Research On Vehicle Fusion Ranging Algorithm Based On Monocular Vision In Dynamic Driving Scenes

Real-time and accurate perception of the environment in dynamic driving scenes is the crucial technology for advanced driving assistant system and autonomous driving.Due to the rich information and low price of vision sensors,most of the environment sensing technologies are based on vision.However,in dynamic driving scenarios,the attitude angle change of the selfvehicle and the dynamic motion of the perception objects will affect the visual perception algorithm results.Therefore,how to make full use of video stream information to achieve highprecision vehicle perception in dynamic driving scenes has become the significant research content of vision-based vehicle perception algorithms.In order to improve the accuracy of vehicle perception in dynamic driving scenes,this paper takes vehicles as the research object and carries out related researches on vehicle ranging algorithms based on monocular vision.The main research contents of this paper are as follows:(1)Construct Light-YOLO vehicle detection network.To improve the running speed of the vehicle detection algorithm on the in-vehicle embedded platform,this paper uses a lightweight feature extraction network based on depthwise separable convolution to replace YOLOv3's backbone network Darknet53 and constructs the Light-YOLO network.Furthermore,by using complete intersection over union(CIoU)to improve the loss function,using K-means++ algorithm to cluster the anchor boxes,and improving the non-maximum suppression(NMS)algorithm,the position prediction accuracy of the vehicle detection bounding boxes is improved.(2)Propose a vehicle tracking algorithm based on the spatiotemporal feature association.To correlate the vehicle information in the video stream and realize spatiotemporal continuous vehicle perception,this paper proposes a vehicle feature association algorithm that combines IoU,Mahalanobis distance and color histogram.The vehicle is tracked by Kalman filtering based on a linear constant velocity model,and the vehicle detection and the tracking bounding boxes is matched through the Hungarian algorithm.This algorithm reduces vehicle tracking missing and ID switching,improves the robustness of vehicle tracking.(3)Propose a vehicle fusion ranging algorithm based on multi-dimensional reference information.To reduce the interference of the dynamic change of the camera attitude angle and the vehicle detection bounding boxes on the ranging algorithm,this paper uses a fast vanishing point detection algorithm based on line segment detection and voting to detect the vanishing point of the road and estimate the camera attitude angles,and a pinhole distance estimation model modified the camera attitude angle is proposed.Subsequently,a multi-dimensional information reference fusion ranging algorithm is proposed.The distance is measured by the improved pinhole model,then the width of the tracked vehicle is calculated using the distance estimation result,and then the relative distance is calculated using the distance estimation model based on the vehicle width.Finally,spatiotemporal continuous vehicle ranging based on vehicle tracking and different reference information is realized.(4)Build the vehicle experimental platform based on the NVIDIA Jetson Xavier to verify the accuracy and operating efficiency of the proposed vehicle detection,tracking and fusion ranging algorithm based on the monocular vision in dynamic driving scenes.Compared with the YOLOv3 network,the average detection precision and recall of the Light-YOLO network are reduced by 3.62% and 4.3%,and the average calculation time per frame is reduced by 12.4ms.Compared with the tracking algorithm named SORT,the MOTA of the proposed tracking algorithm is increased by 2.7%,and the number of ID switching is reduced by 41%.The relative error of the improved pinhole distance estimation model is less than 6%.The result of the fusion ranging algorithm is more stable,and the average operating speed can reach 25 frames per second,which can run on the in-vehicle embedded platform in real-time and meet the needs of vehicle perception in dynamic driving scenes.