Method and device for restoring trajectory of unlicensed vehicle

Abstract

The embodiment of the application provides a licenseless vehicle trajectory restoration method and device, which realizes accurate identification of a vehicle through innovative design of a multi-modal feature fusion model, feature extraction and data alignment. A trajectory screening system is constructed, combined with time and space constraints and path verification, to establish a reliable trajectory restoration mechanism. Verification optimization is introduced, and through reliability evaluation and rule adjustment, the accuracy of the restoration result is ensured. The method effectively solves the shortcomings of traditional technologies in feature fusion, trajectory screening and verification optimization, and provides technical support for licenseless vehicle management.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method and apparatus for reconstructing the trajectory of unlicensed vehicles. Background Technology

[0002] Existing methods for reconstructing the trajectories of unlicensed vehicles have significant shortcomings. Traditional systems perform poorly in feature extraction and data fusion, failing to effectively integrate multi-source information and affecting recognition accuracy.

[0003] Furthermore, existing technologies suffer from bottlenecks in trajectory selection and constraint processing. Most systems lack robust spatiotemporal constraint mechanisms and path reachability verification strategies, resulting in inaccurate trajectory reconstruction.

[0004] The existing system has technical shortcomings in verification and optimization. It lacks in-depth analysis of feature reliability, making it difficult to achieve accurate trajectory repair through dynamic adjustments, thus affecting the reconstruction effect. Solving these problems is of great significance for improving the accuracy of trajectory reconstruction. Summary of the Invention

[0005] To address the problems in the existing technology, this application provides a method and apparatus for reconstructing the trajectory of unlicensed vehicles, which can effectively solve the shortcomings of traditional technologies in feature fusion, trajectory screening and verification optimization, and provide technical support for the management of unlicensed vehicles.

[0006] To solve at least one of the above problems, this application provides the following technical solution:

[0007] Firstly, this application provides a method for reconstructing the trajectory of an unlicensed vehicle, including:

[0008] Vehicle images and location / time data captured by road surveillance cameras are collected. The vehicle images are input into a feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings. The vehicle images are then input into a convolutional neural network to extract deep visual feature vectors. Vehicle size data from radar sensors and identity data from an electronic toll collection system are collected. The structured features, deep visual feature vectors, vehicle size data, and identity data are aligned by timestamps to construct a feature fusion dataset.

[0009] The system performs preliminary screening of vehicles in video data based on structured features, calculates cosine similarity of the depth visual feature vectors for precise matching, reads road topology data from electronic maps, calculates travel time constraints and path accessibility constraints for adjacent capture points, uses the travel time constraints to remove trajectory points that violate driving speed limits, filters trajectory points that cross physical obstacles based on the path accessibility constraints, dynamically adjusts spatiotemporal thresholds according to road type, and fits missing road segments using road connectivity relationships.

[0010] The filtered vehicle trajectory sequences are input into the batch verification interface. The similarity threshold and matching weight are dynamically adjusted based on human feedback. The camera location information is read to construct a feature reliability scoring matrix. The feature reliability score is written into the rule engine. The feature matching priority under different scenarios is adaptively adjusted. The road traffic rules are used to filter trajectory segments that violate traffic rules. The trajectory repair scheme is generated based on road connectivity constraints. A trajectory restoration database containing vehicle features, trajectory segments, and verification records is constructed.

[0011] Furthermore, it also includes: acquiring vehicle capture images from road monitoring cameras, reading latitude and longitude coordinates and timestamp information at the time of image acquisition, preprocessing the vehicle capture images, performing image enhancement and size normalization, reading vehicle type classification dictionary and vehicle color dictionary, inputting the normalized image into a feature extraction model, recognizing the vehicle outline based on an edge detection algorithm, and extracting structured features of vehicle type, color, number of axles, and sunroof markings based on a region segmentation algorithm;

[0012] A convolutional neural network comprising an encoding layer, a pooling layer, and a fully connected layer is constructed. The normalized image is input into the encoding layer to extract local feature maps. The feature maps are input into the pooling layer to reduce the data dimensionality. The dimensionality-reduced features are input into the fully connected layer to generate a deep visual feature vector. Normalization calculation is performed on the deep visual feature vector to generate a feature representation of the vehicle image.

[0013] Furthermore, it also includes: reading roadside millimeter-wave radar data stream, parsing the radar data stream into target detection data, extracting vehicle length, width and height dimension parameters and motion state parameters, reading vehicle passage data collected by the electronic toll collection system, parsing electronic tag information to obtain vehicle identity features, grouping the vehicle dimension parameters and identity features according to the collection location and lane number, and constructing a multi-source data correspondence table;

[0014] Read the structured features and deep visual feature vectors, sort the structured features, deep visual feature vectors, vehicle size parameters, and identity features according to timestamp information, calculate the temporal correspondence between features based on a sliding time window, perform interpolation alignment on the feature data, combine the aligned multimodal features to generate a fused feature vector, and construct a feature fusion dataset.

[0015] Furthermore, it also includes: reading the structured features in the feature fusion dataset, performing a coarse selection of vehicles in the video data based on vehicle type classification and color matching, calculating the cosine similarity of each pair of deep visual feature vectors in the coarse selection results, constructing a candidate vehicle similarity matrix, setting a similarity threshold to filter high-confidence matching pairs, writing the spatiotemporal information of the high-confidence matching pairs into a trajectory data table, and constructing a vehicle motion trajectory sequence based on the trajectory data table;

[0016] Read the road topology data from the electronic map, extract the road connection relationships and traffic rules, calculate the shortest path distance between adjacent capture points, calculate the theoretical minimum travel time based on the road speed limit standard, set the theoretical minimum travel time as the lower limit of the time constraint, compare the lower limit of the constraint with the time interval in the trajectory data table, and remove trajectory points whose time interval is less than the lower limit of the constraint.

[0017] Furthermore, it also includes: reading physical obstacle data from electronic maps, extracting spatial distribution information of medians, rivers, and buildings, calculating a set of reachable paths between trajectory points based on road topology, matching the set of reachable paths with trajectory sequences, identifying trajectory segments that cross physical obstacles, marking the trajectory segments as invalid trajectories and deleting them from the trajectory data table, and filtering abnormal trajectory points based on path reachability constraints;

[0018] Read road attribute information and classify road types into expressways, urban expressways, and urban arterial roads. Calculate the maximum permissible speed based on road capacity and write the maximum permissible speed into a spatiotemporal constraint parameter table. Dynamically set the trajectory point time interval threshold according to road type. Read road connectivity relationships to construct a road network topology map. Generate candidate paths for missing road segments based on the shortest path algorithm. Sort the candidate paths by road level and select the path with the highest weighted score to fit and complete the trajectory.

[0019] Furthermore, it also includes: grouping vehicle trajectory sequences by time window, reading the feature matching results and constraint filtering results of each group of trajectories, constructing a verification data table containing trajectory number, vehicle features, and matching confidence, inputting the verification data table into a batch verification interface, recording manually confirmed mismatched trajectories and missed detection trajectories, calculating feature similarity correction coefficients based on the manual feedback results, applying the correction coefficients to the dynamic adjustment of similarity thresholds, and performing adaptive updates on feature matching weights;

[0020] The system reads the camera deployment location and coverage area, calculates the feature recognition accuracy at each location, calculates the feature reliability score under different lighting and weather conditions, writes the feature reliability score into a scoring matrix, constructs feature weight adaptive rules based on the scoring matrix, deploys the adaptive rules to the rule engine, and dynamically configures the feature matching strategy for different scenarios.

[0021] Furthermore, it also includes: reading the feature reliability scores in the rule engine, sorting the vehicle features in descending order of reliability scores, dynamically allocating feature matching priorities based on the sorting results, reading road traffic management regulations, extracting motor vehicle traffic rules and turning restriction rules, converting the traffic rules into trajectory constraints, identifying trajectory segments that violate one-way traffic, prohibition of left turns, and prohibition of U-turns, writing the violation trajectory segments into the trajectory table to be repaired, and generating alternative paths that comply with traffic rules based on road connectivity.

[0022] A trajectory reconstruction database is constructed by writing vehicle structured features, depth visual feature vectors, and multimodal sensor features into a feature table, trajectory point coordinate sequences, timestamp sequences, and road segment number sequences into a trajectory table, and manual verification records, feature matching records, and trajectory repair records into an operation record table. The relationship between the feature table, trajectory table, and operation record table is established to form a complete trajectory reconstruction dataset.

[0023] Secondly, this application provides a device for reconstructing the trajectory of an unlicensed vehicle, comprising:

[0024] The feature fusion module is used to collect vehicle images and location and time data captured by road monitoring cameras, input the vehicle images into the feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings, input the vehicle images into a convolutional neural network to extract deep visual feature vectors, collect vehicle size data from radar sensors and identity data from the electronic toll collection system, and align the structured features, deep visual feature vectors, vehicle size data, and identity data by timestamps to construct a feature fusion dataset.

[0025] The road analysis module is used to perform preliminary screening of vehicles in video data based on structured features, calculate cosine similarity of the depth visual feature vectors for precise matching, read road topology data from electronic maps, calculate travel time constraints and path accessibility constraints of adjacent capture points, use the travel time constraints to remove trajectory points that violate driving speed limits, filter trajectory points that cross physical obstacles based on the path accessibility constraints, dynamically adjust spatiotemporal thresholds according to road type, and fit missing road segments using road connectivity relationships.

[0026] The trajectory restoration module is used to input the filtered vehicle trajectory sequences into the batch verification interface, dynamically adjust the similarity threshold and matching weight based on human feedback, read camera location information to construct a feature reliability scoring matrix, write the feature reliability score into the rule engine, adaptively adjust the feature matching priority under different scenarios, filter trajectory segments that violate traffic rules using road traffic rules, generate trajectory repair schemes based on road connectivity constraints, and construct a trajectory restoration database containing vehicle features, trajectory segments, and verification records.

[0027] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method for restoring the trajectory of a license plateless vehicle.

[0028] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for reconstructing the trajectory of a vehicle without license plates.

[0029] Fifthly, this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the described method for reconstructing the trajectory of a vehicle without a license plate.

[0030] As described above, this application provides a method and apparatus for reconstructing the trajectory of unlicensed vehicles. Through an innovative design of a multimodal feature fusion model, accurate vehicle identification is achieved via feature extraction and data alignment. A trajectory screening system is constructed, combining spatiotemporal constraints and path verification to establish a reliable trajectory reconstruction mechanism. Verification optimization is introduced, ensuring the accuracy of the reconstruction results through reliability assessment and rule adjustment. This method effectively addresses the shortcomings of traditional technologies in feature fusion, trajectory screening, and verification optimization, providing technical support for the management of unlicensed vehicles. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the method for reconstructing the trajectory of an unlicensed vehicle in an embodiment of this application.

[0033] Figure 2 This is a structural diagram of the unlicensed vehicle trajectory reconstruction device in the embodiments of this application;

[0034] Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0035] Figure label:

[0036] Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0039] To address the problems existing in current technologies, this application provides a method and apparatus for reconstructing the trajectory of unlicensed vehicles. By innovatively designing a multimodal feature fusion model, accurate vehicle identification is achieved through feature extraction and data alignment. A trajectory screening system is constructed, combining spatiotemporal constraints and path verification to establish a reliable trajectory reconstruction mechanism. Verification optimization is introduced, ensuring the accuracy of the reconstruction results through reliability assessment and rule adjustment. This method effectively solves the shortcomings of traditional technologies in feature fusion, trajectory screening, and verification optimization, providing technical support for the management of unlicensed vehicles.

[0040] To effectively address the shortcomings of traditional technologies in feature fusion, trajectory selection, and verification optimization, and to provide technical support for the management of unlicensed vehicles, this application provides an embodiment of a method for reconstructing the trajectory of unlicensed vehicles. See [link to embodiment]. Figure 1 The method for reconstructing the trajectory of unlicensed vehicles specifically includes the following:

[0041] Step S101: Collect vehicle images and location time data captured by road monitoring cameras, input the vehicle images into a feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings, input the vehicle images into a convolutional neural network to extract deep visual feature vectors, collect vehicle size data from radar sensors and identity data from the electronic toll collection system, and align the structured features, deep visual feature vectors, vehicle size data, and identity data according to timestamps to construct a feature fusion dataset;

[0042] Optionally, this embodiment focuses on S101, starting with the monitoring system of a typical urban expressway, integrating camera capture, millimeter-wave radar, and ETC data streams into a "alignable and searchable" feature fusion dataset. First, the video input is processed: intersection and road segment cameras capture cross-line frames at a fixed focal length, along with latitude, longitude, and device clock time. Since the original image sizes and noise distributions vary, this embodiment performs dehazing, color equalization, distortion correction, and size normalization on each frame, preserving the license plate area without relying on its readability. After normalization, the image enters the feature extraction model, which consists of an edge detection branch and a region segmentation branch working together. The former generates stable shape boundaries at the vehicle's outline, while the latter extracts category labels and count information from semantic regions such as the body, window area, and tires. The vehicle type is determined by a combination of outline proportion and headlight shape. Color is determined using a multispectral approximation color dictionary with voting under shadow and nighttime lighting conditions. The number of axles is verified by counting connected components in the wheel area and wheel arch rhythm. The sunroof marking is determined by the dark texture and reflective variations in the roof area. These outputs are encoded into structured feature vectors with clearly defined fields, facilitating subsequent retrieval.

[0043] This embodiment does not rely solely on structured labels, as vehicles without license plates often exhibit high similarity in color and model. Therefore, the same normalized image is fed in parallel into a convolutional neural network to extract deep visual feature vectors. The network employs lightweight encoding layers to capture edges, textures, and logo fragments, pooling layers to suppress scale variations, and fully connected layers to form embedded representations. To avoid scale drift caused by lighting and viewpoint differences, the output vector is L2 normalized for subsequent cosine similarity matching. The inherent connection here is intuitive: structured features provide "coarse binning," while depth vectors perform "fine-tuning" within each bin. The two complement each other in terms of the stable attributes and detailed textures of the same vehicle, reducing the risk of single feature failure.

[0044] On the sensor side, millimeter-wave radar is deployed along the lane, outputting target tracking ID, speed, and RCS reflection information. We focus on extracting estimated length, width, height, and vehicle heading. Since size estimation under radar systems is affected by attitude, this embodiment applies median filtering to the size of the same target across several time slices, using the stable speed and heading segments as a confidence interval. Size data is only written if it falls within this confidence interval. The identity data provided by the ETC system includes the electronic tag ID, channel number, timestamp, and vehicle type code read from the antenna gantry. Vehicles without license plates may not have a valid tag, but we still retain the "not read / pseudo code" state as one of the identity feature dimensions, so it can appear as a difference signal in subsequent fusion.

[0045] Alignment is the core challenge of S101. There are discrepancies between the three time sources: video, radar, and ETC. This embodiment first uses the PTP master clock of the road management center as a reference to estimate the offset of each device from the master clock, and then maps the capture time, radar time, and gantry time onto a unified axis. Spatially, the camera pixel coordinates are projected onto the local road coordinate system through calibration extrinsic parameters. The radar and gantry positions are pre-marked in the GIS, and lane numbers become natural matching anchors. We use a sliding time window to find cross-source correspondences of the same vehicle, calculating temporal consistency and lane consistency scores within the window. If the structured features (vehicle type, color, number of axles) are consistent with the radar size within the tolerance, and the similarity between the depth vector and historical candidates with the same trajectory exceeds a threshold, then it is determined to be the same vehicle. To avoid information loss due to occasional occlusion, samples that do not meet the threshold but are close to the boundary are first interpolated and buffered, waiting for the next capture to supplement the evidence before deciding whether to accept or reject them.

[0046] This embodiment employs a weighted concatenation approach when generating the fusion vector. Specifically, the fusion vector F consists of four parts: structured features S, depth vector V, radar size R, and identity features E. Each part is combined according to a reliability weight w, where w is derived from estimates of device health status, lighting conditions, and lane congestion. To avoid excessive formulaic descriptions, only a minimal expression is used: F = [wS·S, wV·V, wR·R, wE·E], where wS, wV, wR, and wE represent the reliability of the four types of features (structured, depth, size, and identity) at the current moment, and S, V, R, and E are the corresponding feature sub-vectors. The physical meaning of the weight is "the strength of the feature's contribution to the consistency of the same vehicle in the current scene." wV increases during nighttime rain (textures are still comparable), wS decreases during strong midday reflections, and wE is zero when the ETC gantry is not reading a gear, but the "not read" state itself still occupies a space to prevent information from being mistakenly deleted.

[0047] To translate the abstract process into practice, this example presents two real-world scenarios. At the tunnel entrance in the afternoon, the video exhibits severe color drift, with structured color labels repeatedly switching between "silver-gray" and "light gold." Meanwhile, the radar dimensions provide stable length and width values ​​at stable vehicle speeds, the depth vector maintains high similarity at the texture level, and the weighting mechanism allows wR and wV to dominate. The fused vector shows good continuity along the timeline. The other scenario involves an elevated ramp where the ETC gantry and checkpoint are close but asynchronous. Some vehicles experience gaps where the gantry reads the data, but the checkpoint misses it. The time window is filled using trajectory context and lane number. The identity feature E is valid at the gantry point, and at the checkpoint, a very short buffer is extrapolated based on the most recent valid value to prevent sequence breaks.

[0048] Data quality is not constant. This embodiment adds anomaly labeling and write-back at the acquisition end. If the camera detects strong shaking or overexposure, the wS and wV values ​​are immediately reduced within that time window, and a "image quality anomaly" label is written into the fusion record. When the radar merges multiple targets, the size R is marked as "unreliable." The system prefers to mark the window as pending rather than rashly aligning it to the wrong image object. These strategies resolve the contradiction of "unreliable strong constraints and easy drift in license plate vehicle reconstruction," allowing alignment to be based on physical intuition: the continuity of size and speed, the stability of lane constraints, and the relative stability of visual details.

[0049] The output of this embodiment is a time-series entry temporarily assigned by vehicle ID. Each entry stores a fusion vector F, a unified timestamp, spatial coordinates, a list of device sources, and a weight log. The emphasis on the weight log is to provide interpretable context for subsequent steps such as similarity matching and constraint filtering: if a trajectory primarily relies on radar and depth, its extension in nighttime road sections is reasonable; if it primarily relies on structured color, breaks at points of abrupt changes in illumination are understandable. Overall, S101 consolidates scattered multi-source observations into a searchable and verifiable feature fusion dataset, laying a solid foundation for subsequent initial screening, fine matching, and spatiotemporal constraints.

[0050] Step S102: Perform preliminary screening of vehicles in video data based on structured features, calculate cosine similarity of the depth visual feature vectors for precise matching, read road topology data from electronic map, calculate travel time constraints and path accessibility constraints of adjacent capture points, use the travel time constraints to remove trajectory points that violate driving speed limits, filter trajectory points that cross physical obstacles based on the path accessibility constraints, dynamically adjust spatiotemporal thresholds according to road type, and fit missing road segments using road connectivity relationships;

[0051] Optionally, this embodiment focuses on S102, organizing the fragmented appearances of unlicensed vehicles in multi-camera videos into a continuous trajectory verified by rules. The scenario is a corridor where urban expressways and main roads intersect, with varying camera capture frequencies and frequent changes in weather and lighting. The starting point is the initial screening of structured features: the input consists of vehicle type, color, number of axles, and sunroof identification already aligned to the timeline. This step does not aim for a perfect hit, but rather clusters similar-looking targets within the same time window and adjacent spatial range into candidate clusters. This embodiment uses vehicle type and number of axles as strong constraints and color and sunroof as weak constraints because color drift is significant between cameras at different intersections, but vehicle outlines are more stable after scale normalization. The initial screening outputs several candidate trajectory segments, each with a camera ID, timestamp, and spatial coordinates, which then proceed to the subsequent fine-tuning process.

[0052] This embodiment performs deep, precise matching within the candidate dataset. We take the depth visual feature vector of each captured target, calculate the cosine similarity for each pair, obtain a similarity matrix, and set segmented thresholds: a high threshold for nearby camera pairs and a low threshold for cross-region camera pairs to accommodate differences in viewpoint. The discriminative power of the depth vectors comes from cross-domain training. During training, the convolutional network is constrained to focus on details related to identity but not easily affected by lighting, such as bumper texture, grille shape, and window pillar lines, which aligns with the application scenario. To avoid the spread of single-shot misjudgments, similarity is smoothed over time using a sliding window. Only when the average similarity of multiple consecutive frames of the same vehicle pair stabilizes above the threshold are the two segments connected. Similarity only indicates "likeness"; the next step is to ask "whether it's accessible."

[0053] This embodiment introduces road topology and traffic constraints. We read the node-edge structure of the electronic map, where edges include road type, direction, speed limit, and physical obstacle labels. For any adjacent capture points i→j, we calculate the feasible shortest path and its theoretical minimum travel time, with the following criterion: Δt ≥ dmin / vmax, where Δt is the time interval between two captures, dmin is the length of the shortest path from i to j in the road network, and vmax is the maximum allowable speed limit for that path. The physical meaning of this inequality is intuitive: if the lower limit of the time taken for the shortest reachable distance between cameras exceeds the observed time difference, then the pairing is unreliable. Furthermore, if the shortest reachable path does not exist or requires crossing rivers, medians, or enclosed building areas, the path reachability is directly judged as negative, even if the similarity is high. This superposition of the two constraints of "similarity" and "possibility of reach" conforms to the natural laws of real road travel.

[0054] This embodiment requires dynamic spatiotemporal thresholds under complex road conditions. Vehicle speed distribution on highways is high and stable, while urban arterial roads are greatly affected by traffic lights, resulting in wider time interval fluctuations. We provide different Δt tolerance ranges based on road type, with tighter thresholds on highways and wider thresholds on arterial roads, and introduce time-of-day factors: the upper limit of reachable time for the same road segment is increased during morning and evening rush hours. Threshold adjustments are not arbitrary but based on historical quantile intervals derived from sample statistics. The statistics match the actual vehicle speed distribution, neither excessively restricting real vehicles nor allowing unreasonable splicing. This approach is crucial for vehicle tracking at intersections, especially at underpass entrances and exits, where visibility gaps between capture points rely more heavily on time constraints.

[0055] In this embodiment, to address missing or occluded captures, trajectory gaps are also filled through fitting. We construct a candidate path set for each breakpoint pair on the road network topology, using the shortest path as the initial candidate, and then weighting and sorting them according to road level and traffic rules. Weighting factors include road level preference, U-turn and left-turn restrictions, and construction / closure information to avoid fitting paths that violate the rules. If multiple closely reachable paths exist, we check the "unassigned captures" from surrounding cameras as corroborating evidence; the path with more time-consistent unassigned segments receives a higher score. The completed output is not a hard confirmation, but rather annotates the confidence level and awaits approval from the S103 human-machine verification interface, ensuring that the trajectory reconstruction is both coherent and allows for further investigation.

[0056] This embodiment strings the above steps together into a refined matching pipeline and provides two on-site segments for illustration. Segment 1: Rainy night, color features are grayish. Initial screening only retains candidates with the same vehicle type and axle. The depth similarity reaches a high threshold between two adjacent cameras, but the electronic map shows that the two points are separated by central greenery and there is no U-turn, so the accessibility is false. This stitching is discarded, and the trajectory is matched with the next camera point instead. Segment 2: Urban expressway exit, the time interval between capturing A and B is short, and the shortest path length is relatively long. If the main road threshold is used, it may be misjudged. However, we identify the road type as an expressway, with a higher vmax. After lowering the Δt threshold, the pairing is successful. Then there is a missing capture in a tunnel. The system fits the exit C with road network connectivity and finally generates a continuous trajectory and marks it as "mid-section fitting".

[0057] This embodiment incorporates two additional engineering safeguards. First, for vehicles with modified exteriors (such as roof racks or window tinting), depth features may shift. Therefore, we increase the weight of "sunroof markings" and "axle count" in the structured features, and only allow low-similarity stitching on road sections with stricter time constraints, reducing false merging. Second, to address camera positioning errors, we add a confidence radius to the capture point coordinates, and use buffer zone entry nodes instead of precise points in path calculation to prevent false "unreachable" judgments due to deviations of several meters. These seemingly trivial compromises reflect a deep respect for the actual conditions of the roads and equipment.

[0058] The final output of this embodiment is a trajectory sequence that has been filtered to remove speed violations and obstacles, dynamically thresholded, and fitted with road connectivity where necessary. It includes the basis for each stitch: similarity summary, time constraint criteria, reachability results, and candidate path weights. Subsequent batch verification in S103 can quickly locate problematic segments based on this, requiring only visual verification of key areas. This chain combines visual "similarity," map "accessibility," and temporal "timeliness," solving the long-standing problems of unlicensed vehicles easily getting mixed up and the chain breaking under sparse capture and complex road networks. The reliability and interpretability of the trajectory reconstruction better meet the expectations of public security and road administration scenarios.

[0059] Step S103: Input the filtered vehicle trajectory sequence into the batch verification interface, dynamically adjust the similarity threshold and matching weight based on human feedback, read the camera location information to construct a feature reliability scoring matrix, write the feature reliability score into the rule engine, adaptively adjust the feature matching priority under different scenarios, filter trajectory segments that violate traffic rules using road traffic rules, generate a trajectory repair scheme based on road connectivity constraints, and construct a trajectory restoration database containing vehicle features, trajectory segments, and verification records.

[0060] Optionally, this embodiment follows the output of S102, sending the vehicle trajectory sequence, which has already been filtered through spatiotemporal and accessibility constraints, into the batch verification interface. Then, human experience is incorporated into the model parameters and rule engine, and finally, it is solidified into a traceable database. Let's first discuss the input and output of batch verification. The input consists of several candidate trajectories, each composed of a segment sequence, splicing criteria (similarity, time constraint criterion, accessibility results), and fusion feature weight logs. Batch verification unfolds within a scrollable time window. The interface displays trajectories with similar structured features overlaid within the same time period, marking "weak evidence splicing" points. The operator simply selects "keep / split / replace," and the system instantly records mismatched and missing tags, feeding them back to the threshold and weight updater. The key here is the division of labor between humans and machines: humans do not perform frame-by-frame judgments, but only express preferences at the most error-prone connection points, allowing the algorithm to adjust parameters.

[0061] This embodiment translates human feedback into calculable correction values. For splicing pairs marked as mismatched, we review their depth similarity s, structured consistency c, time margin m, and reachability Boolean value r to estimate which type of evidence is more distorted in that scenario. The updater aggregates statistics by camera pair and time period, outputting a set of correction coefficients Δθ for fine-tuning downstream thresholds and weights. To make it more intuitive for readers, we express the threshold update using a simplified relation: τnew = τold + α·Δθ, where τ represents the currently used similarity threshold, α is the learning rate, and Δθ comes from the intensity of the human feedback's indication of "too loose / too tight". The physical meanings of τ, α, and Δθ here are respectively "the minimum similarity requirement for determining whether they are the same vehicle", "the magnitude of the impact of a single feedback on the threshold", and "the tightening or loosening trend indicated by the feedback". This update conforms to common sense: in night and rain, depth features are stable, so the threshold can be slightly loosened; in backlight with strong reflection, it should be tightened to avoid mismatching vehicles with film applied.

[0062] This embodiment reads camera location information and historical recognition results to construct a feature reliability scoring matrix for each location and location pair. The rows of the matrix represent feature types (vehicle type, color, axle, sunroof, depth vector, radar size, ETC identity), the columns represent location IDs or location pairs (A→B), and the cell value is the historical stability score of that feature at that location or across locations. The statistical criteria are divided into weather and time-segmented data. The scoring is not mysterious; it originates from natural statistics in stages S101 and S102: color scores are lower when drifting is frequent at tunnel entrances, depth vector scores are slightly lower at checkpoints with dense traffic of adjacent vehicles of the same brand, and radar size scores are also lowered due to the tendency for confusion during congested periods. After the scoring matrix is ​​written into the rule engine, the system sorts the matching priorities by reliability. For example, for the A→B camera pair, if the color score is below the threshold, vehicle type + axle + depth become the main evidence chain, while color is only used as a reference.

[0063] This embodiment also introduces road traffic rules to filter trajectory segments for compliance. The rule engine has built-in constraints from traffic management regulations, such as one-way traffic, no left turns, no U-turns, and tidal flow lane directions. Any turn in the trajectory sequence that violates the rules is marked as "to be repaired." Instead of immediate deletion, this segment is submitted to the repair engine: it searches the road network connectivity graph for alternative paths that satisfy the rules and are time-feasible. If multiple alternatives exist, it weights them based on road grade, historical traffic light delays, and camera coverage density, selecting the one with the highest score as the repair suggestion. Manual confirmation is then done via a single click on the batch verification interface before the repair is finalized. This in-and-out process adheres to the rules while retaining sufficient flexibility for on-site situations. For example, during temporary traffic control periods, the rule engine loads event packages to relax traffic restrictions for specific time periods.

[0064] This embodiment presents two specific scenarios regarding weight adaptation. On elevated curved sections, vehicle exterior textures are compressed and blurred on distant cameras, causing fluctuations in depth similarity. The scoring matrix lowers the depth weight and prioritizes wheel axle and radar dimensions, so the stitching no longer overly relies on texture. At interchange ramp merging points, clusters of similar white SUVs increase structural ambiguity. The system incorporates the time margin m of adjacent segments with higher weights, accepting only combinations that satisfy "sufficient time," and guides the manual review interface to present only two or three most likely competitors, reducing the burden on the human eye.

[0065] This embodiment ultimately embeds everything into data assets. We construct a trajectory reconstruction database, divided into three main tables and several sub-tables: the feature table stores structured features, depth vectors, and multimodal supplements (radar size, ETC identity); the trajectory table stores coordinate sequences, time series, road segment numbers, and summaries of the splicing basis; the operation record table records manual verification records, threshold and weight version numbers, repair actions, and reasons. The three tables are linked by trajectory IDs and segment IDs. Any modification will generate a new version entry instead of overwriting the original record. The version chain can be traced back to specific manual operations and rule snapshots, ensuring interpretability and auditability requirements.

[0066] This embodiment also considers boundaries and anomalies. In the event of system-level clock drift, the batch verification entry will lower the confidence level of the time constraint for that period and force manual confirmation. For vehicles without license plates and without ETC readings, the identity dimension is empty, but this itself forms a "null value mode," which can actually differentiate signals on certain road sections; the rule engine will have targeted priority configurations. We will also avoid infinitely amplifying the "human-controlled" approach; human bias will be dispersed: the updater uses a moderate learning rate α and only applies significant adjustments to statistically recurring patterns, avoiding one-time bias from skewing the system.

[0067] The practical effects of this embodiment are not described with exaggerated figures, but rather through tangible experience: when law enforcement agencies review a suspected unlicensed vehicle, they can open a trajectory and see why each connection point was established or corrected—which features are more credible, whether there was sufficient time, and whether traffic rules were followed. The database is not a cold, impersonal warehouse; it preserves the chain of evidence and traces of human decision-making, providing a basis for subsequent model training and rule maintenance, forming a closed loop from trajectory generation to compliance correction and then to experience feedback.

[0068] As described above, the method for reconstructing the trajectory of unlicensed vehicles provided in this application can achieve accurate vehicle identification through innovative design of a multimodal feature fusion model, feature extraction, and data alignment. A trajectory screening system is constructed, combining spatiotemporal constraints and path verification to establish a reliable trajectory reconstruction mechanism. Verification optimization is introduced, ensuring the accuracy of the reconstruction results through reliability assessment and rule adjustment. This method effectively addresses the shortcomings of traditional technologies in feature fusion, trajectory screening, and verification optimization, providing technical support for the management of unlicensed vehicles.

[0069] In one embodiment of the method for reconstructing the trajectory of an unlicensed vehicle in this application, it may further include the following:

[0070] Step S201: Collect vehicle images from road monitoring cameras, read the latitude and longitude coordinates and timestamp information when the images are collected, preprocess the vehicle images, perform image enhancement and size normalization, read the vehicle type classification dictionary and vehicle color dictionary, input the normalized image into the feature extraction model, identify the vehicle outline based on the edge detection algorithm, and extract the structured features of vehicle type, color, number of axles and sunroof markings based on the region segmentation algorithm;

[0071] Step S202: Construct a convolutional neural network including an encoding layer, a pooling layer, and a fully connected layer. Input the normalized image into the encoding layer to extract local feature maps. Input the feature maps into the pooling layer to reduce the data dimensionality. Input the dimensionality-reduced features into the fully connected layer to generate a deep visual feature vector. Perform normalization calculation on the deep visual feature vector to generate a feature representation of the vehicle image.

[0072] Optionally, this embodiment focuses on S201 and S202, with the scenario set in a composite road section containing both elevated roads and ground-level auxiliary roads. The system captures vehicle images from checkpoints and roadside PTZ cameras, along with latitude and longitude coordinates and acquisition timestamps. To avoid clock drift affecting subsequent alignment, we first correct the timestamps using offset records from roadside PTP time synchronization, and convert the latitude and longitude of the camera mounting points to local plane coordinates, preserving the camera orientation and calibration extrinsic parameters. This allows us to roughly correlate the pixel dimensions with the actual scale. The quality of the original image is not stable; rain, backlighting, and nighttime lighting all introduce noise. This embodiment selectively performs dehazing (dark channel prior), color shift correction (grayscale world hypothesis and linear color mapping), highlight suppression, and detail enhancement based on frame-level quality assessment results. Finally, the image is uniformly scaled to the training distribution, and mailbox padding is used to ensure the aspect ratio is not shrunk. Size normalization is not for aesthetic purposes, but to ensure that subsequent edge and segmentation networks operate at similar pixel densities, reducing threshold drift caused by scale.

[0073] This embodiment reads a vehicle classification dictionary and a color dictionary. These dictionaries are not static; they are subdivided according to common regional vehicle types, with sedans further subdivided into three-box and two-box, and SUVs into compact and mid-size. The color dictionary has separate mappings for daytime and nighttime use to reduce the bias of mistaking silver for beige in warm-light conditions. The normalized image enters the feature extraction model, which consists of two parallel branches: the edge detection branch starts with Canny and then uses a learned boundary refinement network to suppress background texture, outputting the vehicle's outline and key polylines (bumper, roofline, wheel arches); the region segmentation branch uses a lightweight semantic segmentation network to segment out four main regions: the body, windows, wheels, and roof. The calculation of structured features closely follows these outputs: the vehicle model is determined by a combination of the aspect ratio of the outline, the geometric relationship between the front and rear of the vehicle, and the corner point pattern of the headlights; the color is obtained by voting using the dominant color histogram and stable color clusters after occlusion removal in the body area; the number of axles is not directly counted as circles, but rather the connected components of the wheel area are rhythmically verified in the direction of travel to reduce false detections of shadows and ground reflections; the sunroof indicator is found in the roof area as a closed patch with low reflection and regular edges, combined with consistency confirmation with adjacent frames. The final output is solidified into structured fields, including category code, color code, number of axles, sunroof Boolean bits, and confidence scores for each item, for subsequent retrieval.

[0074] This embodiment focuses on explaining the causal relationship in the combination of segmentation and edge mapping. A stable vehicle outline provides the vehicle's structural framework even when obscured; semantic segmentation excels at identifying "where the body / windows are." The combination of these two elements is crucial to separating a "black-like" shadow from a "true black paint" appearance. For example, on a rainy night, road reflections brighten the lower edge of the car's side, which might be mistaken for a lighter color if viewed solely by appearance. We constrain the color sampling area towards the center of the vehicle by using the position of the wheel arch curve to maintain stability. This approach of mutually correcting geometry and semantics is key to reducing confusion caused by similar appearances in scenarios involving vehicles without license plates.

[0075] This embodiment proceeds to S202, where a convolutional neural network is constructed to generate deep visual features. The network includes an encoding layer, a pooling layer, and a fully connected layer. The encoding layer extracts multi-scale textures and shape edges from stacked residual blocks. Early layers perceive stripe and dot features, while deeper layers focus on details sensitive to distinguishing the interior of a vehicle, such as grille style, headlight cavity structure, and window pillar proportions. The pooling layer uses spatial pyramid pooling to map inputs at different cropping scales to fixed-length representations, avoiding information loss due to differences in the field of view of different cameras. The fully connected layer outputs a fixed-length embedding. During training, a metric learning loss is used to bring "different frames of the same car" closer together and "similar appearances of different cars" more distinct. The input is the aforementioned normalized image and its enhanced flipped sample, and the output is a deep visual feature vector.

[0076] This embodiment performs L2 normalization on the output embedding to obtain a directional representation. The purpose is to decouple similarity calculation from brightness scaling, making subsequent matching using cosine similarity more stable. The effectiveness of the convolutional network depends on the natural association between the input information and the target: the texture and component details we are interested in are most varied at the front and rear of the vehicle, and best able to maintain "self-identity" across lighting conditions. In training and online inference, the model input / output is aligned with the S101 fusion process. The input is a normalized image, and the output is a fixed-length feature representation, which is bound to the structured fields by timestamps when stored in the database.

[0077] This embodiment uses two specific segments to illustrate how the connection between S201 and S202 maintains stability in complex environments. The first segment is a sunset backlight on an elevated ramp, where the structured color fluctuates between "dark gray / black." The edge branch captures the clear outline of the taillights and the rear bumper's crease, and the vehicle is still classified as a "compact SUV." The depth vector maintains high similarity between two adjacent checkpoints, and the two segments are successfully merged in the subsequent screening stage. The second segment is the yellow sodium lamp area inside the tunnel, where the sunroof boundary almost disappears. The segmentation branch gives "uncertainty," and the system marks the sunroof position as low confidence to avoid using it as a hard constraint. However, the depth vector still captures the shape of the A-pillar and the air intake grille, and the matching is not significantly affected.

[0078] This embodiment also considers the boundary conditions of the device and the environment. Slight camera defocusing can cause the loss of high-frequency textures. We add anti-aliasing pre-filtering before the encoding layer to preserve as many low-frequency shapes as possible that can be distinguished. For vehicles with window tinting or color changes, the structured color and depth texture will shift simultaneously. The recording layer will reduce the reliability of the features for that time period and leave it to subsequent spatiotemporal and accessibility constraints to prevent a single visual feature from determining success or failure. The final output is a set of structured features and depth visual embeddings with source time and spatial indexes. These can be used independently or in subsequent steps to be temporally aligned with radar size and ETC identity to form a complete feature fusion dataset. This chain clarifies "what can be seen" and "what should be captured," providing a solid basis for subsequent trajectory reconstruction.

[0079] In one embodiment of the method for reconstructing the trajectory of an unlicensed vehicle in this application, it may further include the following:

[0080] Step S301: Read the roadside millimeter-wave radar data stream, parse the radar data stream into target detection data, extract vehicle length, width and height dimension parameters and motion state parameters, read vehicle passage data collected by the electronic toll collection system, parse electronic tag information to obtain vehicle identity features, group the vehicle dimension parameters and identity features according to the collection location and lane number, and construct a multi-source data correspondence table.

[0081] Step S302: Read the structured features and depth visual feature vectors, sort the structured features, depth visual feature vectors, vehicle size parameters, and identity features according to timestamp information, calculate the temporal correspondence between features based on a sliding time window, perform interpolation alignment on the feature data, combine the aligned multimodal features to generate a fusion feature vector, and construct a feature fusion dataset.

[0082] Optionally, in the implementation process of S301-S302, this embodiment stitches together the "hard physical" information of radar and ETC with the "visual semantics" of the video side to ultimately form a searchable feature fusion dataset. First, consider S301: the roadside millimeter-wave radar is deployed along the lanes and continuously outputs point traces and target tracking streams. In this embodiment, at the edge side, a combination of Kalman filtering and DBSCAN is used to cluster the scattered points into target trajectories. Then, a constant false alarm threshold is used for measurement correlation to obtain target detection data at each moment. For each radar trajectory, we estimate the three-dimensional dimensions of the vehicle: length, width, and height. The length is derived from the radar scattering envelope along the target's direction of travel; the width is obtained from the lateral scattering span and the geometric projection of the installation height; and the height is distinguished by Doppler sweep angle and RCS energy distribution to differentiate between types such as vans and sedans. Motion state parameters include instantaneous velocity, longitudinal acceleration, heading angle, and heading variability, used to identify lane changes and deceleration. State continuity is an important and reliable clue for subsequent time alignment.

[0083] This embodiment reads ETC gantry traffic data in parallel. The original record includes gantry ID, lane number, timestamp, electronic tag ID, and vehicle classification within the toll system. The electronic tag ID is hashed and anonymized to form an identity feature. Considering scenarios where the data is unread due to no license plate or obstruction, we retain the "unread / abnormal verification" status code as a value for the identity feature field to avoid treating missing data as blank. Subsequently, the radar size parameters and ETC identity features are grouped according to "collection location + lane number" to construct a multi-source data correspondence table. The purpose of grouping is twofold: first, gantries and radars are usually on the same pole or cross section, so lane numbers are naturally aligned; second, during peak periods with multiple parallel lanes, lane restrictions can significantly compress the size of potential matching pairs, making the association more stable in subsequent time windows. If there is an inconsistency between the lane markings of the gantry and radar, this embodiment checks based on spatial proximity and speed direction, prioritizing the retention of the radar-side lane judgment and marking it with "lane conflict" in the table for subsequent weight evaluation reference.

[0084] This embodiment converges the correspondence table generated by S301 with the video structured features and depth vectors output by S101 into the alignment link of S302. The first step is sorting and cleaning: all feature entries are sorted according to a unified PTP reference time, and records with abnormal timestamp jumps and duplicate writing are filtered out. The second step is to calculate the temporal correspondence based on a sliding time window. The window width is set according to the road type and the capture / gantry sampling frequency. Within the window, we calculate three types of consistency scores: visual consistency (whether the vehicle type, axle, and sunroof are stable and whether the depth vector similarity crosses the threshold), physical consistency (whether the radar speed extrapolated to the next capture time can reach the camera position, and whether the heading matches), and identity consistency (whether the same tag ID or the continuity of the "unread" mode appears in the same window). The three scores are not simply added together, but are fed into a lightweight discriminative model to output the association probability. This model is trained offline using labeled same / different vehicle pairs. The ranking of feature weights is consistent with real-world patterns: in tunnels and rainy night conditions, physical consistency and depth vector weights are higher; in smooth-traffic areas during the day, structured color and vehicle type are more distinguishable.

[0085] This embodiment faces the problem of inconsistent data granularity during the alignment process. Video is captured event-driven, radar is measured continuously at high frequencies, and ETC involves sparse gantry crossings. We establish a timeline for each candidate vehicle chain and perform constrained interpolation on missing measurement points: position is calculated using physical extrapolation based on velocity and heading; depth vectors are calculated using linear interpolation of the two nearest captures with a "virtual frame" identifier; size parameters are maintained at the median within a stable heading range; and identity features are filled with the most recent valid values ​​without crossing gantry sections. Interpolation is not intended to create non-existent information, but rather to provide continuous input for subsequent trajectory fitting and similarity calculations. All interpolated samples carry confidence weights in the fusion vector to avoid excessive influence on decision-making.

[0086] This embodiment assembles multimodal features into a single-step fusion vector and constructs a feature fusion dataset at the session level. For ease of understanding, we provide a weighted expression for the fusion, but only retain one necessary formula to clarify the meaning of the variables: F_t = [wS(t)·S_t, wV(t)·V_t, wR(t)·R_t, wE(t)·E_t], where t is a time on a unified time axis, S_t is a structured feature vector (encoding of vehicle type, color, axle, and sunroof), V_t is a depth visual feature vector, R_t is a vector concatenated with radar size and motion state, and E_t is an embedded representation of ETC identity features; wS(t), wV(t), wR(t), and wE(t) are the credibility weights of the four types of features at time t, derived from the evaluation of device health, self-stability, and scene labels (weather, time period, and road type). The connotation of weight is "the contribution of the feature to the judgment of the same vehicle in the current context". When it rains and noise is reduced, wV and wR are higher. When there is interference from colored light boxes, wS decreases. When the gantry is empty, wE is set to zero and maintains the continuity of identity in the vector as a placeholder.

[0087] This example provides two real-world scenarios to illustrate the rationale behind the above logic. On an elevated highway during a light nighttime drizzle, the video shows color drift and repeated shifts in structured colors. The radar speed is stable, the dimensions are consistent, and the depth vector is also stable. The sliding window discrimination model uses physical and depth data as primary evidence. Even with the window spanning two camera locations, it still provides a high correlation probability. Interpolation only fills in virtual frames in the middle, resulting in continuous fused data. During the day, under an overpass, the gantry and checkpoint are close together, and traffic is dense. Vehicles of the same color and type are seen driving side-by-side. While the depth similarity is high for both vehicles, physical consistency eliminates one possibility: based on the radar heading variability and lane number, the other pair cannot merge across lanes within Δt, lowering the correlation probability and leading to abandonment of the pairing.

[0088] This embodiment also handles anomalies and boundaries. When camera shake causes depth vector drift, the device monitoring module lowers the wV value of that segment and prompts for manual annotation; when radar shows target overlap, the reliability of R_t decreases, so we extend the sliding window and wait for the next checkpoint to capture and re-certify; when temporary ETC maintenance results in a full hour of "unread" data, E_t does not participate in the decision-making process, but the "continuously unread" pattern actually forms a recognition feature, which will be used as a difference signal in S102 later. All these states are written into the meta-fields of the multi-source correspondence table and the fused dataset to ensure that subsequent filtering, matching, and repair can reuse the context.

[0089] The output of this embodiment is a feature fusion dataset organized using temporary vehicle session IDs, containing time series F_t, alignment quality metrics, interpolation markers, and a list of data sources. Its core problem is to place sensor observations with different sampling frequencies and reliability within the same spatiotemporal framework and provide a credibility profile that conforms to natural driving patterns. Subsequently, when performing similarity matching and spatiotemporal constraints, S102 no longer needs to repeatedly process cross-source alignment details; it can directly consume this fused data, reducing engineering complexity and making the interpretation chain for human-machine verification smoother.

[0090] In one embodiment of the method for reconstructing the trajectory of an unlicensed vehicle in this application, it may further include the following:

[0091] Step S401: Read the structured features in the feature fusion dataset, perform a coarse selection of vehicles in the video data based on vehicle type classification and color matching, calculate the cosine similarity of each pair of deep visual feature vectors in the coarse selection results, construct a candidate vehicle similarity matrix, set a similarity threshold to filter high-confidence matching pairs, write the spatiotemporal information of the high-confidence matching pairs into the trajectory data table, and construct a vehicle motion trajectory sequence based on the trajectory data table;

[0092] Step S402: Read the road topology data of the electronic map, extract the road connection relationship and traffic rules, calculate the shortest path distance between adjacent capture points, calculate the theoretical minimum passage time based on the road speed limit standard, set the theoretical minimum passage time as the lower limit of the time constraint, compare the lower limit of the constraint with the time interval in the trajectory data table, and remove trajectory points whose time interval is less than the lower limit of the constraint.

[0093] Optionally, this embodiment builds upon the feature fusion dataset generated in S101-S202, performing a three-step process in S401: "coarse selection—fine matching—table placement." Each captured image contains structured features (vehicle type, color, number of axles, sunroof identifier, and confidence level), depth visual vector, unified timestamp, and spatial coordinates. In the coarse selection stage, we use vehicle type classification as the strong key and color as the secondary key, clustering targets within the same time window and adjacent camera coverage areas into candidate buckets. The reasoning is straightforward: vehicle type and axle count determine the overall shape and have high stability after scale normalization, while color is more affected by lighting disturbances across locations and is only used to reduce the search radius rather than directly determining identity. For colors with low confidence or boundary conditions (such as yellow light at tunnel entrances), the color key is automatically downweighted to avoid grouping vehicles with similar appearances into different buckets.

[0094] In this embodiment, an exact match of depth vectors is initiated within the candidate buckets. The cosine similarity is calculated pairwise for the depth visual feature vectors of each target to construct a candidate vehicle similarity matrix, and piecewise thresholds are set according to the geometric relationship of camera pairs: the thresholds for adjacent camera pairs with similar viewing angles are relatively high, and the thresholds for long-distance camera pairs across intersections are appropriately lowered to tolerate the scale difference. The similarity is not determined by a single frame. We introduce a sliding time window to calculate the steady-state mean of the cross-frame similarity of the same vehicle. Only the matching pairs that satisfy the threshold for several consecutive frames are regarded as "high-confidence". The internal logic of this criterion is consistent with the scenario: it is difficult for the details of the same vehicle to change qualitatively in a short period of time, and the directional features of the texture and component layout will remain close in the embedding space, and accidental noise should not change the conclusion.

[0095] In this embodiment, the spatio-temporal information of high-confidence matching pairs is written into the trajectory data table. Each written item includes the start and end capture IDs, time interval, spatial coordinates, camera pair, and a summary of matching evidence (similarity mean, variance, threshold version used, color / model consistency identifier). To elevate from matching pairs to trajectories, we regard the matches as edges and the captures as nodes, perform a topological sort in time, and search for the longest path sets that do not intersect or weakly intersect to form a sequence of vehicle motion trajectories. When a fork occurs, the main chain is preferentially retained according to the strength of evidence and time continuity, and the branches are temporarily stored in a candidate state. After the spatio-temporal constraint check in S402, it is decided whether to incorporate or remove them.

[0096] In this embodiment, it enters S402, attaching the hard constraints of the map to the soft evidence of vision. The system reads the road topology of the electronic map and extracts the connection relationships and traffic attributes (one-way, left turn prohibited, road type, speed limit) of nodes and edges. For any adjacent capture points i→j in the trajectory data table, the shortest reachable path length dmin on the road network is calculated, and the theoretical minimum travel time tmin is estimated based on this. The criterion for time consistency follows common sense: a vehicle cannot "teleport" at a speed exceeding the speed limit. Therefore, we compare the observed time difference Δt with tmin. If Δt < tmin, this trajectory point pair is marked as violating the time constraint and is removed; if multiple segments are violated in a chain, the entire segment is removed from the main chain and falls into the repair pool to avoid error accumulation. Here, tmin is not arbitrary; it comes from the multiplication and division relationship between the road speed limit and the shortest path length, implicitly representing the physical lower limit of the fastest reachable speed of the vehicle within the rules.

[0097] This embodiment further refines the engineering details to ensure that the time constraints are neither too strict nor too lenient. First, the shortest path calculation considers road type and temporary event packets. For example, if construction closures render the conventional shortest path unreachable, the system automatically selects a suboptimal reachable path and adjusts the dmin accordingly, thereby increasing the tmin and avoiding misjudgments. Second, recognizing meter-level errors in camera positioning, we introduce a buffer radius to snap the pixel projection point to the centerline of the nearest road segment before calculating dmin, avoiding the illusion of "too short a path, too small a tmin" caused by geometric jitter. These two steps bring the time constraints closer to the actual traffic conditions.

[0098] This embodiment provides two segments to illustrate how the combined force of S401-S402 works. On a weekday morning, a white hatchback appears alternately on elevated roads and ground-level auxiliary roads. A coarse selection clusters images with colors matching the car model. If the depth similarity exceeds a high threshold between two adjacent checkpoints, an edge is written. Subsequently, a pair of far-distance matches appears, with similarity barely reaching a low threshold. However, the map calculates that dmin is large and tmin is much larger than Δt. This pair is discarded by S402, keeping the main chain clean. The other segment is a rainy night. Color confidence is low and is downweighted. Two segments are created by splicing car model and depth. Due to construction closing the regular path, the system switches to auxiliary roads. tmin is adjusted upwards according to the new path, establishing the relationship between Δt and tmin. The trajectory continues across points, and the reason for the "event package taking effect" can be reviewed in the evidence summary.

[0099] This embodiment also handles boundary and uncertain states. When encountering "appearance clones" under dense traffic flow, the depth matrix may exhibit multiple approximate maximum values. We break this stalemate by using temporal continuity and consistency of the preceding and following velocity vectors. For gaps caused by missing captures, the ends of high-confidence edges are preserved, but the middle section is written as a "missing placeholder" and submitted for subsequent connectivity fitting or manual verification. For rejected time-violation points, we do not discard all evidence directly, but retain their depth similarity and structured fields in the candidate pool. If a midpoint is captured later and meets the tmin condition, the original evidence can be revived. The entire process is traceable within the version chain of the trajectory data table.

[0100] The output of this embodiment is a sequence of vehicle motion trajectories generated through coarse selection and deep refinement, and then purified by time lower limit constraints. Each edge and node is labeled with sufficient evidence. The next step of human-machine verification and rule adaptation can then revolve around these "interpretable breakpoints," rather than blindly searching through the entire video. For case handling and road administration, what they see is not just "connected," but also "why connected and why dismantled." This interpretability comes precisely from the logical chain of S401 and S402 that intertwines visual similarity with the physical constraints of the road network.

[0101] In one embodiment of the method for reconstructing the trajectory of an unlicensed vehicle in this application, it may further include the following:

[0102] Step S501: Read the physical obstacle data in the electronic map, extract the spatial distribution information of the median strip, river and building, calculate the set of reachable paths between trajectory points based on the road topology relationship, match the set of reachable paths with the trajectory sequence, identify the trajectory segments that cross physical obstacles, mark the trajectory segments as invalid trajectories and delete them from the trajectory data table, and filter abnormal trajectory points based on path reachability constraints.

[0103] Step S502: Read road attribute information, classify road types into expressways, urban expressways, and urban arterial roads, calculate the maximum allowable speed based on road capacity, write the maximum allowable speed into the spatiotemporal constraint parameter table, dynamically set the trajectory point time interval threshold according to road type, read road connectivity relationships to construct a road network topology map, generate candidate paths for missing road segments based on the shortest path algorithm, sort the candidate paths by road level with weights, and select the path with the highest weighted score to fit and complete the trajectory.

[0104] Optionally, this embodiment focuses on S501-S502, placing the trajectories that have passed the initial screening and time limit test under the constraints of the physical world to identify and fill in any gaps. We first import obstacle data directly related to vehicle traffic from the electronic map, including the geometric segments of the central median, polygons of rivers, boundaries of buildings and enclosed areas, and the traffic attributes of bridges, tunnels, and U-turns. Obstacles and roads are not isolated layers; this embodiment performs a topology cleaning during import: the intersection of the road centerline and obstacle boundaries is used to mark traversable structured elements (bridges, culverts) as "traversable," while the rest remain "blocked." This causal step is crucial; only by eliminating situations where "it appears to be blocked but actually has a bridge" can the accessibility calculation avoid mistakenly blocking normal traffic.

[0105] This embodiment calculates a set of reachable paths between trajectory point pairs. Specifically, it attaches the captured points to the nearest road centerline node and performs a directed shortest path search based on the road topology. However, instead of selecting a single path, it limits the maximum detour ratio and the upper limit of the number of nodes, enumerating several feasible paths to form a candidate set. Then, it compares the set with the trajectory sequence one by one: if any captured point pair has no solution in the candidate set, or can only be connected by crossing an insurmountable obstacle, it is determined that the segment crosses an obstacle. The obstacle-crossing segment is marked as an invalid trajectory and deleted from the trajectory data table. Simultaneously, the operation log is written with reasons such as "obstacle type, suspected camera mismatch / time drift," facilitating subsequent human-machine verification and tracing the root cause. Unlike the time constraint of S402, this emphasizes the hard boundary of spatial closure, which is logically more "absolute." Even with high depth similarity, it must obey the natural rule that "rivers cannot flow across."

[0106] This embodiment takes into account the imperfections of map and positioning, and introduces buffering and redundancy correction. Camera extrinsic parameter errors can cause the adsorption point to shift. We set a meter-level buffer zone outside the road centerline. Trajectory points within the buffer zone are still considered to be on the road. Obstacle boundaries are smoothed using morphological expansion and contraction to remove jagged small-scale pseudo-blockages. For linear obstacles (medians), we further combine the U-turn / crossing attributes of the road segment. If adjacent points are connected and require crossing the median but there is a legal U-turn nearby, the system will invalidate "direct crossing" but retain "detour" as a candidate for subsequent fitting. This is neither unfair nor overly lenient.

[0107] This embodiment proceeds to S502, focusing on fitting and completing the spatiotemporal thresholds and missing road segments. The road attribute file provides road types (highway, expressway, arterial road) and traffic capacity. Based on this, we calculate the maximum permissible speed and write it into the spatiotemporal constraint parameter table. The parameter table is not constant; it loads different versions according to time period and weather. The "maximum permissible" will be lowered during peak hours and rain / snowy periods. We set time interval thresholds for trajectory points according to type: tighter for high-speed roads and wider for arterial roads. These thresholds are then filled back into the validation field of the trajectory table, ensuring that every subsequent stitching is traceable. The significance of the threshold lies in answering "on this road and at this time, how far apart can two points still be connected by a single vehicle?" It represents the boundary of the feasible region for the next step of path fitting.

[0108] This embodiment constructs a topology graph using road network connectivity to generate candidate paths for missing segments in the trajectory. The algorithm employs a multi-constraint shortest path: the basic cost is geometric length, superimposed with road grade cost, turning cost, and historical delay cost. We calculate a comprehensive score for each candidate path: G = λ1·L + λ2·Cturn + λ3·Cdelay + λ4·Cgrade, where L is the path length, Cturn is the turning penalty (prohibited or discouraged turns are weighted higher), Cdelay is the delay estimate based on historical signals and traffic flow, Cgrade is the road grade cost (lower grades have higher costs), and λ is the scene weight, given by road type and time period configuration. The physical meaning of each parameter directly corresponds to the driver's actual choice preference: shorter routes with fewer illegal turns, less delay, and higher road grades are more "reasonable." We sort the routes in ascending order of score and select the lowest-scoring route as the fitted solution. If several solutions have similar scores, we search for supporting evidence of "unassigned captures," prioritizing those that exist.

[0109] This embodiment does not treat the fitting as a definitive conclusion, but instead writes the fitted segments into a trajectory table with a confidence level. The confidence level is based half on path scoring and half on spatiotemporal fit: the minimum travel time of the fitted path must fall within an upper and lower limit range (the upper limit comes from peak delay statistics), and the speed and heading at the beginning and end of the fitted segment should be compatible with the known segments before and after it. Candidates with incompatibilities (e.g., the end segment requires a sharp deceleration to make up the time) will be downgraded or discarded. Thus, the completion process not only considers the "shortest" path but also whether it is "runnable and whether the run is natural."

[0110] This embodiment uses two on-site segments to implement the above process. Segment one is a corridor where the river and the expressway run parallel. Visually, the two capture depths are very similar, but on the map, only a straight line across the river can connect them, and there is no bridge. S501 directly judges it as an obstacle crossing and deletes the edge, and the trajectory returns to the previous reliable point. S502 then searches for reachable candidates in the same side road network, and finally selects a path that detours along the auxiliary road and turns around at the next bridgehead. The confidence level is moderate and awaits batch verification and confirmation. Segment two is a main urban road under construction and closed. The original conventional passage is disconnected, the time interval is relatively large, the road capacity in the parameter table is reduced, the shortest path cuts to a back street, the turning penalty of the candidate is slightly higher, but the overall travel time is consistent with the field observation, the fit is valid, and the circumstantial camera provides a frame of "unassigned capture", so the confidence level is increased.

[0111] This embodiment also does not avoid boundaries. Outdated map versions can cause road segments that don't actually cross obstacles to be misjudged. After each obstacle crossing determination, we check the most recent map change records. If the area is within the change window, we directly mark that segment as "map uncertain" and submit it for manual review. For suction deviations caused by large-angle side shots from the camera, we use "multiple road projection candidates + shortest normal distance" to ensure stability, avoiding forcibly sucking points into the wrong lane. After two rounds of filtering and supplementation (S501 and S502), the trajectory table contains chains that can be validated by the physical world. Every deletion and supplementation has a source. The next step of human-machine verification and rule updates only requires looking at a few segments marked "obstacle crossing / fitting," naturally reducing the workload and providing frontline police officers with more confidence in the results.

[0112] In one embodiment of the method for reconstructing the trajectory of an unlicensed vehicle in this application, it may further include the following:

[0113] Step S601: Group the vehicle trajectory sequence by time window, read the feature matching results and constraint filtering results of each group of trajectories, construct a verification data table containing trajectory number, vehicle features, and matching confidence, input the verification data table into the batch verification interface, record manually confirmed mismatched trajectories and missed detection trajectories, calculate the feature similarity correction coefficient based on the manual feedback results, apply the correction coefficient to the dynamic adjustment of the similarity threshold, and perform adaptive update of the feature matching weight;

[0114] Step S602: Read the camera deployment location and coverage area, calculate the feature recognition accuracy at each location, calculate the feature reliability score under different lighting and weather conditions, write the feature reliability score into the scoring matrix, construct feature weight adaptive rules based on the scoring matrix, deploy the adaptive rules to the rule engine, and dynamically configure the feature matching strategy for different scenarios.

[0115] Optionally, this embodiment builds upon the trajectory sequence and evidence summary formed in S401-S502, implementing the human-machine closed loop of S601 and the scene adaptation of S602. First, the trajectories are grouped by time windows. The window size is not hard-coded but determined by road type and camera density. Short windows are used for expressways to facilitate timely correction, while relatively long windows are used for main roads to encompass traffic light cycles. Within each group, two types of information are read: one is the feature matching result, including structured field consistency, mean and variance of deep similarity, and cross-frame stability; the other is the constraint filtering result, including time lower limit criteria, accessibility and obstacle crossing conclusions, and the fitting and completion process. Based on this, we construct a verification data table, with the trajectory number as the primary key, accompanied by vehicle feature snapshots and segment-level matching confidence. This exposes the possible errors and the reasons for these connections. The batch verification interface only presents connection points whose scores fall within the gray area, allowing manual selection to retain, split, or replace them, and allows marking "missed detections requiring correction" time periods.

[0116] This embodiment translates manual feedback into corrections that can be digested by the parameter layer. For edges identified as mismatches, we backtrack their evidence vectors, decomposing them into depth similarity s, structured consistency c, time margin m, reachability identifier r, and camera pair context u (location pair, time period, weather). We aggregate and statistically analyze these within the same u to obtain the trend of "which type of evidence is prone to bias in this context." To apply this trend to the threshold and weights, we calculate the feature similarity correction coefficient Δθ(u) and update it online using a moderate learning rate after offline cold start to avoid excessive oscillations caused by a single feedback. Formally, this can be written as τnew(u) = τold(u) + α·Δθ(u), where τ is the similarity threshold in this context, α is the learning rate, and Δθ(u) represents the net indication strength of the manual feedback regarding "should be tightened / should be relaxed." At the same time, the matching weights wS and wV are also adaptively updated according to the sign and magnitude of Δθ(u): if mismatch is caused by color drift, wS is reduced and wV is increased; if mismatch is caused by texture blur, the opposite adjustment is made, and a one-time explanation is generated on the interface to record "why the change is made".

[0117] This embodiment does not take the shortcut of simply widening the threshold to handle missed detections. For manually marked missed detection segments, we first examine whether the constraints are too strict, such as tmin not being adjusted synchronously after the construction event takes effect, or adsorption error causing dmin to be too small. If the visual constraints are indeed too strict, then the threshold is slightly adjusted downwards, and verification through an A / B shadow queue is required in the next window of the same context u: the system reruns a shadow matching with the new threshold without affecting the mainstream waterline, and the threshold update is only "positive" if the mismatch rate does not worsen. In this way, the coupling between manual and parameter follows natural rules: the cause of the error must be correct, and the adjustment must be verified.

[0118] In this embodiment, we switch to step S602 to begin creating a reliability profile for the location and scene. We read the camera deployment location, orientation, focal length, and coverage area, link them to historical recognition records, and calculate the accuracy of each feature item at each location under different lighting and weather conditions. Specific dimensions include "daytime / nighttime / backlight" and "sunny / rainy / foggy / snowy." We also separately calculate the cross-domain stability of the location pair (A→B). Accuracy is not arbitrary; it comes from the refinement of the previous trajectory steps and the "truth values" verified manually: confirmed correct splicing accumulates positive samples, and incorrect splicing is accumulated as negative samples. We normalize these statistics to the [0,1] interval to obtain the feature reliability score, which is written into the scoring matrix R. The matrix element R(f,u) represents the reliability of feature f in context u.

[0119] This embodiment constructs adaptive feature weight rules based on the scoring matrix R. The rules are divided into two layers: the bottom layer is the weight mapping w(f,u)=g(R(f,u)), where g is a monotonic mapping and truncates extreme scores to avoid weight collapse; the upper layer is the strategy switching. When the R of certain key features in a certain context is lower than the threshold, the matching process switches to the "main evidence chain" dimensionality reduction strategy. For example, in the case of backlighting in a tunnel, color is relegated to a secondary reference, and depth + vehicle type + time margin become a hard constraint combination. After the rules are generated, they are deployed to the rule engine. For each candidate match processed by the execution end, R is queried according to its location pair and scene label, and w and strategy branches are calculated online without manual intervention. To avoid frequent jitter, the rolling update of R adopts a time decay window, and the influence of old data gradually weakens. When the weather changes suddenly, it only takes a few hours to reflect the weight, but it will not fluctuate drastically on a minute-by-minute basis.

[0120] This example provides two real-world scenarios to illustrate why this adaptive system is closer to reality. On a rainy night, an elevated camera at point A→B shows historical statistics indicating that the color R-value is low and the depth R-value is high for this pair. Therefore, the system automatically prioritizes depth and wheel axle, significantly reducing mismatches during batch verification. On a sunny, backlit main road from B→C, the depth occasionally becomes unstable due to directional reflections, causing the R-value to drop slightly. The system then prioritizes time margin and vehicle type, slightly tightening the threshold to prevent collisions when white SUVs are clustered together. In both scenarios, the behavior of the rule engine is "clearly explainable" to frontline colleagues, not a black box.

[0121] This embodiment records each adjustment of thresholds and weights, the version number of each scoring matrix, the effective time window, and the triggered summary of human feedback in the operation log. The trajectory reconstruction database thus has two additional tables: a parameter version table and a scene scoring table, which are respectively linked to the trajectory table and the feature table. Any future disputes can be reviewed to determine "why this threshold was set and what feedback and statistics were used as the basis." Regarding engineering boundaries, if human annotation is scarce in the short term, Δθ(u) is not updated; if there is a significant hardware change at a certain point (lens replacement), the system sets its R to "cold start," triggering a conservative strategy period where a human-machine hybrid approach is used to create a new profile.

[0122] In one embodiment of the method for reconstructing the trajectory of an unlicensed vehicle in this application, it may further include the following:

[0123] Step S701: Read the feature reliability scores in the rule engine, sort the vehicle features in descending order of reliability scores, dynamically allocate feature matching priorities based on the sorting results, read the road traffic management regulations, extract motor vehicle traffic rules and turning restriction rules, convert the traffic rules into trajectory constraints, identify trajectory segments that violate one-way traffic, no left turn, and no U-turn rules, write the violation trajectory segments into the trajectory table to be repaired, and generate alternative paths that comply with traffic rules based on road connectivity.

[0124] Step S702: Construct a trajectory reconstruction database. Write vehicle structured features, depth visual feature vectors, and multimodal sensor features into the feature table. Write trajectory point coordinate sequences, timestamp sequences, and road segment number sequences into the trajectory table. Write manual verification records, feature matching records, and trajectory repair records into the operation record table. Establish the association between the feature table, trajectory table, and operation record table to form a complete trajectory reconstruction dataset.

[0125] Optionally, this embodiment focuses on steps S701-S702, transforming the "which features are more reliable in which situations" derived from previous steps into an executable matching strategy. When non-compliant trajectories are detected, alternative paths are generated according to traffic rules, and the complete evidence chain is ultimately written into the trajectory reconstruction database. The process begins by reading the feature reliability score matrix from the rule engine. This matrix records structured features, depth vectors, radar size, and the historical stability of ETC identity based on point ID, camera pairs, time period, and weather slices. This embodiment sorts vehicle features within a single processing window in descending order of score, providing dynamic priorities: for example, at tunnel entrances, where color drift is significant, the color ranking shifts to the back, while depth and axle ranking rise; in daytime, smooth road sections, vehicle type and color are more effective. Priority doesn't just affect which is calculated first, but directly impacts the weight of the evidence chain in the matching decision. Low-scoring items only serve to fine-tune the scoring and provide anomaly alerts, preventing weak evidence from being used as a determinant.

[0126] This embodiment translates road traffic management regulations into machine-readable trajectory constraints. We extract rules such as one-way traffic, no left turns, no U-turns, and tidal flow from the regulation database and bind them to the edge and node attributes of the electronic map, forming a "turning feasibility table." The trajectory sequence is verified segment by segment against the table. Any segment showing reverse traffic in one direction, a left turn at a no-left-turn node, or a U-turn in a no-U-turn section is marked as a "violation trajectory segment" and added to the trajectory table to be repaired. Here, even strong visual evidence must obey traffic rules, because the goal of reconstruction is a "truly feasible" driving history, not just a patchwork of similarities.

[0127] This embodiment does not simply delete the trajectory segment to be repaired, but generates alternative paths based on road network connectivity. The generation of alternative paths involves two constraints: first, it fully satisfies the turning feasibility table and does not violate prohibited edges; second, it satisfies time feasibility, with the minimum travel time of the alternative path not exceeding the upper limit of the observation time interval (the upper limit is calculated based on road type and time period). Candidate paths are sorted by comprehensive score, which takes into account drivers' natural preferences: schemes with higher road segment levels, fewer turns, and lower historical delays are preferred. If the scores of the top two categories are close, we further examine surrounding unassigned camera footage and radar trajectory fragments as corroborating evidence, prioritizing schemes supported by external evidence. After manual confirmation with a single click on the interface, the alternative path is added to the table and replaces the non-compliant segment; the original non-compliant segment is retained as a historical version with a repair reason attached.

[0128] This example illustrates how priorities and rules interact in two scenarios. On a rainy night, on an elevated highway ramp, the color reliability score is low. The system uses vehicle type and depth as primary evidence to construct the path from A to B, but the trajectory makes a left turn at point B, which the rule engine interprets as a no-left-turn rule. Alternative path search generates a feasible path from B' to C at the adjacent U-turn point. The path is feasible in terms of time, and there is an unassigned capture frame near B', thus the repair is successful. Another scenario is a main urban road during morning rush hour, where white cars are clustered together, resulting in high depth vectors for multiple vehicles. This example significantly increases the weight of time margin and radar size based on the score. A previously favored straight path is abandoned due to insufficient time, and a right-turn detour is selected instead to avoid "visual congestion."

[0129] This embodiment transitions to step S702, where all intermediate products are stored in the database. We construct three main tables: a feature table, a trajectory table, and an operation record table. The feature table fields include structured features (vehicle type, color, axle, sunroof, and confidence level), depth visual feature vectors, and multimodal sensor features (radar size and motion status, ETC identity), and record the source location, timestamp, and feature reliability weights. The trajectory table stores coordinate sequences, time sequences, road segment number sequences, and an evidence summary for each spliced ​​or repaired segment (feature priority used, similarity statistics, time and rule verification results, and alternative path ID). The operation record table contains manual verification instructions, mismatch / missing match tags, threshold and weight version numbers, rule snapshot numbers, and explanations of the reasons for repair actions.

[0130] This embodiment establishes a strict relationship between the three tables. The trajectory ID runs through the feature table and the trajectory table, and the segment ID is used to link single-segment matching and repair actions with the operation record table; any changes are inserted into a new row with an incrementing version number, and old records are frozen for easy playback and auditing. To ensure data integrity, we perform two types of checks before data is entered into the database: timeline consistency (checking for cross-source time alignment drift, and adding a "time risk" label to abnormal segments) and rule consistency (confirming that each turn in the trajectory is feasible under the rule snapshot). Data that fails to meet the requirements is not included in the main chain and is moved to the review area.

[0131] This embodiment also considers a closed-loop data usage model. The database publishes two views: one for training and one for operations. The training view extracts paired positive and negative samples based on "high-confidence trajectories" to further refine the metric learning of deep features. The operations view automatically writes back to the rule engine based on the reliability of statistical features at different locations and time periods, completing the rolling update of the scoring matrix. To avoid parameter oscillations caused by a single misjudgment, we set a moderate learning rate and minimum sample threshold in the write-back channel. Only statistically recurring deviations change the priority order, maintaining system stability.

[0132] This embodiment outputs a traceable trajectory reconstruction dataset: it not only shows "which road it connects to," but also clearly states "why it connects in this way, where it was rejected by the rules, and how it was corrected." In actual work orders for public security inspections, the person in charge can open a trajectory and directly see the basis and supporting screenshots for the alternative path, naturally exposing the location of the suspicious points; in the daily work of model maintenance, the scoring matrix corresponds one-to-one with the rule snapshots, and reviewing the error correction a week ago allows the reproduction of the judgment at that time. These details weave the algorithm with the world of roads, making it both like an engineering project and a case file.

[0133] To effectively address the shortcomings of traditional technologies in feature fusion, trajectory filtering, and verification optimization, and to provide technical support for the management of unlicensed vehicles, this application provides an embodiment of an unlicensed vehicle trajectory reconstruction device for implementing all or part of the aforementioned unlicensed vehicle trajectory reconstruction method. See [link to embodiment]. Figure 2 The unlicensed vehicle trajectory reconstruction device specifically includes the following components:

[0134] The feature fusion module 10 is used to collect vehicle images and location and time data captured by road monitoring cameras, input the vehicle images into a feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings, input the vehicle images into a convolutional neural network to extract deep visual feature vectors, collect vehicle size data from radar sensors and identity data from the electronic toll collection system, and align the structured features, deep visual feature vectors, vehicle size data, and identity data according to timestamps to construct a feature fusion dataset.

[0135] The road analysis module 20 is used to perform preliminary screening of vehicles in video data based on structured features, calculate cosine similarity of the depth visual feature vectors for precise matching, read road topology data from electronic maps, calculate travel time constraints and path accessibility constraints of adjacent capture points, use the travel time constraints to remove trajectory points that violate driving speed limits, filter trajectory points that cross physical obstacles based on the path accessibility constraints, dynamically adjust spatiotemporal thresholds according to road type, and fit missing road segments using road connectivity relationships.

[0136] The trajectory restoration module 30 is used to input the filtered vehicle trajectory sequence into the batch verification interface, dynamically adjust the similarity threshold and matching weight based on human feedback, read the camera location information to construct a feature reliability scoring matrix, write the feature reliability score into the rule engine, adaptively adjust the feature matching priority under different scenarios, filter trajectory segments that violate traffic rules using road traffic rules, generate a trajectory repair scheme based on road connectivity constraints, and construct a trajectory restoration database containing vehicle features, trajectory segments, and verification records.

[0137] As described above, the unlicensed vehicle trajectory reconstruction device provided in this application can accurately identify vehicles through feature extraction and data alignment by innovatively designing a multimodal feature fusion model. It constructs a trajectory screening system, combining spatiotemporal constraints and path verification to establish a reliable trajectory reconstruction mechanism. Verification optimization is introduced, and the accuracy of the reconstruction results is ensured through reliability assessment and rule adjustment. This method effectively solves the shortcomings of traditional technologies in feature fusion, trajectory screening, and verification optimization, providing technical support for the management of unlicensed vehicles.

[0138] From a hardware perspective, in order to effectively address the shortcomings of traditional technologies in feature fusion, trajectory filtering, and verification optimization, and to provide technical support for the management of unlicensed vehicles, this application provides an embodiment of an electronic device for implementing all or part of the aforementioned unlicensed vehicle trajectory reconstruction method. The electronic device specifically includes the following components:

[0139] The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the unlicensed vehicle trajectory reconstruction device and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the unlicensed vehicle trajectory reconstruction method and the unlicensed vehicle trajectory reconstruction device in the embodiments, the content of which is incorporated herein, and repeated details will not be repeated.

[0140] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0141] In practical applications, the method for reconstructing the trajectory of unlicensed vehicles can be partially executed on the electronic device side as described above, or all operations can be completed in the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed in the client device, the client device may further include a processor.

[0142] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0143] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0144] In one embodiment, the function of reconstructing the trajectory of a vehicle without a license plate can be integrated into a central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0145] Step S101: Collect vehicle images and location time data captured by road monitoring cameras, input the vehicle images into a feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings, input the vehicle images into a convolutional neural network to extract deep visual feature vectors, collect vehicle size data from radar sensors and identity data from the electronic toll collection system, and align the structured features, deep visual feature vectors, vehicle size data, and identity data according to timestamps to construct a feature fusion dataset;

[0146] Step S102: Perform preliminary screening of vehicles in video data based on structured features, calculate cosine similarity of the depth visual feature vectors for precise matching, read road topology data from electronic map, calculate travel time constraints and path accessibility constraints of adjacent capture points, use the travel time constraints to remove trajectory points that violate driving speed limits, filter trajectory points that cross physical obstacles based on the path accessibility constraints, dynamically adjust spatiotemporal thresholds according to road type, and fit missing road segments using road connectivity relationships;

[0147] Step S103: Input the filtered vehicle trajectory sequence into the batch verification interface, dynamically adjust the similarity threshold and matching weight based on human feedback, read the camera location information to construct a feature reliability scoring matrix, write the feature reliability score into the rule engine, adaptively adjust the feature matching priority under different scenarios, filter trajectory segments that violate traffic rules using road traffic rules, generate a trajectory repair scheme based on road connectivity constraints, and construct a trajectory restoration database containing vehicle features, trajectory segments, and verification records.

[0148] As described above, the electronic device provided in this application, through an innovative design of a multimodal feature fusion model, achieves accurate vehicle identification via feature extraction and data alignment. A trajectory screening system is constructed, combining spatiotemporal constraints and path verification to establish a reliable trajectory reconstruction mechanism. Verification optimization is introduced, ensuring the accuracy of the reconstruction results through reliability assessment and rule adjustment. This method effectively addresses the shortcomings of traditional technologies in feature fusion, trajectory screening, and verification optimization, providing technical support for the management of unlicensed vehicles.

[0149] In another embodiment, the unlicensed vehicle trajectory reconstruction device can be configured separately from the central processing unit 9100. For example, the unlicensed vehicle trajectory reconstruction device can be configured as a chip connected to the central processing unit 9100, and the unlicensed vehicle trajectory reconstruction method function can be realized through the control of the central processing unit.

[0150] like Figure 3 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3 For components not shown, please refer to existing technologies.

[0151] like Figure 3 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0152] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0153] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0154] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0155] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device for communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0156] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0157] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored audio via the speaker 9131.

[0158] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the unlicensed vehicle trajectory reconstruction method with a server or client as the execution subject in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the unlicensed vehicle trajectory reconstruction method with a server or client as the execution subject in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0159] Step S101: Collect vehicle images and location time data captured by road monitoring cameras, input the vehicle images into a feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings, input the vehicle images into a convolutional neural network to extract deep visual feature vectors, collect vehicle size data from radar sensors and identity data from the electronic toll collection system, and align the structured features, deep visual feature vectors, vehicle size data, and identity data according to timestamps to construct a feature fusion dataset;

[0160] Step S102: Perform preliminary screening of vehicles in video data based on structured features, calculate cosine similarity of the depth visual feature vectors for precise matching, read road topology data from electronic map, calculate travel time constraints and path accessibility constraints of adjacent capture points, use the travel time constraints to remove trajectory points that violate driving speed limits, filter trajectory points that cross physical obstacles based on the path accessibility constraints, dynamically adjust spatiotemporal thresholds according to road type, and fit missing road segments using road connectivity relationships;

[0161] Step S103: Input the filtered vehicle trajectory sequence into the batch verification interface, dynamically adjust the similarity threshold and matching weight based on human feedback, read the camera location information to construct a feature reliability scoring matrix, write the feature reliability score into the rule engine, adaptively adjust the feature matching priority under different scenarios, filter trajectory segments that violate traffic rules using road traffic rules, generate a trajectory repair scheme based on road connectivity constraints, and construct a trajectory restoration database containing vehicle features, trajectory segments, and verification records.

[0162] As described above, the computer-readable storage medium provided in this application, through an innovative design of a multimodal feature fusion model, achieves accurate vehicle identification via feature extraction and data alignment. A trajectory screening system is constructed, combining spatiotemporal constraints and path verification to establish a reliable trajectory reconstruction mechanism. Verification optimization is introduced, ensuring the accuracy of the reconstruction results through reliability assessment and rule adjustment. This method effectively addresses the shortcomings of traditional technologies in feature fusion, trajectory screening, and verification optimization, providing technical support for the management of unlicensed vehicles.

[0163] Embodiments of this application also provide a computer program product capable of implementing all steps in the unlicensed vehicle trajectory reconstruction method described above, where the execution subject is a server or client. When executed by a processor, this computer program / instruction implements the steps of the unlicensed vehicle trajectory reconstruction method. For example, the computer program / instruction implements the following steps:

[0164] Step S101: Collect vehicle images and location time data captured by road monitoring cameras, input the vehicle images into a feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings, input the vehicle images into a convolutional neural network to extract deep visual feature vectors, collect vehicle size data from radar sensors and identity data from the electronic toll collection system, and align the structured features, deep visual feature vectors, vehicle size data, and identity data according to timestamps to construct a feature fusion dataset;

[0165] Step S102: Perform preliminary screening of vehicles in video data based on structured features, calculate cosine similarity of the depth visual feature vectors for precise matching, read road topology data from electronic map, calculate travel time constraints and path accessibility constraints of adjacent capture points, use the travel time constraints to remove trajectory points that violate driving speed limits, filter trajectory points that cross physical obstacles based on the path accessibility constraints, dynamically adjust spatiotemporal thresholds according to road type, and fit missing road segments using road connectivity relationships;

[0166] Step S103: Input the filtered vehicle trajectory sequence into the batch verification interface, dynamically adjust the similarity threshold and matching weight based on human feedback, read the camera location information to construct a feature reliability scoring matrix, write the feature reliability score into the rule engine, adaptively adjust the feature matching priority under different scenarios, filter trajectory segments that violate traffic rules using road traffic rules, generate a trajectory repair scheme based on road connectivity constraints, and construct a trajectory restoration database containing vehicle features, trajectory segments, and verification records.

[0167] As described above, the computer program product provided in this application, through innovative design of a multimodal feature fusion model, achieves accurate vehicle identification via feature extraction and data alignment. A trajectory screening system is constructed, combining spatiotemporal constraints and path verification to establish a reliable trajectory reconstruction mechanism. Verification optimization is introduced, ensuring the accuracy of the reconstruction results through reliability assessment and rule adjustment. This method effectively addresses the shortcomings of traditional technologies in feature fusion, trajectory screening, and verification optimization, providing technical support for the management of unlicensed vehicles.

[0168] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for reconstructing the trajectory of an unlicensed vehicle, characterized in that, The method includes: Vehicle images and location / time data captured by road surveillance cameras are collected. The vehicle images are input into a feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings. The vehicle images are then input into a convolutional neural network to extract deep visual feature vectors. Vehicle size data from radar sensors and identity data from an electronic toll collection system are collected. The structured features, deep visual feature vectors, vehicle size data, and identity data are aligned by timestamps to construct a feature fusion dataset. The process involves: performing preliminary screening of vehicles in video data based on structured features; calculating cosine similarity between deep visual feature vectors for precise matching; reading road topology data from an electronic map; calculating travel time constraints and path reachability constraints between adjacent capture points; and using the travel time constraints to eliminate trajectory points that violate speed limits. This includes: reading structured features from a feature fusion dataset; performing a coarse selection of vehicles in the video data based on vehicle type classification and color matching; calculating cosine similarity between pairs of deep visual feature vectors in the coarse selection results; constructing a candidate vehicle similarity matrix; setting a similarity threshold to filter high-confidence matching pairs; writing the spatiotemporal information of the high-confidence matching pairs into a trajectory data table; and constructing a vehicle motion trajectory sequence based on the trajectory data table. The process also involves reading road topology data from an electronic map, extracting road connectivity and traffic rules, calculating the shortest path distance between adjacent capture points, calculating the theoretical minimum travel time based on road speed limits, setting the theoretical minimum travel time as the lower limit of the time constraint, comparing the lower limit of the constraint with the time interval in the trajectory data table, and eliminating trajectory points with time intervals less than the lower limit of the constraint. Based on the path reachability constraints, the trajectory points that cross physical obstacles are filtered out, the spatiotemporal thresholds are dynamically adjusted according to the road type, and the missing road segments are fitted using road connectivity relationships. The filtered vehicle trajectory sequences are input into the batch verification interface. The similarity threshold and matching weight are dynamically adjusted based on human feedback. The camera location information is read to construct a feature reliability scoring matrix. The feature reliability score is written into the rule engine. The feature matching priority under different scenarios is adaptively adjusted. The road traffic rules are used to filter trajectory segments that violate traffic rules. The trajectory repair scheme is generated based on road connectivity constraints. A trajectory restoration database containing vehicle features, trajectory segments, and verification records is constructed.

2. The method for reconstructing the trajectory of an unlicensed vehicle according to claim 1, characterized in that, The process involves collecting vehicle images and location / time data captured by road surveillance cameras, inputting the vehicle images into a feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings, and then inputting the vehicle images into a convolutional neural network to extract deep visual feature vectors, including: Vehicle images are captured from road surveillance cameras. The latitude and longitude coordinates and timestamp information at the time of image capture are read. The captured vehicle images are preprocessed, and image enhancement and size normalization are performed. The vehicle type classification dictionary and vehicle color dictionary are read. The normalized image is input into the feature extraction model. The vehicle outline is identified based on the edge detection algorithm. The structured features of vehicle type, color, number of axles and sunroof are extracted based on the region segmentation algorithm. A convolutional neural network comprising an encoding layer, a pooling layer, and a fully connected layer is constructed. The normalized image is input into the encoding layer to extract local feature maps. The feature maps are input into the pooling layer to reduce the data dimensionality. The dimensionality-reduced features are input into the fully connected layer to generate a deep visual feature vector. Normalization calculation is performed on the deep visual feature vector to generate a feature representation of the vehicle image.

3. The method for reconstructing the trajectory of an unlicensed vehicle according to claim 1, characterized in that, The vehicle size data collected from the radar sensor and the identity data from the electronic toll collection system are combined, and the structured features, depth visual feature vectors, vehicle size data, and identity data are aligned by timestamps to construct a feature fusion dataset, including: Read the roadside millimeter-wave radar data stream, parse the radar data stream into target detection data, extract vehicle length, width and height dimension parameters and motion state parameters, read vehicle passage data collected by the electronic toll collection system, parse electronic tag information to obtain vehicle identity features, group the vehicle dimension parameters and identity features according to the collection location and lane number, and construct a multi-source data correspondence table. Read the structured features and deep visual feature vectors, sort the structured features, deep visual feature vectors, vehicle size parameters, and identity features according to timestamp information, calculate the temporal correspondence between features based on a sliding time window, perform interpolation alignment on the feature data, combine the aligned multimodal features to generate a fused feature vector, and construct a feature fusion dataset.

4. The method for reconstructing the trajectory of an unlicensed vehicle according to claim 1, characterized in that, The process of filtering trajectory points that traverse physical obstacles based on the path reachability constraints, dynamically adjusting spatiotemporal thresholds according to road type, and fitting missing road segments using road connectivity relationships includes: Read the physical obstacle data in the electronic map, extract the spatial distribution information of the median strip, river and building, calculate the set of reachable paths between trajectory points based on the road topology, match the set of reachable paths with the trajectory sequence, identify the trajectory segments that cross physical obstacles, mark the trajectory segments as invalid trajectories and delete them from the trajectory data table, and filter abnormal trajectory points based on path reachability constraints. Read road attribute information and classify road types into expressways, urban expressways, and urban arterial roads. Calculate the maximum permissible speed based on road capacity and write the maximum permissible speed into a spatiotemporal constraint parameter table. Dynamically set the trajectory point time interval threshold according to road type. Read road connectivity relationships to construct a road network topology map. Generate candidate paths for missing road segments based on the shortest path algorithm. Sort the candidate paths by road level and select the path with the highest weighted score to fit and complete the trajectory.

5. The method for reconstructing the trajectory of an unlicensed vehicle according to claim 1, characterized in that, The process of inputting the filtered vehicle trajectory sequences into a batch verification interface, dynamically adjusting the similarity threshold and matching weight based on human feedback, reading camera location information to construct a feature reliability scoring matrix, and writing the feature reliability score into the rule engine includes: The vehicle trajectory sequence is grouped by time window, and the feature matching results and constraint filtering results of each group of trajectories are read. A verification data table containing trajectory number, vehicle features and matching confidence is constructed. The verification data table is input into the batch verification interface to record the manually confirmed mismatched trajectories and missed detection trajectories. Based on the manual feedback results, the feature similarity correction coefficient is calculated and applied to the dynamic adjustment of the similarity threshold. The feature matching weight is adaptively updated. The system reads the camera deployment location and coverage area, calculates the feature recognition accuracy at each location, calculates the feature reliability score under different lighting and weather conditions, writes the feature reliability score into a scoring matrix, constructs feature weight adaptive rules based on the scoring matrix, deploys the adaptive rules to the rule engine, and dynamically configures the feature matching strategy for different scenarios.

6. The method for reconstructing the trajectory of an unlicensed vehicle according to claim 1, characterized in that, The method adaptively adjusts the feature matching priority under different scenarios, filters trajectory segments that violate traffic rules using road traffic rules, generates trajectory repair schemes based on road connectivity constraints, and constructs a trajectory restoration database containing vehicle features, trajectory segments, and verification records, including: Read the feature reliability scores from the rule engine, sort the vehicle features in descending order of reliability scores, dynamically allocate feature matching priorities based on the sorting results, read the road traffic management regulations, extract motor vehicle traffic rules and turning restriction rules, convert the traffic rules into trajectory constraints, identify trajectory segments that violate one-way traffic, no left turn, and no U-turn rules, write the violation trajectory segments into the trajectory table to be repaired, and generate alternative paths that comply with traffic rules based on road connectivity. A trajectory reconstruction database is constructed by writing vehicle structured features, depth visual feature vectors, and multimodal sensor features into a feature table, trajectory point coordinate sequences, timestamp sequences, and road segment number sequences into a trajectory table, and manual verification records, feature matching records, and trajectory repair records into an operation record table. The relationship between the feature table, trajectory table, and operation record table is established to form a complete trajectory reconstruction dataset.

7. A device for reconstructing the trajectory of an unlicensed vehicle, characterized in that, The device includes: The feature fusion module is used to collect vehicle images and location and time data captured by road monitoring cameras, input the vehicle images into the feature extraction model to extract structured features such as vehicle type, color, number of axles, and sunroof markings, input the vehicle images into a convolutional neural network to extract deep visual feature vectors, collect vehicle size data from radar sensors and identity data from the electronic toll collection system, and align the structured features, deep visual feature vectors, vehicle size data, and identity data by timestamps to construct a feature fusion dataset. The road analysis module is used to perform preliminary screening of vehicles in video data based on structured features. It calculates cosine similarity of the deep visual feature vectors for precise matching, reads road topology data from an electronic map, calculates travel time constraints and path reachability constraints for adjacent capture points, and uses the travel time constraints to eliminate trajectory points that violate speed limits. This includes: reading structured features from the feature fusion dataset; performing a coarse selection of vehicles in the video data based on vehicle type classification and color matching; calculating the cosine similarity of each pair of deep visual feature vectors in the coarse selection results; constructing a candidate vehicle similarity matrix; setting a similarity threshold to filter high-confidence matching pairs; and then... The spatiotemporal information of the high-confidence matching pairs is written into the trajectory data table, and a vehicle motion trajectory sequence is constructed based on the trajectory data table. The road topology data of the electronic map is read, the road connection relationship and traffic rules are extracted, the shortest path distance between adjacent capture points is calculated, the theoretical minimum passage time is calculated based on the road speed limit standard, the theoretical minimum passage time is set as the lower limit of the time constraint, the lower limit of the constraint is compared with the time interval in the trajectory data table, and trajectory points with time intervals less than the lower limit of the constraint are eliminated. Based on the path accessibility constraint, trajectory points that cross physical obstacles are filtered out, the spatiotemporal threshold is dynamically adjusted according to the road type, and the missing road segments are fitted using road connectivity relationships. The trajectory restoration module is used to input the filtered vehicle trajectory sequences into the batch verification interface, dynamically adjust the similarity threshold and matching weight based on human feedback, read camera location information to construct a feature reliability scoring matrix, write the feature reliability score into the rule engine, adaptively adjust the feature matching priority under different scenarios, filter trajectory segments that violate traffic rules using road traffic rules, generate trajectory repair schemes based on road connectivity constraints, and construct a trajectory restoration database containing vehicle features, trajectory segments, and verification records.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for restoring the trajectory of an unlicensed vehicle as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for reconstructing the trajectory of an unlicensed vehicle as described in any one of claims 1 to 6.

Images