Study On Pedestrian Detection And Tracking Based On Camera And LIDAR Fusion

Pedestrian detection is a key technology in autonomous driving perception system.Although the current vision-based pedestrian detection has obtained a very good detection effect,the camera is sensitive to light and shadow,meanwhile it is unable to provide precise location information,which is difficult to meet autonomous driving task.So this dissertation combines image and LIDAR to complete the task of pedestrian detection and tracking,tries to reduce dependence of vision-based pedestrian detection and achieves stable tracking results at the same time.The main work in this dissertation includes:1.The experiment analyzes three aspects of training a pedestrian classifier,that are features,classifier and samples,which provides guidance for the training of the classifier.According to the characteristic of an on-board camera,a training and testing framework based on constraints of camera projection has been put forward.At the same time,in view of the shortcomings of the single template classifier,this paper classified pedestrian from the aspect of feature and human experience,and train several pedestrian submodels,which improve the detection effect significantly.I have made performance tests and comparison of our algorithm in our pedestrian detection database,and the experimental result shows that this algorithm's average Miss-Rate achieves 6.06%,and has been applied in the environmental perception system of our autonomous car.2.This paper proposes a new description method which added part model to describe the pedestrian detection problem.After formally describing the location relationship of the master model and part models,the paper decomposes the pedestrian score into detection score and position constraint score,and transforms the final detection problem into a zero-one programming problem,and finally using Hungary algorithm to solve.3.We propose a framework of pedestrian detection and tracking which fuse camera and LIDAR.In this framework,LIDAR-based pedestrian segmentation is regarded as weak classifier,and vision-based pedestrian classifier as strong classifier.And with a tracking module,the final detection is given by fusing multiple sensor information in multiple frames together with a voting strategy.Our experiment shows that the fusion framework put forward in this chapter can give more stable detection and tracking results,and performs better on some miss detection situation.