Traffic incident detection method based on intelligent network connection

Abstract

The invention discloses a traffic incident detection method based on intelligent network connection, and belongs to the technical field of traffic incident detection. Comprising the following steps: acquiring meteorological data, road topology data, roadside sensing data of a roadside unit and vehicle data of a plurality of networked vehicles in a target area; wherein the vehicle data comprises vehicle microscopic behavior data and inter-vehicle interaction data; obtaining a spatial-temporal feature vector of the target area according to the vehicle data, the meteorological data and the road topology data; if a single vehicle signal reported by the networked vehicle is received, the networked vehicle reporting the single vehicle signal is used as an abnormal vehicle, multi-level verification is performed on the single vehicle signal according to the vehicle data and the roadside sensing data, and a verification result is obtained; inputting the verification result and the spatial-temporal feature vector of the abnormal vehicle into a pre-trained lightweight event classification model to obtain an event classification result; and generating standardized information according to an event classification result, and pushing the standardized information to related objects in a differentiated manner.

Description

Technical Field

[0001] This invention belongs to the field of traffic incident detection technology, specifically relating to a traffic incident detection method based on intelligent connected vehicles. Background Technology

[0002] With the acceleration of urbanization, traffic congestion and frequent safety accidents have become serious social problems. Rapid and accurate detection of traffic incidents (such as accidents, breakdowns, congestion, and dangerous driving) is a key prerequisite for effective traffic management and improving road safety and efficiency. Currently, traffic incident detection technology has undergone several generations of development. Traditional detection methods mainly rely on fixed infrastructure, which has obvious technical limitations and is difficult to adapt to the complex and ever-changing needs of modern traffic scenarios. The first generation of technology, represented by inductive detectors such as loop coils and microwave radar, has a fixed detection range, limited functionality, high failure rate, and cannot acquire rich vehicle behavior information. The second generation of technology is based on video surveillance and analyzes road conditions through computer vision algorithms; however, this method is severely limited by environmental factors such as lighting, weather, and obstruction, with performance deteriorating sharply at night and in rainy or foggy weather. It also suffers from limitations in installation perspective, high computational complexity, and privacy risks.

[0003] With the continuous development of intelligent transportation technology, various improved traffic incident detection technologies have emerged in the industry. However, none of them have fundamentally solved the core defects of traditional technologies, as follows: Existing Technology 1: Based on fixed detectors, abnormal events are detected by installing cameras on the roadside and using image recognition algorithms. Its defects are: greatly affected by weather, lighting, and obstruction; limited coverage (only within the camera's field of view); and a simple event judgment logic, resulting in a high false alarm rate. Existing Technology 2: Based on floating car data, road conditions are inferred from the trajectory data of some vehicles. Its defects are: single data source; low update frequency; insufficient sample representativeness (not all vehicles), leading to high detection latency and poor sensitivity to small-scale events. Existing Technology 3: Traditional vehicle-road cooperative simple reporting; some patents propose vehicles actively reporting abnormal events (such as emergency braking). Its defects are: reliance on the instantaneous judgment of a single vehicle; lack of cross-validation of multi-source data; prone to generating a large number of false alarms due to false alarms from vehicle sensors or differences in driving behavior; and low system reliability.

[0004] Therefore, there is an urgent need for a traffic incident detection method based on intelligent connected vehicles to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention provides a traffic incident detection method based on intelligent connected vehicles to solve the limitations of existing technologies, such as single data source and single judgment logic, as well as the problems of high false alarm rate and high false alarm rate.

[0006] To achieve the above objectives, the present invention provides a traffic incident detection method based on intelligent connected vehicles, comprising the following steps: Acquire meteorological data, road topology data, roadside perception data of roadside units, and vehicle data of multiple connected vehicles within the target area; among which, vehicle data includes: vehicle micro-behavior data and inter-vehicle interaction data; The spatiotemporal feature vector of the target area is obtained based on vehicle data, meteorological data, and road topology data; If a single vehicle signal is reported by a connected vehicle, the connected vehicle that reported the single vehicle signal will be regarded as an abnormal vehicle. The single vehicle signal will be verified at multiple levels based on vehicle data and roadside perception data to obtain the verification result. The verification results and the spatiotemporal feature vectors of the abnormal vehicles are input into a pre-trained lightweight event classification model to obtain the event classification results; the event classification results include: type, confidence level, and impact range assessment. Standardized information is generated based on the event classification results and pushed to relevant parties in a differentiated manner. These relevant parties include: traffic management centers, rescue units, and connected vehicles located upstream of the abnormal vehicle.

[0007] As an embodiment of the present invention, a spatiotemporal feature vector of a target area is obtained based on vehicle data, meteorological data, and road topology data, including: Vehicle data, meteorological data, and road topology data are mapped to the same spatiotemporal coordinate system; Feature extraction is performed on the mapped vehicle data, meteorological data, and road topology data to obtain spatiotemporal feature vectors.

[0008] As an embodiment of the present invention, multi-level verification of a single vehicle signal is performed based on vehicle data and roadside perception data to obtain verification results, including: If a single vehicle signal is reported by a connected vehicle, vehicle data of connected vehicles adjacent to the abnormal vehicle is obtained. Determine if there are at least a preset number of adjacent connected vehicles with abnormal correlation; if not, determine that the individual vehicle signal of the abnormal vehicle is invalid. If present, the single-vehicle signal of the abnormal vehicle is deemed valid. Spatial verification of the abnormal vehicle is performed based on roadside perception data to obtain the verification result.

[0009] As one embodiment of the present invention, standardized information is generated based on the event classification results and pushed to relevant objects in a differentiated manner, including: Generate standardized information based on the event classification results; It sends an alarm to the traffic management center, sends detailed incident information to the rescue unit, and sends a warning to connected vehicles upstream of the abnormal vehicle.

[0010] The beneficial effects of this invention are as follows: 1. By using a three-level cross-verification mechanism of "single vehicle triggering + neighboring vehicle verification + roadside confirmation", the problem of high false alarm rate in traditional methods or simple reporting is fundamentally solved.

[0011] 2. Full-time and high real-time coverage: Utilizing mobile connected vehicles as "mobile sensors" breaks the geographical limitations of fixed detection equipment, achieving full-segment coverage of roads without blind spots. Data is communicated directly via C-V2X with extremely low latency (milliseconds), representing a significant improvement in real-time performance compared to methods relying on floating car GPS data transmission (minutes).

[0012] 3. Strong anti-interference capability: It integrates vehicle micro-behavior with roadside macro-perception, and is not sensitive to adverse weather (such as fog and rain) and changes in lighting (night), overcoming the inherent defects of pure video detection solutions.

[0013] 4. High level of intelligence: It adopts a hybrid architecture of "rule initial screening + model fine judgment". The model can learn the event characteristics in complex scenarios. It is more adaptable than methods that simply rely on rule thresholds (such as traditional vehicle-road cooperative patents) and can distinguish more detailed event types.

[0014] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0015] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0016] like Figures 1-2 As shown, this invention provides a traffic incident detection method based on intelligent connected vehicles, comprising: Acquire meteorological data, road topology data, roadside perception data of roadside units, and vehicle data of multiple connected vehicles within the target area; among which, vehicle data includes: vehicle micro-behavior data and inter-vehicle interaction data; The spatiotemporal feature vector of the target area is obtained based on vehicle data, meteorological data, and road topology data; If a single vehicle signal is reported by a connected vehicle, the connected vehicle that reported the single vehicle signal will be regarded as an abnormal vehicle. The single vehicle signal will be verified at multiple levels based on vehicle data and roadside perception data to obtain the verification result. The verification results and the spatiotemporal feature vectors of the abnormal vehicles are input into a pre-trained lightweight event classification model to obtain the event classification results; the event classification results include: type, confidence level, and impact range assessment. Standardized information is generated based on the event classification results and pushed to relevant parties in a differentiated manner. These relevant parties include: traffic management centers, rescue units, and connected vehicles located upstream of the abnormal vehicle.

[0017] Specifically, through vehicle-to-everything (C-V2X) wireless communication technology, heterogeneous time-series data is continuously collected from multiple connected vehicles and roadside units within the target area, including the following data: 1. Vehicle micro-behavioral data: from onboard sensors, including but not limited to: vehicle ID, location, speed, acceleration (especially rapid acceleration / deceleration), steering wheel angle (sharp steering), braking status, light status (hazard lights), and airbag deployment signal; 2. Vehicle-to-vehicle interaction data: vehicle-to-vehicle (V2V) communication messages, such as vehicle status and forward collision warning in basic safety messages; 3. Roadside perception data: from roadside unit (RSU) fusion perception, including holographic traffic flow status, whether there are stationary obstacles in specific areas, and information such as pedestrians or non-motorized vehicles entering motor vehicle lanes; 4. Meteorological data: specific regional meteorological data from meteorological departments, including information such as illumination, rainfall, temperature, and wind speed; 5. Road topology data: for events within the high-precision map area, detailed road topology data can be obtained through the high-precision map to accurately locate the detailed location information of the event, down to the lane level.

[0018] This technical solution does not rely on a single data source or simple reporting, but instead constructs a dynamic decision-making model that is collaborative across "end-edge-cloud" and undergoes multi-level cross-validation. Its core lies in: aggregating and deeply mining the full amount of real-time interaction data between intelligent connected vehicles and roadside intelligent units (RSUs), and automatically identifying genuine abnormal traffic events from massive amounts of normal data through spatiotemporal alignment, feature fusion, and rule-model hybrid reasoning.

[0019] The spatiotemporal feature vector of the target area is obtained based on vehicle data, meteorological data, and road topology data, including: Vehicle data, meteorological data, and road topology data are mapped to the same spatiotemporal coordinate system; Feature extraction is performed on the mapped vehicle data, meteorological data, and road topology data to obtain spatiotemporal feature vectors.

[0020] Specifically, data from different sources and timestamps are uniformly mapped to the same spatiotemporal coordinate system. Spatiotemporal feature vectors are extracted for each target area (e.g., road segment, intersection). These feature vectors include: 1. Behavioral anomaly index: based on vehicle micro-behavioral data, the proportion and intensity of vehicles engaging in sudden braking, sharp turns, and abnormal parking within a unit of time; 2. Trajectory conflict density: based on the predicted vehicle trajectory and real-time location, the density of potential conflict points (e.g., rear-end collision, side-impact risk); 3. State change consistency: analyzing whether multiple adjacent vehicles experience state changes (e.g., continuous sudden braking) at similar times and locations, which is crucial for distinguishing between individual abnormal behavior and group events (e.g., accidents). In this invention, for each target area (e.g., a circular area with a radius of 200 meters centered on a roadside unit, or a road segment with a length of 500 meters), within each time window (e.g., 1 second), the system calculates a spatiotemporal feature vector based on aligned multivariate data (vehicle data, meteorological data, and road topology data). This vector consists of the following three core dimensions and their sub-features: Behavioral Abnormality Index (BAI) This index is used to quantify the degree of abnormality in vehicle driving behavior within a target area. It is obtained by fusing statistical characteristics of abnormal behavior of individual vehicles with weighted aggregation. The specific calculation steps are as follows: (1) Detection of abnormal behavior of a single vehicle: For each connected vehicle in the area, determine in real time whether it has triggered the following three types of abnormal events: Sudden deceleration event: vehicle deceleration Exceeding the preset deceleration threshold (e.g., -3.5 m / s²) is marked as a rapid deceleration; Sudden steering event: When the vehicle accelerates laterally Exceeding the preset lateral acceleration threshold (e.g., 3.0 m / s²) or steering wheel angular velocity Exceeding the preset steering wheel angular velocity threshold (e.g., 100° / s) is marked as a sharp turn; Abnormal parking event: When the vehicle speed Less than Furthermore, the hazard lights activation or engine shutdown signal lasts for a duration exceeding the preset dwell time. If the dwell time is 10 seconds or less, mark it as abnormal. (2) Regional anomaly count: within the time window Within the target area, count the number of vehicles that experienced any of the above-mentioned abnormal events. and the total number of vehicles in the target area. At the same time, record the severity weight of all abnormal events within the window (e.g., rapid deceleration weight = 1, sharp turn weight = 1.5, abnormal dwell weight = 3.0).

[0021] (3) Formula for calculating the index: In the formula, This is a normalization constant (e.g., 3.0). The sum of the weights of all exception events within the window. For the first The weight of each abnormal event, This is an index of abnormal behavior. The value range is [0, 1], and the larger the value, the more abnormal the driving behavior in the area.

[0022] Trajectory Conflict Density (TCD) This metric measures the density of potential collision risks between vehicle trajectories within a region. Based on real-time vehicle position and velocity vectors, it predicts future short-term trajectories and calculates the number of collision points. The specific implementation method is as follows: (1) Vehicle motion prediction: For each vehicle pair Assuming it is predicted in the future Calculate the minimum distance between the two trajectories while maintaining the current speed and heading angle for 2 seconds. and the arrival time to reach that distance. ; (2) Conflict detection: If there is any arrival time Make ( For a safe distance, for example Rice, of which, For vehicles China Vehicle speed, For vehicles China Vehicle If the speed is such that a potential conflict occurs, then it is considered a potential conflict. (3) Density calculation: statistical time window Within the region, the total number of potential conflicts between all vehicles. Based on the total number of potential conflicts The trajectory collision density is calculated as follows: The trajectory conflict density is calculated by dividing the number of vehicle pairs with potential conflicts by the maximum number of possible conflict pairs, resulting in a normalized trajectory conflict density. This represents the total number of vehicles within the target area.

[0023] State mutation consistency (SMC) This indicator is used to identify whether multiple vehicles experience abrupt state changes within a similar time and space, thereby distinguishing between individual anomalies and group events (such as chain-reaction rear-end collisions). The calculation method is as follows: (1) Vehicle state change flag: For each vehicle, if its speed decreases by more than a preset speed decrease threshold within 0.5 seconds, the flag is issued. (e.g. 8) ), or deceleration The absolute value exceeds the preset deceleration threshold. If so, then the vehicle has undergone a "state change"; (2) Spatiotemporal clustering: Input all event points (location coordinates + timestamp) marked as state abrupt changes into the DBSCAN clustering algorithm, with parameters set as follows: neighborhood radius Meters, time Seconds, minimum neighborhood number .

[0024] (3) Calculation of the consistency index: , This represents the number of mutant vehicles in the largest cluster after clustering. For state change consistency, i.e., the proportion of vehicles in the largest cluster to the total number of vehicles in the region, The higher the value, the more likely it is that multiple vehicles experienced sudden behavioral changes at almost the same time in similar locations, which is most likely caused by a traffic incident.

[0025] Final spatiotemporal feature vector: For each detection region and each time window, the system outputs a three-dimensional vector. The input to the classification model consists of the input features, as well as optional supplementary features (such as average velocity, density, timestamps, etc.).

[0026] Among these methods, multi-level verification of single-vehicle signals is performed based on vehicle data and roadside perception data to obtain verification results, including: If a single vehicle signal is reported by a connected vehicle, vehicle data of connected vehicles adjacent to the abnormal vehicle is obtained. Determine if there are at least a preset number of adjacent connected vehicles with abnormal correlation; if not, determine that the individual vehicle signal of the abnormal vehicle is invalid. If present, the single-vehicle signal of the abnormal vehicle is deemed valid. Spatial verification of the abnormal vehicle is performed based on roadside perception data to obtain the dynamic verification result.

[0027] Specifically, if a single-vehicle signal is reported by a connected vehicle, the connected vehicle that reported the signal is considered an abnormal vehicle. The single-vehicle signal is an abnormal signal actively reported by the receiving vehicle (such as an emergency braking warning or collision signal). Then, neighboring vehicles (connected vehicles adjacent to the abnormal vehicle within the target area) collaboratively verify the signal. Upon receiving a single-vehicle signal, real-time data from N (N≥2) neighboring vehicles around the incident location is immediately retrieved. If more than M (M≥1, configurable) neighboring vehicle data show abnormal correlation (e.g., neighboring vehicles in the same direction also exhibit sudden braking or detour trajectories; vehicles in the opposite lane slow down and observe), the single-vehicle signal is preliminarily confirmed as valid, filtering out false alarms caused by sensor malfunctions or driver error. Finally, roadside global perception verification is performed using roadside perception data from roadside units (such as camera and millimeter-wave radar fusion) to visually / spatially confirm the initially screened event, obtaining the verification result. For example, this verifies whether there is physical evidence such as stationary vehicles, scattered objects, or crowds at the incident location.

[0028] The verification results and the spatiotemporal feature vectors of the abnormal vehicles are input into a pre-trained lightweight event classification model to obtain the event classification results. The event classification results include: type, confidence level, and impact range assessment. Standardized information is generated based on the event classification results and then pushed to relevant individuals in a differentiated manner, including: Generate standardized information based on the event classification results; It sends an alarm to the traffic management center, sends detailed incident information to the rescue unit, and sends a warning to connected vehicles upstream of the abnormal vehicle.

[0029] Specifically, the extracted spatiotemporal feature vectors are input into a pre-trained lightweight event classification model (small neural network); the model outputs the event type, confidence level, and predicted scope of impact.

[0030] Then, the confirmed event information (type, location, confidence level, recommendations) is generated into standardized messages and pushed to relevant parties in a differentiated manner through RSU or cellular networks; that is, warnings are sent to upstream vehicles of the event (suggesting slowing down or changing lanes); alarms and handling suggestions are sent to the traffic management center; and event details are sent to emergency / rescue units.

[0031] Secondly, the verification results and the spatiotemporal feature vectors of the abnormal vehicles are input into a pre-trained lightweight event classification model to obtain the event classification results; the event classification results include: type, confidence level and impact range assessment.

[0032] The specific implementation process of the lightweight event classification model is as follows: The "pre-trained lightweight event classification model" described in this invention is implemented using the gradient boosting tree (LightGBM) model.

[0033] 1. Training dataset construction (1) Data source: positive samples were extracted from historical real traffic incident records, and negative samples were randomly sampled from normal traffic flow; among them, positive samples include: rear-end collisions, breakdowns, pedestrians running over, road debris, etc.; negative samples are ordinary traffic flow segments in which no incidents occurred.

[0034] (2) Sample labeling: Each sample consists of the following three parts: Input features: the spatiotemporal feature vectors extracted in the preceding steps. And additional features: regional average velocity velocity variance Vehicle density 1. Whether a single vehicle signal exists (Boolean value); 2. Whether the single vehicle signal is valid when verified by a neighboring vehicle (Boolean value); 3. Roadside verification result (Boolean value). Output labels: Event type (encoded as an integer: 0-no event, 1-rear-end collision, 2-breakdown, 3-congestion, 4-dangerous driving), confidence level (0-low, 1-medium, 2-high), and range of influence (numerical value, in meters).

[0035] 2. Model Training Process (1) Feature preprocessing: Z-score standardization is performed on all continuous features to make the mean 0 and the variance 1; Boolean features are not processed.

[0036] (2) Model parameter configuration.

[0037] (3) Training: Supervised learning is performed using the training set to minimize the multi-class cross-entropy loss function. The training is evaluated on the validation set every 10 rounds. Training is stopped when the validation set loss does not decrease for 20 consecutive rounds (early stopping method).

[0038] (4) After training is completed, the model file is obtained and deployed on the edge computing server.

[0039] The beneficial effects of this technical solution are: 1. By adopting a three-level cross-verification mechanism of "single vehicle triggering + neighboring vehicle verification + roadside confirmation", the problem of high false alarm rate in traditional methods or simple reporting is fundamentally solved.

[0040] 2. Full-time and high real-time coverage: Utilizing mobile connected vehicles as "mobile sensors" breaks the geographical limitations of fixed detection equipment, achieving full-segment coverage of roads without blind spots. Data is communicated directly via C-V2X with extremely low latency (milliseconds), representing a significant improvement in real-time performance compared to methods relying on floating car GPS data transmission (minutes).

[0041] 3. Strong anti-interference capability: It integrates vehicle micro-behavior with roadside macro-perception, and is not sensitive to adverse weather (such as fog and rain) and changes in lighting (night), overcoming the inherent defects of pure video detection solutions.

[0042] 4. High level of intelligence: It adopts a hybrid architecture of "rule initial screening + model fine judgment". The model can learn the event characteristics in complex scenarios. It is more adaptable than methods that simply rely on rule thresholds (such as traditional vehicle-road cooperative patents) and can distinguish more detailed event types.

[0043] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A traffic incident detection method based on intelligent connected vehicles, characterized in that, Includes the following steps: Acquire meteorological data, road topology data, roadside perception data of roadside units, and vehicle data of multiple connected vehicles within the target area; among which, vehicle data includes: vehicle micro-behavior data and inter-vehicle interaction data; The spatiotemporal feature vector of the target area is obtained based on vehicle data, meteorological data, and road topology data; If a single vehicle signal is reported by a connected vehicle, the connected vehicle that reported the single vehicle signal will be regarded as an abnormal vehicle. The single vehicle signal will be verified at multiple levels based on vehicle data and roadside perception data to obtain the verification result. The verification results and the spatiotemporal feature vectors of the abnormal vehicles are input into a pre-trained lightweight event classification model to obtain the event classification results; the event classification results include: type, confidence level, and impact range assessment. Standardized information is generated based on the event classification results and pushed to relevant parties in a differentiated manner. These relevant parties include: traffic management centers, rescue units, and connected vehicles located upstream of the abnormal vehicle.

2. The traffic incident detection method based on intelligent connected vehicles according to claim 1, characterized in that, Based on vehicle data, meteorological data, and road topology data, a spatiotemporal feature vector of the target area is obtained, including: Vehicle data, meteorological data, and road topology data are mapped to the same spatiotemporal coordinate system; Feature extraction is performed on the mapped vehicle data, meteorological data, and road topology data to obtain spatiotemporal feature vectors.

3. The traffic incident detection method based on intelligent connected vehicles according to claim 1, characterized in that, Based on vehicle data and roadside perception data, multi-level verification of single-vehicle signals is performed to obtain verification results, including: If a single vehicle signal is reported by a connected vehicle, vehicle data of connected vehicles adjacent to the abnormal vehicle is obtained. Determine if there are at least a preset number of adjacent connected vehicles with abnormal correlation; if not, determine that the individual vehicle signal of the abnormal vehicle is invalid. If present, the single-vehicle signal of the abnormal vehicle is deemed valid. Spatial verification of the abnormal vehicle is performed based on roadside perception data to obtain the verification result.

4. The traffic incident detection method based on intelligent connected vehicles according to claim 1, characterized in that, Standardized information is generated based on the event classification results and then pushed to relevant individuals in a differentiated manner, including: Generate standardized information based on the event classification results; It sends an alarm to the traffic management center, sends detailed incident information to the rescue unit, and sends a warning to connected vehicles upstream of the abnormal vehicle.

Images