Research On Traffic Conflict Analysis And Prediction At Intersection Based On Trajectory Extraction

In recent years,with the continuous implementation of the "Green Travel" concept,the travel modes of residents have shown a diversified trend.While traditional motor vehicles are still widely used,non-motor vehicles,especially convenient shared bicycles and electric bicycles,are favored by residents traveling in short or medium distance.Grade intersection is the key node of road traffic.A variety of traffic flows gather here and form traffic bottleneck.The bad habits of some traffic participants,such as running through red light and Preemptive driving,lead to frequent traffic conflicts and even traffic accidents.In order to improve the intersection safety,based on traffic conflict technology and machine learning algorithm,this paper studies the conflicts between right turn motor vehicles and non-motor vehicles and conflicts between left turn vehicle and going straight vehicle at grade intersections.Firstly,the trajectory data of traffic conflict is obtained.Through the actual investigation and shooting of the conflict video at the intersection,PFTrack is used to extract the trajectory pixel coordinates of both sides of the conflict.Then the homogeneous coordinate conversion method is used to obtain the real-world coordinates of the trajectory.The accuracy analysis shows that the distance error after conversion is no more than 1%.According to the trajectory data,the changes of motion parameters in the process of traffic conflict are analyzed.Secondly,the severity of traffic conflict is classified.From the four dimensions,including time to collision,post-encroachment time,proportion of stopping distance and deceleration rate to avoid crash,an index system to measure traffic conflict severity is established,and each index is calculated based on trajectory data.Then K-medoids clustering algorithm is used to divide the conflict samples into mild conflict,general conflict and serious conflict.By calculating the Adjusted Rand Index and Silhouette Coefficient,it is verified that K-medoids clustering effect is good.Then,the influence factors of traffic conflict are analyzed.From four perspectives,including intersection,motor vehicle,non-motor vehicle and conflict scene,13 factors that may affect the traffic conflict severity are put forward and quantified.Using the random forest model,according to the contribution of each factor in the classification of traffic conflict severity,the factors are divided as strong impact factors,weak impact factors and basically no impact factors.The strong impact factors and weak impact factors are statistically analyzed.Finally,the traffic conflict severity is predicted.The prediction object is the trajectory of the turning vehicle.Setting the spatial position,the motion parameters,turning reference point and strong impact factors of the first three moments as the input and the spatial position of the current moment as the output,the Long Short-Term Memory(LSTM)model is established and optimized.Then,the severity of the conflict is predicted.Using the continuous prediction method,the turning vehicle trajectory with default information is predicted precisely.Combined with the existing clustering research,the severity of each conflict is predicted.The results show that the prediction accuracy of the severity of the conflict between right turn vehicle and non-motor vehicle and the conflict between left turn vehicle and going straight vehicle is 84.64% and91.49% respectively.This paper can provide theoretical support for reducing the occurrence of traffic conflicts of turning vehicle and the conflicts severity,and improving the safety of motor vehicles and non-motor vehicles passing through intersections.