Vehicle Detection And Motion Trend Estimation Based On Components

Vehicle detection and motion trend analysis are the key links in the application research of driverless vehicle,and they are the foundation and guarantee of road environment perception and motion planning decisions.However,the vehicle occlusion and other problems often encounter in the detection process that affect the accuracy of vehicle detection and the reliability of collision threat situation analysis to a certain extent,thus threatening the safe driving of unmanned vehicles.In view of this,this paper focuses on the vehicle detection problem in the case of near occlusion,and conducts in-depth research on vehicle detection methods and vehicle motion parameter estimation methods based on vehicle components,and proposes methods to solve the problem.The main work of this paper is as follows:1.A vehicle component detection method based on YOLOv3 network is proposed.Firstly,According to the different perspectives of the vehicle,eight orientations are defined,and for each orientation,different types of large and small parts of the vehicle are defined and marked.Based on this,a image dataset of vehicle detection based on vehicle parts is constructed;Then,an anchor box selection algorithm based on K-means++ is given,and the anchor box in the YOLOv3 network is reallocated;Finally,through multiple multi-scale training of the YOLOv3 network model,an ideal result of vehicle component detection is obtained;At the same time,through the comparative experiment,we analyse the improvement of the small component detection effect based on the anchor obtained by this method is more obvious than that based on the original anchor in the YOLOv3 network.So it is more suitable for vehicle component detection in this dataset.2.A vehicle detection method based on vehicle components is proposed.Firstly,according to the relative position relationship between vehicle components and vehicle center points,a two-dimensional joint Gaussian distribution model based on relative distance is constructed,and the histogram statistical method is used to estimate the model parameters;Then,using Hough voting idea for reference,a weighted voting mechanism of vehicle center points based on multi-component information is proposed,and on this basis,the algorithm of determining the best center area of NMS based on variance is further proposed;Finally,a vehicle component classification algorithm based on IOU threshold is given,and vehicle detection confidence and vehicle orientation information are obtained by probability independence hypothesis and maximum filtering respectively.In order to obtain a more accurate vehicle bounding box,a vehicle bounding box fusion method using Bayesian filtering to fuse multi-component information is also proposed.The experimental results show that the proposed method has both interpretability and high recognition ability of deep learning methods.What is more,it is ideal for the occlusion problem of nearby vehicles,and has the effect of reducing false alarms and compensating for missed detection.The AP value can reach 82.6%,which is better than the YOLOv3 algorithm.3.A component-based vehicle motion parameter estimation method is given.Starting from the safety of unmanned vehicles,using the vehicle detection and vehicle component information in this paper,the collision time estimation method based on the optical flow method,the vehicle heading angle and vehicle azimuth angle of the threatened vehicle in front of the unmanned vehicle in the camera coordinate system are given respectively.Then,the collision threat situation is constructed based on three vehicle motion parameters: the collision time,the vehicle azimuth angle and the vehicle heading angle.The experimental results show that the collision threat situation constructed by the method in this paper is helpful for the unmanned vehicle to understand the surrounding scenes and provides a decision basis for vehicle collision risk assessment.