Offline tracking with object permanence

Offline tracking with object permanence

Liu, Stan (TU Delft Mechanical, Maritime and Materials Engineering)

Caesar, H.C. (mentor)

Delft University of Technology

Mechanical Engineering | Vehicle Engineering | Cognitive Robotics

2023-08-25

Learning-based approaches are widely applied in the perception system of autonomous vehicles. Thus, a large amount of labeled data are needed to train these data-hungry models. To reduce the expensive labor cost for manual labeling autonomous driving datasets, an alternative is to automatically annotate the datasets using a trained offline perception system. Previous works mainly focused on generating accurate object trajectories in 3D space from point cloud sequence data. However, point cloud sequences might be partially missing due to occlusion. Such occlusion scenarios in the datasets are common yet underexplored in offline autolabeling. In this paper, we propose an offline tracking model that focuses on the occlusion cases of vehicle tracks. It leverages the concept of object permanence which means objects continue to exist even if they are not observed anymore. The model contains two main parts: a re-identification (Re-ID) module that takes online tracking results as input and associates the identities of the tracklets, and an offline motion estimator that recovers the fragmented tracks under occlusion. Both modules innovatively use the vectorized high-definition map (HD map) as one of the inputs to refine the tracking results with occlusion. The model can significantly reduce the number of identity switches and false positives compared to the original online tracking results.

offline tracking
Autonomous driving
Occlusion

http://resolver.tudelft.nl/uuid:728d6807-d478-42d5-b70d-8eccfd232da0

Student theses

master thesis

© 2023 Stan Liu