Object tracking aims at estimating the state of moving objects based on remote measurements. To evaluate online algorithms in automotive systems, ground truth data must be acquired, which is a time-consuming and expensive approach. We propose a novel offline approach to generate ground truth data from existing sensor measurements using CAD models. In our approach, we provide error bounds for the localization of the objects based on the measurement noise of a single laser beam and the sensitivity of the point cloud registration. To estimate accurate kinematic states of the vehicle, we apply an extended Rauch-Tung-Striebel smoother on the stored measurements. In experiments with real sensor data, we demonstrate that the performance of the proposed approach is superior to DGPS within the near range.
Ground truth data plays an important role in validating perception algorithms and in developing data-driven models. Yet, generating ground truth data is a challenging process, often requiring tedious manual work. Thus, we present a post-processing approach to automatically generate ground truth data from environment sensors. In contrast to existing approaches, we incorporate raw sensor data from multiple vehicles. As a result, our cooperative fusion approach overcomes drawbacks of occlusions and decreasing sensor resolution with distance. To improve the alignment precision for raw sensor data fusion, we include mutual detections and match the jointly-observed static environment to support differential global positioning system localization. We further provide a new registration algorithm, where all point clouds are moved simultaneously, while restricting the transformation parameters to increase the robustness against misalignments. The benefits of our raw sensor data fusion approach are demonstrated with real lidar data from two test vehicles in different scenarios.
For the evaluation of autonomous driving systems, this paper provides a new approach of generating reference data for multiple extended object tracking. In our approach, we apply a forward-backward smoother for objects with star-convex shapes based on the Labeled Multi-Bernoulli (LMB) Random Finite Set (RFS) and recursive Gaussian processes. We further propose to combine a robust birth policy with a backward filter to solve the conflict between robustness and completeness of tracking. Thereby, cluster candidates are evaluated based on a quality measure to only initialize objects from more reliable clusters in the forward pass. Missing states will then be recovered by the backward filter through post-processing the unassociated data after the smoothing process. Simulations and real-world experiments demonstrate superior performance of the proposed method in both cardinality and individual state estimation compared to naive LMB filter and smoother for extended objects.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.