Research On Object Detection Based On LiDAR Point Cloud In Autonomous Driving

In recent years,with the development of artificial intelligence and the raise of requirements on safe driving,autonomous driving has become the research hotspot in academic and industrial circles.Autonomous cars consist of three parts:perception system,decision system and control system.Perception system is responsible for obtaining the position and pose of car and information of the surrounding environment.Perception system is the base of decision system and control system,and it plays an important role in autonomous driving.This thesis mainly researches object detection technology in perception system of autonomous car.Based on point cloud data obtained by LiDAR,it implements object detection on surrounding environment of autonomous car utilizing features of point cloud data.It researches object detection technology based on clustering and deep learning,and discusses how to improve the speed and accuracy of algorithms.Following is the work of this thesis.Firstly,this thesis states the definition of object detection,and introduces machine learning algorithms,including traditional machine learning and deep learning,about object detection.It introduces one-stage detectors and two-stage detectors in object detection,and compares advantages and disadvantages of these two detectors.Secondly,this thesis states clustering-based object detection algorithms,which consists of ground segmentation and point cloud clustering.When discussing ground segmentation,thesis compares two different segmentation algorithms,and analyzes how to segment ground more accurately.In point cloud clustering,it discusses how to make algorithms adjusting to different environments and satisfying real-time requirements in autonomous driving applications.Then,this thesis studies deep-learning-based object detection algorithms.It discusses how to order unordered point cloud data,and constructs a 3D fully convolutional network detector to conduct object detection using deep learning.It conducts several improvements to boost the performance of the detector.Meanwhile,aiming at defects of 3D convolution,this thesis constructs a 2D fully convolutional network detector to accelerate detecting speed.Finally,this thesis conducts experiments to verify these detection algorithms.In addition,this thesis compares advantages and disadvantages of these two algorithms and states their available conditions.This thesis mainly focuses on perception system in autonomous cars and studies object detection based on LiDAR point cloud.This thesis targets on making detection faster,more accurate and more feasible.At the same time,this thesis develops object detection programs based on Robot Operating System in Linux,which applies theory to practical work.