Research Of Object Tracking Based On 3D Laser Point Cloud

Object tracking technology plays an important role in practical applications,such as robot,automatic car and video surveillance,thus it has been widely studied at home and abroad.At present,most of the object tracking algorithms focus on 2D image data,these methods try to track objects by obtaining the 2D bounding box of the target object on the image.However,because 2D images cannot obtain the distance information of the target object and are easily affected by the light or other factors,they are often limited in practical application.Meanwhile,the object tracking methods based on 3D LIDAR data usually based on "tracking-by-detection" framework,which heavily rely on the results of front-end detector.When the detector fails,the tracker will not be able to track effectively because of no input,so this framework cannot track the target object robustly and continuously.Therefore,based on the above point,this thesis studies the online realtime tracking algorithm of 3D target object based on 3D LIDAR point cloud.The main contents of this thesis are as follows:(1)This thesis introduces the research algorithm and related work of object tracking technology,including correlation filtering and deep learning methods.(2)To solve the problem of 3D person tracking,this thesis proposes an end-to-end network based on Siamese network named Point Siamese Network(PSN).Different from the "tracking-by-detection" framework,this method is based on the idea of similarity.It computes the similarity between the first frame point cloud of target person and the current frame point cloud,and predict the 3D coordinate of target person by computing the cosine similarity score of each point in the current point cloud.Meanwhile,by adding an attention module to guide feature extraction,it could filter noise and improve the tracking performance.Because it does not depend on the front-end detector,it could track the target person independently and robustly.By comparing with UKF tracking algorithm which has been widely used,this thesis verifies that PSN has higher accuracy and robustness than UKF,no matter in the dataset we collect or in the KITTI dataset.(3)To solve the problem of accurate 3D object tracking,this thesis proposes a novel network named Point Siamese Region Proposal Network(Point-SiamRPN)for 3D LIDAR point cloud object tracking based on regression.The algorithm first uses deep network to extract high dimensional feature from the two frame input point clouds.Meanwhile,this thesis proposes two types of cross correlation modules which are used to fuse features and compute similarity.The two modules can compute the similarity between two point cloud features efficiently.The similarity feature is also seen as the weight of the current frame point cloud feature.After the weighted features are obtained,the 3D bounding box of the target is tracked by regression of the Region Proposal Network(RPN),thus the object could be tracked.This thesis verifies and tests the method on KITTI and H3D datasets and compares it with the other algorithms.The experiments show that the network proposed has a competitive performance compared with the stateof-the-art method.(4)Finally,the work of this thesis is summarized,and the future research is prospected.