HD Map Information for Three-Dimensional Multi-Object Tracking

Author(s)
Kothari, Yash
Advisor(s)
Editor(s)
Associated Organization(s)
Supplementary to:
Abstract
The breakdown of 3DMOT into its sub-parts allows us to identify gaps in the research that are applicable to our problem area. While the ability to detect objects is being worked on passively through multiple studies (including ours), questions still remain about how the algorithm of 3DMOT can be improved beyond just working on improving the quality of 3-D Object Detection algorithms. Different ways to associate bounding boxes of the same object and contrasting methods of estimating the motion of an object to understand where to look for it in the future are both significant problems that showcase the lack of a robust 3DMOT algorithm that can handle both smooth and abrupt motion. Gao et al. (2020) provide a unique method of encoding feature information from HD Maps (lane geometry information) that might be a useful additional feature in providing 3DMOT algorithms with supervision on regions of interest that may contain objects to be tracked. Overall, the process of improving 3DMOT algorithms relies on improving association supervision, solving both smooth and robust motion cases and providing a good estimate on how an object continues to move over time. Our method incorporates the above ideologies and aims to utilize the information from HD Maps to improve the performance of 3DMOT algorithms, with an emphasis on decreasing identity switches and increasing multi-object tracking accuracy
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis
Rights Statement
Rights URI