Blind area event inference method using vehicle trajectory boundary distribution association

Abstract

This invention relates to a blind spot event inference method based on vehicle trajectory boundary distribution correlation, comprising the following steps: S1, for road segments containing road perception blind spots, setting up an upstream perception zone, a road perception blind spot, and a downstream perception zone, and collecting vehicle perception data from the upstream and downstream perception zones; S2, preprocessing the collected vehicle perception data, performing trajectory correlation, and stitching together short trajectory data to obtain complete vehicle trajectory data; S3, statistically analyzing basic vehicle motion information and presetting discrimination conditions for the existence of special traffic events within the blind spot; S4, inferring the specific location blocks of special traffic events within the blind spot; S5, adaptively training and calibrating the special traffic event recognition model to achieve long-term continuous optimization of the model's recognition performance. This invention can dynamically collect blind spot situation data, timely and accurately identify special traffic events within the blind spot, and effectively avoid traffic congestion, accidents, and secondary accidents caused by unclear blind spot event information.

Description

Technical Field

[0001] This invention relates to the field of road safety, and more specifically, to a method for inferring blind spot events using the correlation of vehicle trajectory boundary distribution. Background Technology

[0002] While road perception and recognition technologies are constantly improving, roadside sensing cannot cover the entire road. Significant blind spots remain in some roads or sections of roads, meaning these sections lack corresponding traffic facilities and cannot provide timely and accurate traffic information. Special traffic events within these blind spots often cause traffic congestion, accidents, or secondary accidents due to the lack of information. Current traffic incident detection technologies largely rely on roadside sensor equipment and radar-visual fusion technology, typically covering road sections with complete facilities, using traffic incident detection as the primary indicator.

[0003] In the event of a traffic incident in a blind spot, the lack of traffic sensing equipment makes it impossible to obtain relevant information in a timely manner, thus delaying the handling of subsequent traffic incidents. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a blind spot event inference method that uses the distribution of vehicle trajectory boundaries to expand the dimensions and coverage of road traffic information acquisition, dynamically collect blind spot situation data, and timely and accurately identify special traffic events in blind spots. This provides comprehensive road condition information for traffic safety management and effectively avoids traffic congestion, accidents and secondary accidents caused by unclear blind spot event information.

[0005] The technical solution adopted by this invention to solve its technical problem is: to construct a blind spot event inference method based on the correlation of vehicle trajectory boundary distribution, comprising the following steps: S1. For road sections containing blind spots in road perception, an upstream perception zone, a blind spot in road perception, and a downstream perception zone are set up. Vehicle perception data in the upstream and downstream perception zones are collected through roadside equipment. Traffic events in the blind spot are inferred from the trajectory information of any area in the upstream and downstream perception zones. The identification status of traffic events in the blind spot in road perception includes "there is a special traffic event" and "there is no special traffic event". S2. Preprocess the collected vehicle perception data, remove data that is irrelevant to the vehicle position and summarize the data that is related to the trajectory, use the spatiotemporal matching method to associate the trajectory, and stitch together the short trajectory data to obtain complete vehicle trajectory data containing vehicle position, speed and acceleration information. S3. Collect basic vehicle motion information, including traffic status information and abnormal behavior information, from complete vehicle trajectory data. At the same time, based on vehicle trajectory characteristics, preset the discrimination conditions for the existence of special traffic events in the blind spot. S4. Based on the statistical vehicle motion information, extract low-dimensional trajectory feature parameters, construct high-dimensional trajectory behavior feature vectors using multilayer perceptron (MLP), input the high-dimensional trajectory behavior feature vectors into the classifier model to determine whether there are special traffic events in the blind spot, and infer the specific location block of the special traffic event in the blind spot based on the upstream and downstream coupling degree. S5. Adaptive training and parameter calibration are performed on the special traffic event recognition model to achieve long-term continuous optimization of the model's recognition performance.

[0006] According to the above scheme, in step S1, the vehicle perception data includes vehicle motion state, trajectory type and driving behavior information; the roadside equipment data includes at least one of the perception data of roadside radar data and vehicle-mounted GPS data; special traffic events are events that cause the road to be unusable, including traffic accidents and road debris.

[0007] According to the above scheme, in step S2, the spatiotemporal matching method adopts a unified time reference and spatial coordinates to align vehicle trajectories collected by different sampling methods and different devices to the same spatiotemporal coordinate system, thereby eliminating time deviation and spatial offset and realizing accurate correspondence of vehicle trajectories in time and space; the vehicle trajectory includes information on position, speed and acceleration. According to the above scheme, in step S3, the basic state variables for extracting basic vehicle motion information include average vehicle speed, vehicle distance, number of vehicles in lane, blind spot travel time, and acceleration / deceleration. The criteria for identifying special traffic events within the pre-defined blind spot are any of the following: ① The vehicle changes lanes before entering the blind spot, or the number of vehicles in the lane changes significantly; ② The vehicle decelerates before entering the blind spot and accelerates after leaving the blind spot. According to the above scheme, in step S4, the low-dimensional trajectory feature parameter is P[ ], including upstream trajectory feature parameters Downstream trajectory characteristic parameters upstream and downstream related characteristic parameters ; Among them, velocity characteristic parameters Including the average rate of change of velocity v. Abnormal deceleration frequency and the average of abnormal deceleration locations Longitudinal feature parameters Including the average vehicle distance change rate and the proportion of dangerous following behaviors ; lateral characteristic parameters Including the average lane change position Lane changing behavior ratio and the mean of lateral offset of the trajectory Upstream and downstream related characteristic parameters Including behavioral consistency and speed deviation .

[0008] According to the above scheme, in step S4, the high-dimensional trajectory behavior feature vector constructed using multilayer perceptron MLP includes the upstream avoidance behavior vector. Downstream recovery behavior vector Upstream and downstream coupling behavior vector ;in, From upstream trajectory feature parameters Obtained through two layers of linear transformation by MLP, Downstream trajectory feature parameters Obtained through two layers of linear transformation by MLP, Depend on , and The high-dimensional input vector after dimensional merging is obtained through a two-layer linear transformation of the MLP; the weight matrix and bias vector in the MLP are determined by initialization-model training-optimization.

[0009] According to the above scheme, in step S4, the upstream avoidance behavior vector is... Downstream recovery behavior vector Upstream and downstream coupling behavior vector By concatenating the features, a comprehensive trajectory behavior feature vector is obtained. ;Will Inputting the classifier model f(x) yields the probability of the existence of blind zone events. Preset classification decision threshold ,when ≥ If a special traffic event is detected in the blind spot, it is determined that there is one; otherwise, it is determined that there is no special traffic event.

[0010] According to the above scheme, the classifier model adopts the XGBoost classifier, and the classification decision threshold is... Continuous optimization is achieved through model training.

[0011] According to the above scheme, in step S4, the location range of special traffic events in the blind zone is inferred based on the upstream and downstream coupling degree, and the vehicle deceleration rate per unit mileage is calculated. The percentage of vehicles returning to the straight lane Lane number deviation rate Implementation: ① Used to determine the position of an event before or after it in the blind zone. The larger the value, the closer the event is to the start of the blind zone. The smaller the value, the closer the event is to the end of the blind zone; ② Used to determine the lane where the incident occurred. The higher the value, the higher the probability of a special traffic incident occurring in that lane; ③ Including the deviation rate of the number of vehicles in the upstream lane Deviation rate of vehicle count in downstream lane , The larger the value and The smaller the value, the greater the probability of a special traffic incident occurring in the corresponding lane.

[0012] According to the above scheme, the adaptive training and parameter calibration process of the special traffic event identification model in step S5 includes the following steps: S501. Obtain labeled sample data of "with special traffic events" and "without special traffic events" in the blind spot through manual calibration to form a labeled training dataset; S502. Use supervised learning to pre-train the model and establish a mapping relationship between trajectory behavior features and event existence state; S503: The model identifies real-time trajectory data and outputs results, while obtaining verification feedback data based on patrol confirmation, equipment detection, or manual reporting information; S504. Based on the consistency between the model recognition results and the verification feedback results, adaptively update and calibrate the model's weight matrix, bias vector, and classification decision threshold τ.

[0013] The blind spot event inference method based on vehicle trajectory boundary distribution correlation of the present invention has the following beneficial effects: 1. This invention can overcome the limitations of blind spot detection and realize event recognition in areas without facilities. It does not require the addition of detection facilities in the blind spot of road perception. By analyzing the distribution characteristics of vehicle trajectories upstream and downstream of the blind spot, it completes the identification of special traffic events. It solves the technical limitation of existing technologies that can only detect road sections with complete facilities, effectively expands the coverage of road traffic event detection, and is suitable for all road discontinuous perception scenarios. 2. This invention can improve the timeliness and accuracy of blind spot event recognition, reduce traffic risks, achieve precise correlation of vehicle trajectories through spatiotemporal matching technology, extract multi-dimensional trajectory features and combine them with MLP + XGBoost models for event discrimination, and infer the specific location range of the event within the blind spot. Compared with the "unperceived" state of blind spot events in existing technologies, it can timely and accurately identify special events such as traffic accidents and road debris, avoid traffic congestion, accidents and secondary accidents caused by missing information, and significantly reduce road traffic risks. 3. This invention can optimize the dimensions of traffic information acquisition and improve the data support capability for management and control. It extracts trajectory features from four dimensions: speed, longitudinal, lateral, and upstream and downstream coupling, and maps low-dimensional trajectory data into high-dimensional behavioral features, which enriches the dimensions of road traffic information acquisition. It can dynamically collect traffic situation data in blind spots, provide comprehensive and accurate road condition information for road traffic safety management and control, and improve the scientificity and effectiveness of management and control decisions. 4. This invention can achieve model self-optimization, ensuring long-term stable recognition performance. The model has adaptive training and parameter calibration functions. Combined with manual calibration sample pre-training and real-time verification feedback update, it can continuously optimize model parameters and classification thresholds according to actual road conditions, so that the accuracy and generalization ability of event recognition can be continuously improved in long-term operation. Compared with fixed parameter detection methods, it can adapt to complex and ever-changing road traffic scenarios and ensure the stability of detection performance. 5. This invention can reduce detection costs and improve the practicality of the technology. It relies on existing roadside radar, vehicle GPS and other equipment to collect upstream and downstream data. There is no need to add blind spot detection hardware facilities. It can expand the detection capability without increasing hardware costs, which greatly saves the construction and operation and maintenance costs of traffic perception system. At the same time, the technical process is standardized, easy to connect with existing traffic control system, and has strong operability and implementation. 6. This invention can achieve precise event location and improve emergency response efficiency. By using indicators such as deceleration rate per unit mileage, the proportion of vehicles returning to the correct lane, and the deviation rate of vehicles in the lane, it can accurately infer the location and lane of special traffic events in blind spots. Compared with traditional blind spot-free event location, it can provide accurate location information for traffic emergency response, shorten response time, and improve event handling efficiency. 7. This invention eliminates the need to install additional traffic detection facilities within the blind spot of road perception. By analyzing the vehicle trajectory distribution characteristics upstream and downstream of the blind spot, it can achieve timely and accurate identification of special traffic events within the blind spot and infer the specific location range of the event. The model has an adaptive training function, which can continuously optimize performance according to actual road conditions. Under limited traffic facilities, it effectively expands the dimensions and coverage of road traffic information acquisition, providing comprehensive and real-time road condition data support for road traffic safety management and control. It can significantly reduce the probability of traffic congestion, traffic accidents, and secondary accidents caused by unclear information about events in the blind spot. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of road zoning for the blind spot event inference method based on vehicle trajectory boundary distribution association, as described in this invention. Figure 2This is a schematic diagram of the model structure for identifying special traffic events within blind spots using the blind spot event inference method based on the distribution of vehicle trajectory boundaries, as described in this invention. Figure 3 This is a schematic diagram of the adaptive training process of the model created by the blind spot event inference method based on the distribution of vehicle trajectory boundaries in this invention. Detailed Implementation

[0015] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0016] like Figure 1-3 As shown, the blind spot event inference method of the present invention, which utilizes the correlation of vehicle trajectory boundary distribution, includes the following steps: S1. Perform regional division and data collection, specifically as follows: For roads with blind spots, upstream sensing zones, road sensing blind spots, and downstream sensing zones are set up sequentially according to the vehicle's direction of travel. The upstream sensing zone is the section of road before the vehicle enters the blind spot, where the vehicle will take proactive driving actions based on the road conditions within the blind spot. The road sensing blind spot is the section of road lacking traffic detection facilities and where vehicle data cannot be directly collected. The downstream sensing zone is the section of road after the vehicle leaves the blind spot, where the vehicle will gradually return to normal driving behavior.

[0017] By using roadside equipment such as roadside radar and vehicle-mounted GPS, vehicle perception data is collected from the upstream and downstream perception zones. The data types include vehicle motion status, trajectory type, and driving behavior. The special traffic events to be identified are those that cause the road to become unusable, such as traffic accidents and road debris. The identification status of the road perception blind spot is only divided into two categories: "with special traffic events" and "without special traffic events".

[0018] S2. Perform data preprocessing and trajectory association, specifically including the following steps: S201. Data preprocessing: Collect various types of vehicle data from roadside equipment, remove non-vehicle location data, and retain and summarize only the data related to vehicle trajectories to prepare for subsequent trajectory analysis.

[0019] S202. Trajectory Association: This involves using spatiotemporal matching technology to associate vehicle trajectories. This technology aligns vehicle trajectories collected by different sampling methods and devices to the same spatiotemporal coordinate system by unifying the time base and spatial coordinates, eliminating time deviations and spatial offsets, and achieving precise temporal and spatial correspondence of vehicle trajectories. The spatiotemporally matched short trajectory data is then stitched together to obtain complete vehicle trajectory data. This complete vehicle trajectory data contains at least three core types of information: vehicle position, velocity, and acceleration. Vehicle location information: Obtain the vehicle's lane position in the avoidance zone and recovery zone, as well as lane changes before entering and after exiting the blind spot. Vehicle speed information: Extract vehicle speeds from the upstream and downstream sensing zones. Vehicle acceleration information: Extract vehicle acceleration and deceleration data from the upstream and downstream sensing zones.

[0020] S3. Perform basic vehicle motion information statistics and preset event discrimination conditions. From complete vehicle trajectory data, statistically analyze basic vehicle motion information including traffic status information and abnormal behavior information, and preset discrimination conditions for special traffic events occurring in blind spots. Specifically, this includes the following steps: S301, Extraction of Basic Vehicle Motion Information The extracted basic state variables include average vehicle speed, distance between vehicles, number of vehicles per lane, blind spot travel time, and acceleration / deceleration. The functions of each variable are as follows: Average speed: The average speed of a vehicle in the upstream and downstream sensing zones, reflecting the overall traffic conditions when entering and exiting the blind spot, and can be used to preliminarily determine whether there are special traffic events in the blind spot. Distance: The distance between the front and rear of the vehicle, reflecting the change in distance between vehicles traveling upstream and downstream of the blind spot, and can be used to infer whether a special traffic event in the blind spot was observed before the vehicle entered. Number of vehicles in each lane: The number of vehicles in each lane, reflecting whether a lane-changing strategy was adopted before entering the blind spot, thus inferring the road conditions in the blind spot. Blind spot travel time: The time from when a vehicle enters to when it exits the blind spot, revealing whether the vehicle travels normally within the blind spot. Acceleration / deceleration: The speed changes of the vehicle in the upstream and downstream areas, reflecting whether the vehicle decelerates upon detecting a blind spot event and its speed recovery behavior after exiting the blind spot.

[0021] S302. Event Judgment Conditions Preset: Based on vehicle trajectory data, preset judgment conditions are set for the existence of abnormal traffic events in the blind spot. Meeting any one of these conditions can preliminarily determine the existence of a special traffic event in the blind spot. ① The vehicle changes lanes before entering the blind spot (the vehicle trajectory deviates), or the number of vehicles in a certain lane changes significantly; ② The vehicle decelerates before entering the blind spot and accelerates significantly after exiting the blind spot.

[0022] S4. Event recognition and location reasoning based on neural networks: Based on the basic vehicle motion information collected in step S3, this method extracts low-dimensional trajectory feature parameters, constructs high-dimensional trajectory behavior feature vectors, and inputs them into a classifier model to achieve accurate identification of special traffic events within blind spots. Simultaneously, based on upstream and downstream coupling, it infers the specific location range of special traffic events within the blind spot. Specifically, this includes the following steps: S401, Low-dimensional trajectory feature parameter extraction Extracting low-dimensional trajectory feature parameters P[ This parameter includes upstream trajectory feature parameters. Downstream trajectory characteristic parameters upstream and downstream related characteristic parameters Three categories, including velocity characteristic parameters Longitudinal characteristic parameters and lateral feature parameters The specific calculation method is as follows: Velocity characteristic parameters Including the average rate of change of velocity abnormal deceleration frequency Average of abnormal deceleration locations wait: Average rate of change

[0023]

[0024] In the formula, This represents the speed of a single vehicle within the upstream sensing area. This indicates the speed at which the vehicle began to decelerate. This indicates the speed at which the vehicle comes to a stop and decelerates, or the speed at which it enters a blind spot. This indicates the total number of vehicles in the upstream sensing area.

[0025] abnormal deceleration frequency :

[0026] In the formula, This indicates the number of vehicles that decelerated abnormally in the corresponding lane. This represents the total traffic flow entering the blind spot from the corresponding lane.

[0027] Average of abnormal deceleration locations

[0028] First, fit the deceleration position probability density function using kernel density estimation (KDE). :

[0029] Kernel function Gaussian ,bandwidth According to Silverman's principles Adaptive selection. Then determine the location of the density peak. Final calculation , A significant deviation from the peak deceleration distance during normal driving can be used as a characteristic indicator of the existence of an event.

[0030] Longitudinal feature parameters Including the average vehicle distance change rate Proportion of dangerous following behaviors wait: Average vehicle distance change rate

[0031]

[0032] In the formula, This indicates the distance between the front ends of the vehicles in the upstream sensing area. This indicates the distance between vehicles at normal travel speeds on this road section.

[0033] Proportion of dangerous following behaviors

[0034]

[0035] In the formula, This indicates the number of vehicles whose headway distance decreased in the upstream sensing area.

[0036] Lateral feature parameters Including the average lane change position Lane changing behavior ratio Mean of lateral offset of trajectory wait: Lane change position average :

[0037] In the formula, The total number of vehicles that changed lanes upstream of the blind spot. Indicates the first The distance from the edge of the blind spot when a vehicle changes lanes.

[0038] Lane change behavior ratio :

[0039] In the formula, The number of vehicles that change lanes upstream of the blind spot.

[0040] Mean of lateral offset of trajectory :

[0041] In the formula, Indicates the first The distance by which the vehicle's trajectory deviates laterally.

[0042] Upstream and downstream related characteristic parameters Including behavioral consistency Degree of speed deviation wait: Behavioral consistency :

[0043] In the formula, The overlap length between the upstream and downstream trajectories. The total length of the upstream and downstream trajectories. exist Within the interval, the larger the value, the higher the similarity.

[0044] Speed ​​deviation :

[0045] In the formula, The number of vehicles that decelerated abnormally. For the first The vehicle's speed after decelerating upstream. This represents the speed of the vehicle downstream.

[0046] S402. A high-dimensional trajectory behavior feature vector is constructed using a Multilayer Perceptron (MLP). The MLP maps low-dimensional trajectory feature parameters to a high-dimensional trajectory behavior feature vector, which includes the upstream avoidance behavior vector. Downstream recovery behavior vector Upstream and downstream coupling behavior vector All values ​​are obtained through two layers of forward propagation of a multilayer perceptron (MLP). The weight matrix and bias vector of the MLP are determined following the logic of "initialization - model training - obtaining the optimal value". Upstream avoidance behavior vector : From upstream trajectory feature parameters The upstream avoidance behavior pattern of the vehicle is obtained through linear transformation by MLP. Downstream recovery behavior vector : , from downstream trajectory feature parameters The recovered motion pattern after vehicle avoidance is obtained by linear transformation using MLP. Upstream and downstream coupling behavior vector : ,Depend on The high-dimensional input vector, after dimensional merging, is obtained through MLP linear transformation, representing the overall behavioral pattern of upstream and downstream vehicle interactions. The dimensions are much larger than .

[0047] S403, The classifier model determines the state of blind zone events and the upstream avoidance behavior vector. Downstream recovery behavior vector and upstream and downstream coupling behavior vectors By concatenating the features, a comprehensive trajectory behavior feature vector is obtained. ;Will Input XGBoost classifier model To obtain the probability of the existence of blind zone events .

[0048] Preset classification decision threshold ,when ≥ When, it is determined that a special traffic event exists within the blind spot; when < At that time, it was determined that there were no special traffic incidents within the blind spot, and We will continue to optimize the model through subsequent training to improve the accuracy of the discrimination.

[0049] S404. Event location reasoning based on upstream and downstream coupling, by calculating the vehicle deceleration amplitude per unit mileage. The percentage of vehicles returning to the straight lane Lane number of vehicles deviation rate Three indicators, combined with upstream and downstream coupling, are used to infer the specific location blocks of special traffic events in blind spots: Vehicle deceleration per unit distance

[0050]

[0051] In the formula, This indicates the speed at which the vehicle began to decelerate. This indicates the speed at which the vehicle comes to a stop and decelerates, or the speed at which it enters a blind spot. This indicates the distance the vehicle travels while decelerating.

[0052] The larger the value, the closer the event is to the start of the blind zone. If the value is smaller, the event is closer to the end of the blind zone.

[0053] Percentage of vehicles returning to the straight lane

[0054]

[0055] In the formula, This indicates the number of vehicles that changed lanes again after exiting the blind spot (based on vehicle data trajectory to determine whether a lane change occurred). This indicates the total number of vehicles in the downstream sensing area.

[0056] exist Within the section, a certain lane The higher the value, the greater the probability of a special traffic incident occurring in this lane.

[0057] Lane number of vehicles deviation rate

[0058]

[0059] In the formula, The vehicle deviation rate in a certain lane (upstream / downstream). This represents the total number of vehicles changing lanes upstream / downstream in a given lane. This represents the total number of vehicles entering / leaving the blind spot in the corresponding lane during the corresponding time period.

[0060] exist Within the interval, The larger the value and The smaller the value, the greater the probability of a special traffic incident occurring in that lane.

[0061] S5. Adaptive Training and Parameter Calibration of the Model: To ensure the recognition performance of the model during long-term operation, the special traffic event recognition model of this invention has an adaptive training function. The model parameters are continuously updated and calibrated through a combination of manual calibration and real-time verification. The specific process is as follows: S501. Training dataset construction: Through manual labeling, labeled sample data of "with special traffic events" and "without special traffic events" in the blind spot are obtained to form a labeled training dataset.

[0062] S502. Model pre-training: The model is pre-trained using supervised learning to establish the mapping relationship between trajectory behavior features and event existence state, and to complete the initial parameter setting of the model.

[0063] S503. Real-time identification and feedback data acquisition: After the model is put into operation, it performs online identification on the real-time collected vehicle trajectory data and outputs the results; at the same time, based on road patrol confirmation, data from newly added detection equipment or information reported manually, it obtains actual verification feedback data of blind spot events.

[0064] S504. Model parameter update and calibration: Compare the consistency between the model recognition results and the validation feedback results, and adjust the model's MLP weight matrix, bias vector, and classification decision threshold. Adaptive updates and calibrations are performed to continuously optimize the model's recognition accuracy and generalization ability over long-term operation.

[0065] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for inferring blind spot events using vehicle trajectory boundary distribution correlation, characterized in that, Includes the following steps: S1. For road sections containing blind spots in road perception, an upstream perception zone, a blind spot in road perception, and a downstream perception zone are set up. Vehicle perception data in the upstream and downstream perception zones are collected through roadside equipment. Traffic events in the blind spot are inferred from the trajectory information of any area in the upstream and downstream perception zones. The identification status of traffic events in the blind spot in road perception includes "there is a special traffic event" and "there is no special traffic event". S2. Preprocess the collected vehicle perception data, remove data that is irrelevant to the vehicle position and summarize the data that is related to the trajectory, use the spatiotemporal matching method to associate the trajectory, and stitch together the short trajectory data to obtain complete vehicle trajectory data containing vehicle position, speed and acceleration information. S3. Collect basic vehicle motion information, including traffic status information and abnormal behavior information, from complete vehicle trajectory data. At the same time, based on vehicle trajectory characteristics, preset the discrimination conditions for the existence of special traffic events in the blind spot. S4. Based on the statistical vehicle motion information, extract low-dimensional trajectory feature parameters, use multilayer perceptron (MLP) in machine learning to construct high-dimensional trajectory behavior feature vectors, input the high-dimensional trajectory behavior feature vectors into the classifier model to determine whether there are special traffic events in the blind spot, and infer the specific location block of the special traffic event in the blind spot based on the upstream and downstream coupling degree. S5. Adaptive training and parameter calibration are performed on the special traffic event recognition model to achieve long-term continuous optimization of the model's recognition performance.

2. The blind spot event inference method based on vehicle trajectory boundary distribution association according to claim 1, characterized in that, In step S1, the vehicle perception data includes vehicle motion state, trajectory type, and driving behavior information; the roadside equipment data includes at least one of the perception data of roadside radar data and vehicle-mounted GPS data; special traffic events are events that cause the road to be unusable, including traffic accidents and road debris.

3. The blind spot event inference method based on vehicle trajectory boundary distribution association according to claim 1, characterized in that, In step S2, the spatiotemporal matching method uses a unified time reference and spatial coordinates to align vehicle trajectories collected by different sampling methods and devices to the same spatiotemporal coordinate system, eliminating time deviation and spatial offset, and achieving accurate correspondence of vehicle trajectories in time and space; the vehicle trajectory includes information on position, speed and acceleration.

4. The blind spot event inference method based on vehicle trajectory boundary distribution association according to claim 1, characterized in that, In step S3, the basic state variables extracted from the vehicle's basic motion information include average vehicle speed, vehicle distance, number of vehicles in the lane, blind spot travel time, and acceleration / deceleration. The criteria for identifying special traffic events within the pre-defined blind spot are any of the following: ① The vehicle changes lanes before entering the blind spot, or the number of vehicles in the lane changes significantly; ② The vehicle decelerates before entering the blind spot and accelerates after leaving the blind spot.

5. The blind spot event inference method based on vehicle trajectory boundary distribution association according to claim 1, characterized in that, In step S4, the low-dimensional trajectory feature parameter is P[ ], including upstream trajectory feature parameters Downstream trajectory characteristic parameters upstream and downstream related characteristic parameters ; Among them, velocity characteristic parameters Including the average rate of change of velocity v. Abnormal deceleration frequency and the average of abnormal deceleration locations Longitudinal feature parameters Including the average vehicle distance change rate and the proportion of dangerous following behaviors ; lateral characteristic parameters Including the average lane change position Lane changing behavior ratio and the mean of lateral offset of the trajectory Upstream and downstream related characteristic parameters Including behavioral consistency and speed deviation .

6. The blind spot event inference method based on vehicle trajectory boundary distribution according to claim 5, characterized in that, In step S4, the high-dimensional trajectory behavior feature vector constructed using a multilayer perceptron (MLP) includes an upstream avoidance behavior vector. Downstream recovery behavior vector Upstream and downstream coupling behavior vector ;in, From upstream trajectory feature parameters Obtained through two layers of linear transformation by MLP, Downstream trajectory feature parameters Obtained through two layers of linear transformation by MLP, Depend on , and The high-dimensional input vector after dimensional merging is obtained through a two-layer linear transformation of the MLP; the weight matrix and bias vector in the MLP are determined by initialization-model training-optimization.

7. The blind spot event inference method based on vehicle trajectory boundary distribution association according to claim 6, characterized in that, In step S4, the upstream avoidance behavior vector Downstream recovery behavior vector Upstream and downstream coupling behavior vector By concatenating the features, a comprehensive trajectory behavior feature vector is obtained. ;Will Inputting the classifier model f(x) yields the probability of the existence of blind zone events. Preset classification decision threshold ,when ≥ If a special traffic event is detected in the blind spot, it is determined that there is one; otherwise, it is determined that there is no special traffic event.

8. The blind spot event inference method based on vehicle trajectory boundary distribution association according to claim 7, characterized in that, The classifier model uses the XGBoost classifier, and the classification decision threshold is [not specified]. Continuous optimization is achieved through model training.

9. The blind spot event inference method based on vehicle trajectory boundary distribution association according to claim 8, characterized in that, In step S4, the location range of special traffic events in the blind zone is inferred based on the upstream and downstream coupling degree, and the vehicle deceleration rate per unit mileage is calculated. The percentage of vehicles returning to the straight lane Lane number deviation rate Implementation: ① Used to determine the position of an event before or after it in the blind zone. The larger the value, the closer the event is to the start of the blind zone. The smaller the value, the closer the event is to the end of the blind zone; ② Used to determine the lane where the incident occurred. The higher the value, the higher the probability of a special traffic incident occurring in that lane; ③ Including the deviation rate of the number of vehicles in the upstream lane Deviation rate of vehicle count in downstream lane , The larger the value and The smaller the value, the greater the probability of a special traffic incident occurring in the corresponding lane.

10. The blind spot event inference method based on vehicle trajectory boundary distribution association according to claim 1, characterized in that, In step S5, the adaptive training and parameter calibration process of the special traffic event identification model includes the following steps: S501. Obtain labeled sample data of "with special traffic events" and "without special traffic events" in the blind spot through manual calibration to form a labeled training dataset; S502. Use supervised learning to pre-train the model and establish a mapping relationship between trajectory behavior features and event existence state; S503: The model identifies real-time trajectory data and outputs results, while obtaining verification feedback data based on patrol confirmation, equipment detection, or manual reporting information; S504. Based on the consistency between the model recognition results and the verification feedback results, adaptively update and calibrate the model's weight matrix, bias vector, and classification decision threshold τ.

Images