature constraint, and carrying out normalization on the track point sequence Mapping to a three-dimensional spatiotemporal cube space; The density constraint is obtained by counting the total number of track points in the nodes; Extracting and arranging ordered point sequences of single tracks in grid units, calculating local curvatures of the tracks in nodes, performing cross-track aggregation on the curvatures of a plurality of tracks in the nodes to obtain geometric curvature constraint of the nodes, and performing octree division on grids with density constraint or geometric curvature constraint exceeding a preset threshold to obtain 8 three-dimensional small grids; S203, calculating vehicle track points through multi-level coding and microscopic displacement Leaf grid node and father node, and calculates vehicle track point Relative positional deviations in the belonging leaf network; s204, linearly fusing the leaf grid node characteristics, the father node characteristics and the microscopic displacement characteristics through a multi-scale gating fusion mechanism to generate a geometric grid representation sequence of the vehicle track points.
4. The method for calculating the similarity of the vehicle track based on the dynamic graph alignment and the POI enhancement according to claim 1, wherein the generating the semantically enhanced grid characterization sequence specifically comprises: s301, according to each grid cell Counting the distribution density of the K-type POI data in the space range, and fusing the distribution density with the corresponding preset POI data function weight to form a semantic feature vector; s302, obtaining each track point POI semantic feature vector of leaf node where the POI semantic feature vector is located POI semantic feature vector for parent node Calculating semantic fusion weights through a bilinear attention mechanism; the bilinear attention mechanism carries out outer product operation on POI semantic feature vectors of leaf nodes and father nodes where track points are located, obtains fusion weights through a learnable weight matrix and activation function calculation, carries out weighted fusion on the two semantic vectors, and adds the fusion weights with the geometric grid characterization sequence; S303, carrying out multi-mode fusion on POI semantic feature vectors of leaf nodes where the track points are located and POI semantic feature vectors of parent nodes and geometric semantic features, and combining semantic enhancement characterization of all the track points to generate a semantic enhancement grid characterization sequence.
5. The method for calculating the vehicle track similarity based on dynamic graph alignment and POI enhancement according to claim 1, wherein the step of performing road topology representation learning by adopting a graph annotation force mechanism guided by double probabilities to generate final track road topology semantic representation specifically comprises the following steps: S401, acquiring an initial feature vector corresponding to the structured road sequence by searching, constructing a road sequence initial feature matrix of a current track, and counting transition probability between adjacent road segments and co-occurrence probability of the road segments in a sliding time window from historical road track data; S402, calculating attention coefficients between adjacent road sections, normalizing through Softmax to obtain attention weights, adopting K attention heads to perform parallel calculation, performing linear transformation after all the attention heads are output and spliced, and repeating for L times to obtain final track road topology semantic representation.
6. The method for calculating the vehicle track similarity based on dynamic graph alignment and POI enhancement according to claim 1, wherein the indexes of the cross-modal association weights among the three modes of the GPS modal characteristic, the grid modal characteristic and the road modal characteristic comprise GPS-grid edge weight, grid-road edge weight and road-GPS edge weight; The GPS-grid edge weight is obtained by weighted synthesis of two sub-indexes of space coverage and time synchronism, the space coverage represents the space coverage concentration of GPS points in the grid, and the time synchronism represents the synchronism of an access time mode and a grid history time mode; The grid-road edge weight is obtained by weighted synthesis of three sub-indexes of space alignment, functional consistency and topological constraint intensity, wherein the space alignment represents the geometric coverage length proportion of a road segment in a grid, the functional consistency represents the functional consistency of grid POI semantics and road attributes, and the topological constraint intensity represents the degree of density of the grid penetrated or connected by a road network; The road-GPS edge weight is obtained by weighting and synthesizing three sub-indexes of matching precision, motion rationality and topology constraint satisfaction, wherein the matching precision represents projection precision of map matching, the motion rationality comprises speed rationality and direction consistency, the speed rationality represents the rationality of GPS instantaneous speed and road section speed limit, the direction consistency represents the motion direction consistency, and the topology constraint satisfaction represents the topology connectivity of a matching path.
7. The method for calculating the similarity of vehicle trajectories based on dynamic graph alignment and POI enhancement of claim 6, wherein the attention coefficients of the multi-layer graph attention network are determined by node feature similarity and dynamic edge weights.
8. The method for calculating the similarity of the vehicle track based on the dynamic image alignment and the POI enhancement according to claim 1, wherein the joint loss function expression is: Wherein, the Representing the joint loss function of the joint, Indicating a loss of alignment of the local structure, Representing a global contrast learning penalty, Representing the modal gating regularization loss, 、 Super-parameters representing the contribution of the balance three losses; minimizing joint loss function by back propagation The synchronous optimization track characterizes all the learnable parameters of the network and the dynamic fusion gating network.
9. The method for calculating the similarity of the vehicle track based on the dynamic image alignment and the POI enhancement according to claim 1, wherein the similarity of two tracks to be compared is calculated by adopting cosine similarity, and the expression is as follows: Wherein, the 、 Two tracks to be compared are represented, 、 Respectively represent the tracks 、 Is used to determine the characterization vector of (c), The dot product is represented by a graph of the dot product, Representing vectors Norms, similarity The range of the values is as follows The larger the value, the higher the trajectory semantic similarity.
10. The method for dynamic graph alignment and POI enhanced vehicle trajectory similarity calculation of claim 9, wherein an adjustable temperature parameter is introduced Scaling and shifting the track similarity, wherein the expression is as follows: Wherein, the A parameter of the temperature is indicated and, For controlling the sharpness of the distribution of the score.

Description

Vehicle track similarity calculation method based on dynamic graph alignment and POI enhancement Technical Field The invention relates to the technical field of intelligent transportation, in particular to a vehicle track similarity calculation method based on dynamic graph alignment and POI enhancement. Background Along with the rapid development of global positioning system, mobile communication and Internet of things technologies, the urban traffic system regenerates massive vehicle track data, which not only records the space-time position change of vehicles, but also implies multidimensional information such as driving behavior, road network structure, regional functional characteristics, urban dynamics and the like; The current track similarity calculation can be mainly divided into a geometric distance-based method, a grid-based or grating-based method, a road network matching-based method and a deep learning characterization-based method, however, in a complex and dynamic urban traffic environment, the existing method still faces the following challenges and limitations: Traditional similarity measurement methods, such as dynamic time-warping DTW, longest public subsequence LCSS, edit distance EDR, freche distance Fre CHET DISTANCE and the like, are mainly calculated based on point sequence geometric forms in European space, and can effectively describe shape similarity and time sequence alignment relations of tracks, but generally neglect topology constraints (such as single-line, steering prohibition, overpass layered traffic and the like) and semantic attributes (such as road grade and speed limit information) of an urban road network, often cause misjudgment that geometric similarity is caused, but road network is unreachable or geometric dissimilarity is caused, but the path is equivalent, so that the practicability in real navigation and path pushing is limited; In order to reduce the calculation complexity of continuous track data and enhance the robustness to noise, a gridding or gridding method is widely used, continuous space discretization is carried out, the problem that pure geometric distance or pure point sequence codes are unstable under complex geometric forms such as boundary areas, sharp turns, turning around and the like is effectively solved, however, the traditional method usually adopts uniform gridding division, the resolution is often required to be preset, the too high resolution can lead to lengthy sparse sequences in high-density areas (such as urban centers), the storage and calculation cost is increased, and key details are lost in low-density areas or complex geometric areas (such as sharp turns, roundabout and ramp) when the resolution is too low, so that characterization distortion is caused; The track similarity is not only dependent on the space geometry, but also closely related to the functional semantics of the path region, such as the distribution of POI categories around the starting and ending points, the attribute of the functional region, the activity purpose and the like, the POI data of interest points are important sources for carrying the functional semantics of the region, the existing method generally adopts the POI categories related to the whole track to carry out global aggregation when fusing POI information, or uses the POI of the starting and ending points of the track as the characteristics, the accurate association with the fine granularity space inside the track is lacking, the functional change of the track passing through the region in the advancing process is difficult to be described, and the differentiated influence of different POI categories in the same region on the traveling behavior cannot be distinguished, so that the performance is limited in the retrieval of the similarity based on the traveling intention; Mapping the GPS track to the road network through map matching, and further learning road section representation by using the map neural network is an effective way for introducing topology constraint, however, the conventional road surface representation method based on the map neural network mainly relies on static topological connection relations and road attributes (length, type, lane number and the like) provided by OpenStreetMap and the like, and fails to fully mine and utilize dynamic traffic rules which are exhibited from large-scale historical track data, ignores the behavior priori obtained by statistics from the data, so that the learned road surface representation lacks the depicting capability of an actual traffic mode, and has weaker generalization capability when facing temporary traffic control, congestion detouring and other scenes; the existing vehicle track similarity calculation method still has significant defects in the aspects of self-adaptive multi-scale geometric characterization, fine-granularity semantic fusion, data-driven road network behavior semantic modeling, dynamic robust multi-mode alignment,