Research On Visual Vehicle Detection And Ranging Algorithm For Embedded Lane Change Decision Aid System

Lane change is a common driving behavior,which involves lateral and longitudinal movement of vehicles.In the process of changing lanes,drivers often need to pay close attention to the target lane and the driving environment ahead,which is easy to cause traffic accidents due to negligence.The Advanced Driving Assistant System(ADAS),represented by Lane Change Decision Aid System(LCDAS),can use on-board sensors to accurately perceive road environment in real time,which is conducive to the detection of potential dangers,and timely warning to drivers or take active measures to avoid the occurrence of traffic accidents,has become an important guarantee for driving safety.This paper takes the environmental perception of LCDAS as the research content,and mainly studies the visual vehicle object detection,tracking and ranging algorithm.The detailed work of this paper includes:(1)The Enhancement-tiny YOLOv3 vehicle detection algorithm is proposed.In order to improve the precision of Tiny YOLOv3 model in detecting small vehicle objects,this paper changes the maxpool3 layer of Tiny YOLOv3 to conv4,and adds conv5 layer to compress the feature channel dimension of conv4 layer to reduce invalid parameters,and connects the feature channel dimension of conv6 layer and the new upsample2 layer as the feature branch layer of the feature pyramid network.At the same time,k-means algorithm is used to change the number and size of anchor boxes and the flow chart of vehicle detection algorithm is introduced in detail.(2)Research on vehicle tracking techniques.Without using image information,this paper proposes a vehicle tracking algorithm based on Kalman filtering from the point of motion modeling theory.Assuming that the vehicle object moves uniformly between video sequences,the corresponding relationship between tracking and detection bounding boxes is established according to the Intersection over Union(IOU)of the detection and tracking bounding boxes and Hungarian matching algorithm,and the Kalman filtering algorithm is used to predict and update the position of different matching types of vehicle objects.(3)Research on vehicle ranging techniques based on monocular vision.In this paper,the camera yaw angle and pitch angle are estimated by road vanishing point detection algorithm.The change of pitch angle is taken into account in the Inverse Perspective Mapping(IPM)to establish the dynamic IPM model.The geometric model of camera yaw angle in the top down view of dynamic IPM is established,and the basic ranging model of LCDAS is finally established.Further,according to different installation positions of the side rear and front mounted cameras,judge the orientation of the vehicle object and use the corresponding location information to measure distance.(4)The vehicle detection,tracking and ranging algorithms are tested and validated respectively.This paper builds a vehicle test platform based on the NVIDIA Jetson Xavier embedded development board and 3 USB cameras to verify the feasibility and effectiveness of the vehicle detection,tracking and ranging algorithms.Experiments show that compared with Tiny YOLOv3 model,without tracking algorithm,the mean precision and recall rate of the Enhancement-tiny YOLOv3 model is improved by 4.6% and 7.4% respectively;after adding tracking algorithm,the mean precision and recall rate of the fusion algorithm is improved by 10.6% and 23.6% respectively;with the addition of ranging algorithm,the dynamic and static errors of proposed ranging model are within 7%,and the mean operation speed reaches 28 frames per second.The results show that the proposed algorithm can meet the requirements of real-time and accuracy of embedded LCDAS.