Multi-objects tracking based on 3D lidar and bi-directional recurrent neural networks under autonomous driving
Citation & Export
Hide
Simple citation
Xin, Pujie.
Multi-objects tracking based on 3D lidar and bi-directional recurrent neural networks under autonomous driving. Retrieved from
https://doi.org/doi:10.7282/t3-dgnr-4x58
Export
Description
TitleMulti-objects tracking based on 3D lidar and bi-directional recurrent neural networks under autonomous driving
Date Created2020
Other Date2020-01 (degree)
Extent1 online resource (vii, 51 pages) : illustrations
DescriptionMulti-Objects Tracking (MOT) is an important topic in navigation, where robots or vehicles should interact safely with the moving objects in the environment. The navigation system can hardly make a path plan if there is no position and velocity information of the moving objects. Generally, moving objects tracking includes three stages which are sensor measurement preprocessing, data association, and kinetic states estimation. This thesis presents a new approach to improve the matching precision in the data association stage by combining more characteristics of the targets and their kinetic states detected by sensors. In more details, different perception systems infer different characteristics of the moving objects, which will help distinguish the moving objects, and thus improve the matching precision. However, it is hard to use these specified characteristics of the targets in widely used Single Object Tracking (SOT) strategies such as the Global Nearest Neighbor (GNN) approach, the Joint Probabilistic and Data-association (JPDA) approach, and the Multi-hypothesis Tracking (MHT) approach. Generally, moving targets are viewed as point-like targets in SOT strategies, which means that many characteristics are ignored in the matching process and these SOT methods just associate the data by estimating the probability of each association and select the association with the highest probability. Only the object-to-hypothesis distance was used to compute the probability and the Hungarian algorithm provides an optimal solution to the distance matrix, which is considered as the optimal assignment in the data association. In this thesis, a new method is proposed to calculate the cost matrix considering both the distance matrix and the pose of the moving targets. To compute the new assignment matrix with both distance and pose information, bidirectional Recurrent Neural Network (Bi-RNN) is proposed to input the object-to-hypothesis distance matrix and output the optimal assignment matrix. The loss function of the Bi-RNN is simplified as the mean square error. Multiple Object Tracking Accuracy and Precision, i.e., MOTA and MOTP, are standard and widely used matrix to assess the quality of MOT. In this thesis, they are hence used to evaluate the performance of the proposed tracking method. Experimental datasets, i.e. the KITTI datasets, are used to demonstrate the effectiveness of the new algorithm.
NoteM.S.
NoteIncludes bibliographical references
Genretheses, ETD graduate
LanguageEnglish
CollectionSchool of Graduate Studies Electronic Theses and Dissertations
Organization NameRutgers, The State University of New Jersey
RightsThe author owns the copyright to this work.