Discrimination And Classification Of Traffic Conflicts At Level Intersections Based On Aerial Image

With the continuous improvement of road facilities and the increase of motor vehicle ownership year by year,vehicle driving safety has become a hot issue of people and social concern,there are a large number of motor vehicles in the city roads,traffic accidents always occur.Intersections,as the area where traffic crosses,it has the characteristics of right-of-way distribution,alternating traffic flow in different directions,the operation condition is more complex,it becomes the accident-prone section of urban roads,so many scholars have done a lot of research on safe passage at intersections.Traffic conflict as an indirect evaluation method has been used and expanded by many scholars,but the current road safety situation has put forward new requirements for the comprehensiveness,convenience,and accuracy of conflict,and the development of existing image detection and computer processing technology is expected to enhance the collection of conflict-related data and conflict discrimination methods.In this paper,we mainly focus on the traffic conflicts between vehicles at planar intersections.Firstly,we use the characteristics of high accuracy and simple deployment of UAV photography to collect the original data on the operation of intersections.Then we choose R3 Det network to detect vehicles for the characteristics of variable vehicle scale and angle in aerial images,and take the regression method of combining horizontal anchor frame and rotating anchor frame,and this method has better effect on different angle regression.In this paper,we redesign the anchor frame size according to the aspect ratio of different vehicles,and compare the network performance with the selected single-stage network and the two-stage network for aerial images,the results show that the network has better performance in both detection accuracy and speed.On the basis of obtaining the target position in the image,we correlate the target position based on the detection-tracking strategy,establish the feature set of target motion and appearance.We use Kalman filter to predict and update the target position,and motion features combines with color features to match the prediction result with the detection result,so as to realize the extraction of the original target trajectory.When the original trajectory is obtained,we correct the real relative distance between targets by coordinate transformation,and finally process the trajectory by using mean filtering to reduce the relative offset of target detection position affected by detection accuracy,and extract the parameters related to conflict discrimination on this basis.In the process of conflict discrimination,the traditional method is mainly based on the proximity and reduction speed reaches a certain degree trend of the conflict object,and the conflict is defined by manual ranking.Through the analysis of the definition of conflict,we believe that the conditions of conflict determination are mainly divided into the existence of the possibility of collision and the change of motion state.After that,we obtain the specific time and location of the conflict by the sudden change of motion state,and the sudden change of motion state is judged by the rate of change of deceleration and the rate of change of angular velocity,and the conflict is considered as a continuous process,which includes the comprehensive determination of the start and end time of the conflict.In addition,the severity of conflict was classified based on the abrupt movement state and the continuous urban construction of conflict,which provided a data basis for the evaluation and further study of intersection safety.