Object Tracking Based On Sparse 3D Point Cloud In Urban Environment

Over 1 million people are killed and over 20 million people get injured in road-related accidents each year.Most of these accidents are caused by wrong operation while human beings are driving on the road.Therefore,it is urgent to achieve absolutely safe driving for automobiles and to reduce traffic accidents caused by improper operation.Accurate tracking of the surrounding objects is a prerequisite and contributed to making real-time control decisions.In order to ensure safety,tracking will be one of most important technologies in self-drving research.Based on current methods and literature,this dissertation provides a whole system for real-time tracking of moving targets using Velodyne 16 lidar in urban road scenes.This dissertation also put forward a feasible scheme for the verification of the tracking algorithm to verify the accuracy and real-time performance.The ground extraction and target segmentation in single-frame,targets correlation and tracking in different frame point cloud will be mainly disscussed in this dissertation.We will use the local features of each point in grids combined with the CRF(conditional random field)algorithm to get the optimal Z value of each grid which is used to judge the ground points.Then we will project the 3D point cloud without ground points to a 2D plane to do targets segmentation which is samilar to segmentation in images.A new correlation algorithm according to the target features in different frame will be proposed.The correlation matrix will be used to find a global optimal hypothesis.When the targets are matched correctly,a Kalman filter and a smooth filter will be used to achieve precise tracking.The verification scheme is mainly divided into the verification of the association algorithm and the tracking algorithm.By storing a large number of associated point cloud sequences offline to judge the correctness manually,which is contributed to get an accurate rate in target correlation.Since the self velocity is a known variable,we choose static cars or pedestrians as our targets to vertify accuracy rate of the tracking algorithm.