Search

CN-121997103-A - Vehicle track anomaly detection method and system based on deep learning

CN121997103ACN 121997103 ACN121997103 ACN 121997103ACN-121997103-A

Abstract

The invention relates to the technical field of vehicle track analysis, in particular to a vehicle track abnormality detection method and system based on deep learning, comprising the following steps of obtaining track coordinates and boundary distances to generate symbol offset, constructing a road offset continuous field through interpolation and space embedding, calculating the ratio of the instantaneous speed to the reference speed, generating an unbalance degree parameter, executing vector dimension stretching, extracting a track embedded vector and an abnormal score by using a long-short-period memory network, executing square summation and difference operation, and generating a track energy gradient value. According to the invention, the association between the track and the road geometric constraint is realized by mapping the offset characteristic to the continuous field, the vector scale enhancement speed dynamic perception is recalibrated by using the unbalance degree parameter, the confidence probability curved surface is constructed by cooperating with the energy gradient modulation score and using the tri-linear interpolation, the peak value of the curved surface is positioned, the detection result is output, the positioning noise is restrained, and the abnormal precision under sparse sampling is improved.

Inventors

  • CHEN WEI
  • LONG YUE
  • HUO PENGFEI
  • YU DONGLI
  • WANG BING

Assignees

  • 浙江安防职业技术学院

Dates

Publication Date
20260508
Application Date
20260408

Claims (8)

  1. 1. The vehicle track anomaly detection method based on deep learning is characterized by comprising the following steps of: S1, acquiring a vehicle track point coordinate, a displacement increment length, a sampling time interval, a road left boundary line coordinate sequence and a road right boundary line coordinate sequence through a vehicle-mounted positioning terminal, and calculating a track point left distance value from a vehicle track point coordinate to the road left boundary line coordinate sequence and a track point right distance value from the vehicle track point coordinate to the road right boundary line coordinate sequence to obtain a road distance analysis result; S2, based on the road distance analysis result, calculating a left distance value of the track point and a right distance value of the track point by a difference value to obtain a symbol offset, performing interpolation processing on the symbol offset, and constructing a symbol offset sequence and a road offset continuous field; S3, calculating the ratio of the displacement increment length to the sampling time interval, generating an instantaneous speed value, averaging the instantaneous speed value to obtain a reference speed, calculating a time proportion unbalance parameter by using the instantaneous speed value and the reference speed, and constructing a scale adjustment query vector; The step of obtaining the scale adjustment query vector comprises the following steps: S311, acquiring a displacement increment length and a sampling time interval, performing division ratio operation based on the sampling time interval aiming at the displacement increment length, generating an instantaneous speed value, and performing average value aggregation calculation on a plurality of instantaneous speed values continuously distributed in an acquisition window to obtain a reference speed; S312, calling the instantaneous speed value, the reference speed and the sampling time interval, obtaining the length of a reference road section, the curvature of a track arc and the resonant frequency, executing deviation ratio measurement aiming at the instantaneous speed value and the reference speed, and calculating to obtain a time proportion unbalance parameter; S313, constructing an original query vector based on a vehicle position, a course angle and an acceleration query instruction, calling the time proportion unbalance parameter for each characteristic dimension of the original query vector to execute dimension space linear mapping stretching transformation, calculating the time proportion unbalance parameter and a hidden state characteristic value weight product in the original query vector, and executing vector distribution scale reconstruction mapping processing to obtain a scale adjustment query vector; The formula for acquiring the time proportion unbalance parameter by operation specifically comprises the following steps: ; Wherein, the Representing the time scale imbalance parameters, Representing the value of the instantaneous velocity, Representing the reference velocity of the vehicle, Representing the normalized value of the length of the reference road segment, Represents the normalized value of the curvature of the track arc, Representing the normalized value of the sampling time interval, Representing a normalized value of the resonant frequency; S4, inputting the road deviation continuous field and the scale adjustment query vector into a pre-trained long-short-term memory network model to perform feature extraction, constructing a track embedded vector and an original track abnormal score, calculating a track energy value, calculating the change rate of the track energy value relative to a sampling time interval, and constructing a track energy gradient value; S5, comparing the track energy gradient value with a preset reference gradient threshold value, modulating an original track anomaly score to generate a corrected anomaly score, interpolating and expanding the corrected anomaly score, and establishing a confidence probability curved surface to obtain a vehicle track anomaly detection result.
  2. 2. The deep learning-based vehicle trajectory anomaly detection method of claim 1, wherein the road offset continuous field comprises a topology mapping matrix, a regional deformation tensor and a transverse distribution cloud map, the scale query vector comprises a dynamic attention mask, a feature activation map and a spatiotemporal fusion weight, the trajectory energy gradient value comprises a state mutation scalar, a fluctuation evolution amplitude and a kinetic energy attenuation index, and the vehicle trajectory anomaly detection result comprises a risk critical grade, a behavior violation category and a trajectory failure mode.
  3. 3. The vehicle track abnormality detection method based on deep learning according to claim 1, characterized in that the step of obtaining the road distance analysis result specifically includes: S111, acquiring vehicle track point coordinates, displacement increment length, sampling time interval, a left boundary line coordinate sequence of a road and a right boundary line coordinate sequence of the road through a vehicle-mounted positioning terminal, performing discretization sample alignment processing on the track point coordinates and the displacement increment length according to clock synchronous signals, and processing geographic coordinates by combining a road topological connection relationship, a data frame check rule and a projection transformation algorithm to acquire vehicle track space-time parameter characteristics; S112, based on the space-time parameter characteristics of the vehicle track, calling the coordinates of the vehicle track points and the left boundary line coordinate sequence of the road, searching the point line projection vectors of the coordinates of the vehicle track points in the coverage area of the left boundary line coordinate sequence of the road, calculating the minimum Euclidean distance from the coordinates of the vehicle track points to the matching line segments of the left boundary line coordinate sequence of the road according to the geometric distance judging rule and the coordinate interpolation logic, and generating the left distance value of the track points; S113, according to the space-time parameter characteristics of the vehicle track, a right boundary line coordinate sequence and a sampling time interval of the road are called, dynamic sampling index values of the coordinates of the track points of the matched vehicle in the distribution direction of the transverse axis are calculated, the minimum Euclidean distance between the coordinates of the track points of the vehicle and the coordinates of the right boundary line coordinate sequence of the road is used as a right distance value of the track points, the left distance value of the track points and the right distance value of the track points are arranged based on the left distance value association processing of the track points, and the road distance analysis result is obtained.
  4. 4. The vehicle track abnormality detection method based on deep learning according to claim 3, wherein the road offset continuous field acquisition step is specifically: S211, based on the road distance analysis result, calling a left distance value of a track point and a right distance value of the track point to execute difference operation, determining the lateral azimuth polarity of the track point of the vehicle relative to the central axis of the road, acquiring geometric constraint parameters of the cross section of the road, executing transverse vector mapping aiming at the geographic area where the coordinates of the track point are located, converting a space displacement scalar by combining a preset positive and negative sign judgment criterion, executing polar alignment operation and numerical value polarity correction, and generating a sign offset; S212, acquiring a road mileage sequence parameter based on the symbol offset, calling a linear smooth interpolation function to execute numerical value continuous fitting processing on the symbol offset in the road mileage evolution dimension, eliminating position mutation jump noise among discrete track sampling points, executing resampling processing based on moving average logic on the fitted numerical value, and executing second derivative constraint check and local monotonicity check on the offset change rate to obtain a signed offset sequence; S213, for the signed offset sequence, a space embedding layer weight matrix is called to execute high-dimensional feature nonlinear mapping processing, scalar offset features are converted into high-dimensional vector characterization with space topology attributes, global road topology connection constraint and lane line geometric topology distribution information are associated, each sampling point displacement state is mapped to a global space grid tensor structure, coordinate projection transformation, tensor slice aggregation and feature dimension alignment are executed, and a road offset continuous field is established.
  5. 5. The vehicle track abnormality detection method based on deep learning according to claim 1, wherein the track energy gradient value obtaining step specifically includes: S411, mapping the road offset continuous field and the scale adjustment query vector to a pre-training long-short-term memory network to perform feature extraction, calling a neuron weight matrix and bias parameters to calculate tensor dot products, updating cell state recursion iteration in cooperation with a forgetting gate, an input gate and an output gate, mapping an implicit layer output vector to a fully-connected feature mapping layer by utilizing nonlinear activation processing to aggregate a time sequence long-range dependency relationship, performing dimension reduction projection conversion, and establishing a trace embedded vector and an original trace anomaly score; S412, extracting feature value components of each dimension of a vector space aiming at the track embedded vector, calling vector two-norm calculation logic to process square operation of all dimension components, executing total-dimension square term value aggregation summation by utilizing a linear superposition operator, mapping track points in a high-dimension feature space relative to distribution density of a coordinate origin, establishing track motion state energy characterization attributes, associating time sequence weight distribution matrix enhancement features, and generating track energy values; S413, based on the track energy value, calling the sampling time interval, calculating a first-order discrete difference result of the track energy value of the adjacent track point in the time sequence, calculating a normalized quotient value operation of the first-order discrete difference result with respect to the sampling time interval, obtaining the dynamic evolution change rate of energy along with time, and combining a track kinematics constraint discriminant criterion, a data frame check rule and a smooth filtering logic correction value to obtain a track energy gradient value.
  6. 6. The vehicle track abnormality detection method based on deep learning according to claim 5, characterized in that the vehicle track abnormality detection result obtaining step specifically includes: s511, acquiring a preset reference gradient threshold, calling the track energy gradient value and the reference gradient threshold to execute numerical subtraction operation and offset deviation quantization processing, calculating the deviation amplitude of the track energy evolution state along with time relative to the reference gradient threshold, executing logic threshold discrimination and nonlinear interval mapping conversion, and generating a deviation comparison judging result; S512, acquiring the original track anomaly score based on the deviation comparison judging result, mapping a nonlinear weight correction coefficient for the deviation comparison judging result, calling the original track anomaly score to execute weight product modulation processing, calculating the product operation value of the original track anomaly score and the weight correction coefficient, executing dimension alignment and inhibition correction processing, and generating a corrected anomaly score; s513, according to the corrected abnormal scores, acquiring sampling time intervals, calling a tri-linear interpolation function to execute three-dimensional space continuous numerical value filling aiming at the corrected abnormal scores according to time sequence, mapping score numerical values in space-time dimension probability distribution density, executing smooth surface expansion processing in association with global track time sequence evolution constraint, executing grid node numerical value alignment, and establishing a continuous confidence probability surface; S514, local extremum detection optimizing is carried out on the continuous confidence probability curved surface, the position coordinates of the maximum value distribution of the numerical confidence in the continuous confidence probability curved surface are searched, the peak characteristic point of the continuous confidence probability curved surface is obtained, and the abnormal judgment of the vehicle track is carried out by combining a preset abnormal classification judgment criterion, a risk assessment index and a running state stability index, so that the abnormal detection result of the vehicle track is obtained.
  7. 7. The vehicle track anomaly detection method based on deep learning according to claim 6, wherein the setting process of the reference gradient threshold is specifically that a preset sample data set is called, the energy gradient distribution characteristics of the historical normal track points under a plurality of driving conditions are extracted, and a percentile threshold or a mean weighted variance of statistical distribution is utilized to determine a reference boundary value representing the energy evolution stability as the reference gradient threshold; the process of mapping the nonlinear weight correction coefficient specifically comprises the steps of calling a Sigmoid function or a piecewise power function to construct a nonlinear mapping link from the deviation comparison judging result to a weight scalar, matching corresponding dynamic modulation proportion in a weight saturation interval according to the deviation degree of the deviation comparison judging result relative to zero offset, and carrying out amplitude mapping on the original track anomaly score.
  8. 8. A vehicle track anomaly detection system based on deep learning, characterized in that the system is configured to implement the vehicle track anomaly detection method based on deep learning as claimed in any one of claims 1 to 7, the system comprising: The track distance analysis module is used for acquiring a vehicle track point coordinate, a displacement increment length, a sampling time interval, a road left boundary line coordinate sequence and a road right boundary line coordinate sequence through the vehicle-mounted positioning terminal, calculating a track point left distance value from the vehicle track point coordinate to the road left boundary line coordinate sequence and a track point right distance value from the vehicle track point coordinate to the road right boundary line coordinate sequence, and obtaining a road distance analysis result; The road offset dividing module is used for calculating a left distance value of the track point and a right distance value of the track point based on the road distance analysis result to obtain a symbol offset, carrying out interpolation processing on the symbol offset and constructing a symbol offset sequence and a road offset continuous field; The proportional imbalance analysis module calculates the ratio of the displacement increment length to the sampling time interval, generates an instantaneous speed value, and average value aggregates the instantaneous speed value to obtain a reference speed, calculates a time proportional imbalance parameter by using the instantaneous speed value and the reference speed, and constructs a scale adjustment query vector; the track gradient analysis module is used for inputting the road deviation continuous field and the scale adjustment query vector into a pre-trained long-short-period memory network model to perform feature extraction, constructing a track embedded vector and an original track abnormal score, calculating a track energy value, calculating the change rate of the track energy value relative to a sampling time interval, and constructing a track energy gradient value; And the track anomaly judging module is used for comparing the track energy gradient value with a preset reference gradient threshold value, modulating an original track anomaly score, generating a corrected anomaly score, interpolating and expanding the corrected anomaly score, and establishing a confidence probability curved surface to obtain a vehicle track anomaly detection result.

Description

Vehicle track anomaly detection method and system based on deep learning Technical Field The invention relates to the technical field of vehicle track analysis, in particular to a vehicle track abnormality detection method and system based on deep learning. Background The technical field of vehicle track analysis mainly relates to acquisition, modeling and analysis of space-time motion data of a vehicle in a road network, and the core matters comprise acquisition of basic data such as longitude and latitude coordinates, speed, acceleration, course angle, time stamp and the like of the vehicle through a satellite positioning device, an inertial measurement unit, a vehicle-mounted communication terminal and road side sensing equipment, coordinate conversion, track segmentation, track matching and path reconstruction of continuous track points, and construction of a vehicle motion behavior model by combining road topological structure, traffic rules and historical operation data, so that systematic analysis of vehicle operation states, path selection, running rules and abnormal behaviors is carried out. The conventional vehicle track abnormality detection method is an identification method developed by a pointer in the situation that a vehicle deviates from a conventional driving path, abnormal speed change or abnormal stay behavior and the like in the running process, the technical matters aimed at are that whether the track deviates from a normal driving mode in a certain time period is judged from a large number of continuous track point sequences, the conventional method generally carries out interval division on speed values, acceleration values and direction change angles according to manually set thresholds, track points exceeding a set range are marked as suspicious points, euclidean distance or dynamic time regular distance between tracks is calculated according to a historical track sample counted in advance, the track to be detected is compared with a standard track template point by point, the track to be detected is judged to be abnormal when the integral deviation accumulated value exceeds the set proportion, or the characteristic values such as average speed, maximum yaw angle and stay time length of each section are extracted after the track is segmented, the abnormal type is judged by point by line by point by utilizing a manually constructed rule table, and the identification of the abnormal condition of the track is completed. The traditional detection method relies on manually setting a fixed threshold value to perform interval division on speed and yaw angle, so that the abnormality judgment standard under complex and changeable driving environment lacks self-adaptability, the road geometric topological constraint is ignored to limit the vehicle displacement simply by calculating Euclidean distance between tracks or comparing dynamic time regular distance with a standard template, the static template matching mode is difficult to cover diversified driving styles and long-period time sequence evolution rules, the positioning error sensitivity of discrete sampling points is high, false alarm and false alarm phenomena are easy to occur, and especially when track data are sparse or sampling intervals are uneven, the traditional method cannot effectively mine hidden motion modes, and is difficult to support accurate abnormality recognition requirements under high-dynamic traffic scenes, so that the overall detection robustness and generalization capability are limited. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a vehicle track anomaly detection method and system based on deep learning. In order to achieve the above purpose, the invention adopts the following technical scheme that the vehicle track anomaly detection method based on deep learning comprises the following steps: S1, acquiring a vehicle track point coordinate, a displacement increment length, a sampling time interval, a road left boundary line coordinate sequence and a road right boundary line coordinate sequence through a vehicle-mounted positioning terminal, and calculating a track point left distance value from a vehicle track point coordinate to the road left boundary line coordinate sequence and a track point right distance value from the vehicle track point coordinate to the road right boundary line coordinate sequence to obtain a road distance analysis result; S2, based on the road distance analysis result, calculating a left distance value of the track point and a right distance value of the track point by a difference value to obtain a symbol offset, performing interpolation processing on the symbol offset, and constructing a symbol offset sequence and a road offset continuous field; S3, calculating the ratio of the displacement increment length to the sampling time interval, generating an instantaneous speed value, averaging the instantaneous speed value to obtain a refer