Uncertainty Quantification for Adaptive Measurement Noise in Kalman Filter Based Object Tracking
Ulreich F, Kaup A, Ebert M (2024)
Publication Type: Conference contribution
Publication year: 2024
Publisher: Institute of Electrical and Electronics Engineers Inc.
Conference Proceedings Title: 2024 Sensor Data Fusion: Trends, Solutions, Applications, SDF 2024
Event location: Bonn
ISBN: 9798331527440
DOI: 10.1109/SDF63218.2024.10773809
Abstract
The application of deep learning systems in the perception of autonomous vehicles results in the need for information about the trust in its predictions. In our work, we use methods from the field of uncertainty quantification and probability theory to natively integrate the 3D object detection stage into multi-object tracking for perception. For this purpose, we fine-tune a state-of-the-art sensor fusion model for camera and LiDAR data to output a full probability distribution. We propose a calibration scheme and we quantitatively show that predictive uncertainty can be linked to object occlusion. We furthermore compare predictive uncertainty with post-hoc calculated empirical uncertainty in terms of calibration and tracking performance. Through our approach, we reduce the calibration error of the center x-coordinates about 92% and of the center y-coordinates about 87%. Moreover, we utilize the predictive uncertainty estimate of our detector as the measurement variance in Kalman filtering for multi-object tracking, reducing average tracking position error on the nuScenes data set by about 1.3%.
Authors with CRIS profile
Involved external institutions
How to cite
APA:
Ulreich, F., Kaup, A., & Ebert, M. (2024). Uncertainty Quantification for Adaptive Measurement Noise in Kalman Filter Based Object Tracking. In 2024 Sensor Data Fusion: Trends, Solutions, Applications, SDF 2024. Bonn, DE: Institute of Electrical and Electronics Engineers Inc..
MLA:
Ulreich, Fabian, André Kaup, and Martin Ebert. "Uncertainty Quantification for Adaptive Measurement Noise in Kalman Filter Based Object Tracking." Proceedings of the 2024 Sensor Data Fusion: Trends, Solutions, Applications, SDF 2024, Bonn Institute of Electrical and Electronics Engineers Inc., 2024.
BibTeX: Download