CN-122020000-A - Vehicle track missing data completion method and system based on space-time correlation learning

Abstract

A vehicle track missing data complement method and system based on space-time correlation learning belongs to the field of intelligent traffic and track data processing. The method comprises the steps of preprocessing an original vehicle track, screening similar vehicle tracks based on preset space-time constraints in a space range and a time window corresponding to a missing zone, weighting and fusing the similar tracks according to correlation of track change trends and space distances to obtain similar track features for completion, extracting features of historical tracks before and after the missing zone of a target vehicle by utilizing a time sequence feature learning model to obtain historical track features, dynamically weighting and fusing the similar track features and the historical track features according to the relative length of the missing zone to generate fusion features for completion, and finally outputting the completed vehicle track through a reconstruction network. The invention can improve the completion precision of the long-time missing segment and maintain the time continuity and the space consistency of the track.

Inventors

• LI ZHENLONG
• DONG TIANHAO

Assignees

• 北京工业大学

Dates

Publication Date

20260512

Application Date

20260201

Claims (7)

1. 1. A vehicle trajectory deletion data complement method, comprising: (1) Acquiring vehicle track data, and performing denoising processing, abnormal point identification and missing segment marking on the vehicle track data to obtain a target track sequence containing track missing segments; (2) Screening candidate tracks meeting preset conditions on spatial proximity and time synchronism with the target track from the multi-vehicle track data based on the spatial position and the time range corresponding to the missing segment in the target track sequence, and constructing a similar track set; (3) According to the track change trend correlation and the space distance between the similar track set and the target track, weighting and fusing the similar tracks to generate a similar track feature sequence; (4) Invoking an offline pre-trained time sequence generation model, and performing feature learning on the historical tracks before and after the target track missing segment to generate a historical track feature sequence for representing a target vehicle time sequence change rule; (5) Dynamically adjusting the weights of the historical track feature sequence and the similar track feature sequence in the completion process according to the relative length of the track missing segment, and fusing the historical track feature sequence and the similar track feature sequence to generate a completion feature sequence; (6) And reconstructing vehicle track points corresponding to the missing segments based on the complement feature sequences, and outputting complete vehicle track data after complement.
2. 2. The method of claim 1, wherein: Step S1, track data acquisition and pretreatment based on Kalman filtering S11, track data acquisition Acquiring original vehicle track data from a vehicle positioning terminal, a road monitoring system or a vehicle-road cooperative infrastructure, wherein the track data at least comprises vehicle identification information, a track acquisition time stamp and a spatial position coordinate of a vehicle at a corresponding moment, wherein the spatial position coordinate is expressed in terms of longitude and latitude; S12, track denoising processing based on Kalman filtering Modeling the spatial position change of the vehicle into a discrete time state space model, wherein a state vector is used for describing the motion state of the vehicle at the current moment and at least comprises the position and speed information of the vehicle, and an observation vector is a longitude and latitude coordinate value acquired by track acquisition; s13, identifying abnormal points and marking missing fragments After the Kalman filtering denoising processing is completed, the smooth track data is further subjected to abnormal point identification and missing segment marking, track points which are obviously deviated from the normal running rule of the vehicle are identified according to the speed change amplitude and the space position jump degree between adjacent track points of the vehicle and the time interval change condition between adjacent time stamps, when the conditions that the speed jump and the space position jump between the adjacent track points exceed the reasonable running range or the time interval is abnormally increased are detected, the corresponding track points are judged to be abnormal points, and if the abnormal points or the data state continuously appears in time, the continuous interval is marked to be track missing segments, and the start-stop time and the corresponding length information of the missing segments are recorded.
3. 3. The method of claim 1, wherein: step S2, similar track screening based on space-time constraint Introducing a similar track screening method based on space-time constraint aiming at the condition that a target vehicle has track missing fragments; S21 spatial neighborhood screening Constructing a space search window by taking a space position corresponding to the missing segment of the target vehicle as a center, and screening candidate track points or track segments in the space neighborhood range from track sequences of other vehicles; S22 time window constraint Screening vehicles with track records in the same or adjacent time windows in the time range corresponding to the missing segments to ensure time synchronism among tracks; S23, track change trend correlation calculation Calculating the correlation between the candidate track and the target track based on the track position increment sequence, namely the longitude and latitude variation; measuring the track change trend by adopting a pearson correlation coefficient to obtain a correlation score of the candidate track; And finally screening to obtain a similar track set which is related to the target track on the aspects of spatial proximity, time synchronism and change trend.
4. 4. The method of claim 1, wherein: Step S3, similar track feature fusion based on correlation and distance Constructing space reference features of the screened similar track set in a weighted fusion mode; the weight calculation formula of the similar track is as follows: Wherein: Represent the first Correlation coefficients of the similar tracks and the target track; Represent the first The spatial distance between the similar track and the target track; is a correlation enhancement coefficient; Is the spatial distance attenuation coefficient; Number of similar trajectories; by the weight mechanism, tracks with high correlation and close space distance occupy larger weight in the fusion process, and similar track feature sequences for track completion are generated.
5. 5. The method of claim 1, wherein: step S4, learning historical track characteristics based on TimeGAN Introducing TimeGAN a time sequence generation model to perform feature learning on the historical track for describing the time evolution rule of the target vehicle; S41, historical track sequence construction and model input Aiming at the condition that a track missing segment exists in a target vehicle, continuous track data before and after the missing segment is intercepted respectively to be used as a history track input sequence; the historical track input sequence is used as input of TimeGAN model encoders; s42, model convergence Judging whether the training is converged or not by monitoring the change condition of the model total loss function in the pre-training process of TimeGAN models, and when the relative change rate of the model total loss function in the continuous K-wheel training iteration is lower than a preset threshold epsilon or the total loss function is not reduced within a preset round number range, considering that the model training reaches a convergence state, wherein the relative change rate threshold epsilon is 10 < -3 >, and the continuous round number K is 10 < -20 >; S43, generating based on TimeGAN historical track characteristics And sending the history track input sequence into a TimeGAN model which is finished by offline pre-training, and modeling the time sequence dependency relationship of the target vehicle track through the processes of encoding, generating and decoding.
6. 6. The method according to claim 1, wherein the step S5 is a dynamic weighted fusion based on the deletion ratio as follows: after the similar track feature fusion step and the history track feature learning step are completed, similar track feature sequences are obtained respectively And historical track feature sequence Wherein the historical track feature sequence The method is used for representing the time evolution trend of the target vehicle based on the historical track of the target vehicle, and the similar track feature sequence For characterizing a reference motion characteristic of an adjacent vehicle in a spatial dimension; S51 deletion proportion calculation The corresponding total length of the complete track of the target vehicle in the current completion process is set as The length of the trace deletion fragment is meter Meter, the track deletion ratio is defined as: S52, dynamic weight calculation Based on the deletion ratio Dynamic fusion weights are respectively distributed for the historical track feature sequences and the similar track feature sequences, and the weights of the historical track features The definition is as follows: Similar trajectory feature weights The method comprises the following steps: Wherein, gamma is a weight adjusting parameter; S53, feature fusion In obtaining dynamic weight And (3) with Then, for the history track characteristic sequence With similar track feature sequences And (3) carrying out weighted fusion to generate a complement feature sequence for track reconstruction, wherein the calculation mode is as follows: Wherein, the Representing the complement feature sequence after fusion.
7. 7. The method according to claim 1, wherein the system used in the method comprises: the data preprocessing module is used for performing track denoising, abnormal point identification and missing interval marking; a second similar track screening module for screening candidate tracks according to space-time constraint and calculating correlation; the feature learning module is used for executing time sequence learning of similar track fusion and historical tracks; a dynamic fusion module for adaptively adjusting fusion weights of different features according to the length of the missing segment; And (V) a result output module for generating and outputting the completed track data.

Description

Vehicle track missing data completion method and system based on space-time correlation learning Technical Field The invention relates to the technical field of intelligent traffic and moving target track data processing, in particular to a vehicle track data-oriented deletion complement and space-time sequence reconstruction method and system. The method particularly relates to technologies such as track data preprocessing, similar track screening, feature learning based on space-time correlation, a time sequence generation model, dynamic weighting fusion in a track complement process and the like. Background With the development of Intelligent Traffic Systems (ITS), vehicle-road coordination and moving object monitoring technologies, vehicle track data has become an important foundation for traffic management, path planning, operation monitoring and unmanned decision. However, in the actual collection process, the vehicle track is often affected by various environmental and equipment factors, and problems such as noise, jump points, interruption, large-area loss and the like are generated. For example, GPS signals drift due to weather, building shielding, tunnel or overhead structure, acquisition equipment may cause track interruption due to power consumption, communication interruption or hardware faults, and meanwhile, track point distribution is uneven and time sampling rate is inconsistent due to the change of different vehicle running states and road geometries. Aiming at track noise and loss, the prior art mainly comprises a filtering method, an interpolation method and a time sequence prediction method based on deep learning. The traditional Kalman filtering, particle filtering and Savitzky-Golay filtering can reduce track fluctuation to a certain extent, but are mainly applicable to linear or short-period noise suppression, and are difficult to effectively recover for long-time and large-area track missing fragments. Although the conventional methods such as linear interpolation, spline interpolation, polynomial interpolation and the like can handle small-scale missing, the complement result is difficult to maintain kinematic consistency under complex road geometric scenes, and the real change trend of the track cannot be accurately restored. In recent years, deep learning methods such as long-short-term memory network (LSTM), recurrent Neural Network (RNN), generation countermeasure network (GAN) and the like are applied to track prediction and sequence complement tasks, and achieve better effects in short-term loss prediction. However, the method has certain limitations that firstly, a single time sequence model is difficult to capture time sequence dependence and space correlation of vehicle tracks, secondly, the model is easy to generate trend deviation when the road structure information or the adjacent vehicle behavior information is lacked, thirdly, when the segments are lost for a long time, the accumulated error of the traditional depth model is obviously amplified, and the complement tracks are difficult to keep consistent with the real tracks. In addition, although the prior research proposes an interpolation strategy based on similar tracks, reference track integration is generally carried out by adopting a simple distance threshold or unweighted average mode, and the correlation difference among tracks, the attenuation relation of spatial proximity degree and a dynamic weight distribution mechanism of multi-source information are ignored. The lack of efficient similar trajectory screening and weighted fusion makes the complement results insufficient to meet continuity, smoothness and space-time consistency requirements in complex traffic environments. In summary, the prior art has the disadvantages of difficulty in satisfying the requirements of track noise suppression, long missing segment completion, space-time feature extraction, multi-source track fusion and the like, and has obvious shortcomings in complex traffic scenes, and a track data completion method and system capable of effectively processing large-scale missing, fusing multi-vehicle space-time features and ensuring track continuity are needed. Disclosure of Invention 1. Object of the invention Aiming at the problems that noise interference and large-area deletion are commonly existed in the existing vehicle track data and the space-time consistency is difficult to maintain by the traditional method, the invention provides a track deletion data complement method and a track deletion data complement system based on space-time correlation learning. The method aims to realize joint modeling of the track time sequence characteristics and the spatial similarity, improve the completion precision of the large-span missing segment, enhance the continuity and stability of the completion result and improve the applicability under complex traffic scenes. The invention realizes the simultaneous utilization of the historical track and