Vehicle Detaction And Tracking Algorithm In Urban Congestion Environment

In recent years,with the rapid development of autonomous driving technology,intelligent vehicles will be gradually popular in people's lives.Urban environment with numerous vehicles is one of the most important application scenarios in the unmanned driving field.In this complex environment,small objects are difficult to detect and vehicles are densely occluded,which seriously affects the perception performance.In this paper,fusing the information collected from camera and lidar,a relatively efficient and accurate detection and tracking algorithm is designed to perceive the 3D position,motion state and trajectory of other vehicles around.Thus,the planning accuracy and the driving safety of intelligent vehicles are ensured.There are two sources of experimental data in this paper: the first one is the KITTI benchmark and the second one is the urban road scene in Shanghai collected by the experimental intelligent vehicle.This vehicle perception algorithm will be introduced in the following three modules.Firstly,there is an image-based 2D object detection module.To solve the scale variance problem,an efficient hierarchical convolutional neural network named FMLA-CNN is proposed.The feature fusion strategy and multi-level alignment strategy(MLA)are separately used in the first stage and the second stage of the network to promote small object detection.The MLA strategy introduced in this paper can effectively reserve hierarchical spatial location information,thus promoting multi-scale object detection.This 2D detection algorithm achieves excellent results on the KITTI rankings and the effectiveness of small object detection are verified through actual scene comparison.The second is 3D object detection module.Aiming at the vehicle occlusion problem,a detection scheme which fuses image features and point cloud bird-view features is introduced.In this paper,we present a 3D object detection network called PTF-AVOD,which contains a perspective transform fusion module.The proposed perspective transform fusion method is a fine pixel-level correlation fusion,which makes it possible to combine two modal information more effectively.A loss function for dense 3D object detection is designed for the network by referring to the Repulsion Loss in 2D detection,so as that the accurate detection of direction angle is realized.The experiments show that the 3D detection performance of the PTF-AVOD network on occluded vehicles is significantly improved compared with the AVOD(Aggregate View Object Detection)network.Thirdly,as for the multi-target tracking module,a correlation filter tracker(Point-KCF)is proposed,which takes raw 3D point cloud as input and tracks vehicles in all directions on the bird view.A multi-target tracking algorithm system based on Point-KCF is designed to complete multi-target tracking of vehicles on the bird-view of point cloud.The system uses the PTF-AVOD network mentioned above to detect 3D targets every four frames while online tracking based on Point-KCF tracker are applied to three consecutive frames.The fusion results of 3D detection are used to initialize and update the results of tracker.Finally,the correlation matrix is redesigned and the Kuhn-Munkras algorithm is used to correlate data between different frames.The experimental results show that this perception algorithm accurately predicts the location,the trajectory and the angle of the vehicles on the bird view in real time.