Search

CN-122020437-A - Time sequence multi-scale abnormal track detection method and system based on POI risk propagation

CN122020437ACN 122020437 ACN122020437 ACN 122020437ACN-122020437-A

Abstract

The application relates to the technical field of abnormal track detection, and provides a time sequence multi-scale abnormal track detection method and system based on POI risk propagation. Dividing a target area into grids, carrying out grid risk feature analysis based on risk propagation, determining grid risk feature representation, dividing a track into track segments with multiple time-space scales, and fusing the features of the track segments with multiple time-space scales according to the residence time of the track in each grid and the grid risk feature representation to obtain the comprehensive abnormal features of the track fused with multiple scales. By the method, the problem of insufficient association of the track semantics and the risk degree of the region is solved, association of track points, POIs and the risk of the region is realized, the problem of insufficient characterization of local detail features of the track under different scales is solved, and abnormal track detection considering local and global features is realized.

Inventors

  • LI PEIYUE
  • ZHANG YONGFEI

Assignees

  • 北京航空航天大学

Dates

Publication Date
20260512
Application Date
20251203

Claims (10)

  1. 1. The time sequence multi-scale abnormal track detection method based on POI risk propagation is characterized by comprising the following steps of: Dividing a target area into grids, carrying out grid risk feature analysis based on risk propagation, and determining grid risk feature representation ; Dividing the track into track segments with multiple time-space scales, and according to the stay time of the track in each grid And grid risk feature representation Fusing the characteristics of the multi-time-space-scale track fragments to obtain multi-scale fused track comprehensive abnormal characteristics 。
  2. 2. The method of claim 1, wherein the step of determining the position of the substrate comprises, Space-time periodic fusion is carried out on the semantic codes of the grid POIs, and the space-time information-fused space-time POI space-time semantic feature vectors are obtained ; Risk propagation-based grid neighborhood risk aggregation according to grid POI space-time semantic feature vectors Aggregating the regional risks of the grids and the neighborhood thereof to obtain a grid risk feature vector ; According to grid risk feature vectors Monitoring training grid risk feature representation through multi-layer perceptron using alert data 。
  3. 3. The method of claim 2, wherein the step of determining the position of the substrate comprises, Projecting the independent heat vector of each POI type code into POI feature vectors, and aggregating all POI feature vectors in each grid to form a grid POI semantic code ; Space-time periodic encoding of a grid, embedding the periodic time of the grid Space embedding Grid POI semantic coding Fusion is carried out through a plurality of layers of perceptrons to obtain a grid POI space-time semantic feature vector fused with space-time information 。
  4. 4. The method of claim 2, wherein the formula is as follows: aggregating the grid and the regional risk of its neighborhood; In the formula, A grid risk feature vector representing the risk of the current grid aggregation neighborhood region, Representing a non-linear activation function, Representing current grid POI spatiotemporal semantic feature vectors Is a linear transformation matrix of grid risk features of (a), Representing current grid POI spatiotemporal semantic feature vectors A linear transformation matrix of grid risk features of a neighborhood of (a); Representing the first of the current grid POI space-time semantic feature vectors of the neighborhood; Representing the number of neighborhoods of the current grid.
  5. 5. The method of claim 1, wherein the step of determining the position of the substrate comprises, Dwell time in each grid according to trajectory And grid risk feature representation Detecting regional risk exposure of the track to obtain risk exposure fusion characteristics of each grid through which the track passes To determine a sequence of feature vectors of the trajectory ; Dividing the track into track segments with multiple time-space scales, and according to the characteristic vector sequence of the track Determining dynamic characteristics of track fragments under different time-space scales ; Dynamic characterization of multi-time-space scale track segments through a trans-former encoding network Coding to obtain a multi-scale track feature vector sequence And for the multi-scale track characteristic vector sequence Performing attention pooling to obtain multi-scale fused track comprehensive abnormal characteristics Wherein, the method comprises the steps of, In the formula, On the scale of the track The number of non-overlapping track segments divided down, L is the number of all feature vectors of the track at multiple scales, Are all positive integers, the total number of the two is equal to the positive integer, Respectively representing the trace feature vectors of the trace at different scales.
  6. 6. The method of claim 5, wherein the step of determining the position of the probe is performed, Residence time of the track in each grid Vector representation is formed by multi-layer perceptron to obtain dwell time code of track in grid ; Encoding dwell time of track in grid Grid risk feature representation Fusion is carried out to obtain risk exposure fusion characteristics of the track on the grid ; Risk exposure features to grids through a two-way long and short term memory network Constructed feature vector sequence Processing to obtain a characteristic vector sequence of the track Wherein, the method comprises the steps of, , Is a positive integer representing the number of grids through which the track passes.
  7. 7. The method of claim 5, wherein the step of determining the position of the probe is performed, Dividing a track into Non-overlapping segments and are dimensioned In terms of space-time scale Feature vector sequences of respective combined trajectories Each track segment is encoded through a long-short-term memory network to obtain a scale Feature vector of lower track segment Wherein, the method comprises the steps of, In the formula, The number of grids through which the track passes; is a positive integer; The initial feature vectors of all track segments are integrated by a transducer encoder according to the space-time scale to obtain the dynamic features of the track segments under different space-time scales 。
  8. 8. The method of claim 5, wherein the formula is as follows: Determining trajectory feature vectors Weighting weights of (2) ; In the formula, As a parameter vector that can be learned, Is the first track at different scales A plurality of trajectory feature vectors; l is the number of feature vectors of all tracks of the track under multiple scales, Is a positive integer.
  9. 9. The method as recited in claim 1, further comprising: constructing a comprehensive space-time reasoning model to detect a time sequence multi-scale abnormal track based on POI risk propagation, wherein the loss function of the comprehensive space-time reasoning model is as follows: In the formula, To integrate the focus loss of the spatiotemporal inference model, To integrate the risk entropy loss of the spatio-temporal inference model, The overall loss of the comprehensive space-time reasoning model; is super parameter and takes the value of ; Is a positive integer representing the number of grids through which the track passes; based on grid risk feature vector for comprehensive space-time reasoning model The predicted grid risk entropy of the grid, The real risk entropy under the corresponding space-time condition is obtained through calculation of the historical risk event; The focus parameter is indicated as such, Representing the class balancing weighting terms, For synthesizing space-time reasoning model pair real category Is used for estimating the probability of a failure.
  10. 10. A POI risk propagation based time series multi-scale anomaly trajectory detection system, comprising: A grid risk analysis unit configured to divide the target region into grids, perform a grid risk feature analysis based on risk propagation, and determine a grid risk feature representation ; A multiscale anomaly detection unit configured to divide the trajectory into multiple time-space scale trajectory segments and to determine a residence time of the trajectory in each grid based on the residence time of the trajectory And grid risk feature representation Fusing the characteristics of the multi-time-space-scale track fragments to obtain multi-scale fused track comprehensive abnormal characteristics 。

Description

Time sequence multi-scale abnormal track detection method and system based on POI risk propagation Technical Field The application relates to the technical field of abnormal track detection, in particular to a time sequence multi-scale abnormal track detection method and system based on POI risk propagation. Background Abnormal trajectories generally refer to trajectories that stay, wander, or detour for long periods of time, unlike conventional patterns of behavior. Abnormal trajectory detection is one of the core tasks in the field of spatiotemporal data mining, aimed at identifying trajectories from large-scale trajectory data that deviate significantly from conventional behavior patterns. These anomalies may reflect traffic congestion, city infrastructure failures, personal suspicious behavior, and may even be early signals of threat to public safety. Traditional abnormal track detection methods, such as methods based on distance, density or statistical models, focus on analyzing the statistical significance of the spatiotemporal geometric features (such as speed, direction, shape) or data distribution of the track, but consider less the rich semantic information contained in the track, and this limitation makes it difficult for traditional methods to achieve ideal effects when identifying anomalies related to specific site functions, individual behavioral intentions or complex scene contexts. Points of interest (Point of Interest, POIs for short) as functional nodes (such as restaurants, malls, office buildings, subway stations, etc.) bearing specific semantic information in the geographic space can inject rich context information and semantic connotations into the trajectory data. By associating the track with the POI, the method is helpful to more deeply analyze the activity type, trip purpose and behavior mode reflected by the track, thereby enabling the definition and detection of the semantic-based track abnormality. For example, a trajectory is spatially continuous and normal in speed, but if the sequence of POIs it accesses is logically inconsistent with normal patterns (e.g., going directly from a train station to a remote park, and then to a military regulatory region), or there is long stay behavior during non-business hours for certain types of POIs (e.g., banks), the trajectory may be considered an abnormal trajectory based on POI semantics. At present, the abnormal track detection of the fusion POI has the key problems that firstly, the existing method represents the regional risk through single POI matching on the characteristic depiction of the regional risk, the POI is specifically regarded as an isolated functional node, the association is carried out according to the distance relation between the track point and the POI, the association uncertainty caused by the staggered distribution of the track and the POI is ignored, meanwhile, the space diffusion effect of the POI risk is not considered, the association of track semantics and the risk degree of the region is insufficient, the abnormal behavior usually occurs in one or more parts of the track, the existing method processes time sequence data through a mean processing method, the time sequence characteristic of the track under different scales is specifically ignored, only characteristic mean value pooling is carried out on each scale segment, and the local detail characteristic depiction of the track is insufficient, so that the abnormal track detection model is low in accuracy. Thus, there is a need to provide a solution to the above-mentioned deficiencies of the prior art. On the feature depiction of the regional risk, the regional risk is depicted through multi-POI aggregation, namely a POI risk spreading mechanism is introduced, a grid risk representing method based on risk spreading is designed, and the relationship among the track, the POI and the regional risk is established. In the track feature representation, the track segments are subjected to time sequence feature representation, namely the time sequence features of the track segments with different scales are extracted, and the local and global features of the track are considered. Disclosure of Invention The application aims to provide a time sequence multi-scale abnormal track detection method and system based on POI risk propagation, so as to solve or alleviate the problems in the prior art. In order to achieve the above object, the present application provides the following technical solutions: The application provides a time sequence multi-scale abnormal track detection method based on POI risk propagation, which comprises the steps of dividing a target area into grids, carrying out grid risk feature analysis based on risk propagation, and determining grid risk feature representation Dividing the track into track segments with multiple time-space scales, and according to the residence time of the track in each gridAnd grid risk feature representationFusing the ch