Uncertainty Estimation for 3D Object Detection and Tracking

Publikation: Bog/antologi/afhandling/rapportPh.d.-afhandling


3D object detection and tracking are important perception tasks for robotics and autonomous driving, which allow the system to analyze its surroundings, plan movement and interact with objects, animals or humans. These systems are mostly required to be autonomous to have freedom of movement and be able to work without the direct human control, which imposes a limitation on the computational capabilities of devices that can be used to process the inputs. Such limitations lead to a need for both fast and accurate methods, as the inaccuracy in predictions can lead to damages to property, health or life. However, the accurate neural networks are prone to overestimate their confidence, which leads to "silent errors" and may make an illusion of a controlled environment, while the system actually fails to adequately react to the inputs. This can be mitigated by the use of explicit uncertainty estimation, which is capable of not only signaling that something went wrong during perception, but also improve the quality of the perception methods. In this thesis, Uncertainty Estimation for 3D Object Detection and Tracking, I first provide an overview of 3D object detection and tracking, uncertainty estimation methods and their applications, then I describe the technical contributions to improve the real-world applicability of the state-of-the-art methods by improving the voxel-based 3D object detection inference for embedded devices, proposing a novel method for real-time embedded 3D single object tracking, improving the uncertainty estimation methods and applying the proposed uncertainty estimation methods to 3D single and multiple object tracking, improving the performance of those methods. These contributions have made progress in the fields of 3D object detection, tracking and uncertainty estimation and provided the directions for future research.
ForlagAarhus University
StatusUdgivet - maj 2023


Dyk ned i forskningsemnerne om 'Uncertainty Estimation for 3D Object Detection and Tracking'. Sammen danner de et unikt fingeraftryk.