Object Recognition Based On Sparse 3d Point Cloud In Urban Environment

Autonomous driving technologies have received more and more attention in the last decade.Object perception system is one of the important technologies for an autonomous vehicle,it can detecte environment and provide information for control system.A Light Detection and Ranging(Li DAR)is widely used in autonomous driving research,because it can provide more accurate 3D information of the environment and work properly without external environmental conditions.A 64-beam Li DAR is widely adopted for localization and object perception in most autonomous driving researches.However,its high price limits broader ap plications of autonomous driving technologies,a new perception method based on an affordable 16-beam Li DAR for autonomous driving applications is necessary.So,this dissertation is focused on the implementation of object classification based on 16-beam Li DAR for autonomous driving applications.A three-stage perception system based on Li DAR is proposed,which consists of pre-processing of raw data,segmentation and classification.Pre-processing of raw data consists of coordinate transformation and ground estimation.Raw data of Li DAR are stored in order point clouds after coordinate transformation and being divided into different frames,ground estimation is implemented on the order point cloud to find the points belonging to ground.The point cloud is di vided into separate points sets using suitable segmentation algorithm in segmentation process.These points sets are converted into a feature vector and classified by classifiers.Free space constraint and shape information of objects are introduced to red uce segmentation errors in grid image segmentation process.After building a grid image,an image dilation is adopted to connect pixels belonging to the same object.At the same time,the missing readings in the bottom part of the point cloud are interpolated to obtain a reliable free space constraint.According to the constraint,those reasonless occupied pixels in dilation image are corrected.Subsequently,corrected gird image and a region growing algorithm are used to label the point cloud.And some segments are merged based on their shape information.The algorithm is evaluate using data gathered in real urban environment,and the result show that the segmentation method can reduce segmentation errors.Features wildly used in classifying point clouds are tested on the data sets gathered from 16-beam Li DAR,the most suitable feature combination is selected for final classifier.In order to reduce time spent on labeling samples,the track sequences of objects are manually labeled before training classifier s.Single feature classifiers are trained to estimate whether or not this feature is suitable for the data of 16-beam Li DAR,and those achieving the best performance are chosen for finial classifiers.Objects in urban environment are divided into four categories: car,truck,pedestrian and background.Four individual one-versus-all classifiers are trained respectively.Final class decisions are made based on the maximal probability outputs from four classifiers.