Accurate estimation on the perception performance is an essential prerequisite for autonomous driving safety. State-of-the-art perception models get excellent performance on existing well-labelled dataset, while suffering the lack of reliable confidence in varying scenarios. This paper is dedicated to investigate an online perception performance estimation, in particular, to calculate the precision rate of each detected object adaptively without manually labelled annotations. A parameterized relationship is constructed between the confidence output by perception models and the precision performance. And then the relationship can be online updated according to a robust multiple hypothesis tracking (MHT) system iteratively. The proposed method is applied on widely studied Mask-RCNN on the KITTI dataset, to get precision estimations accordant to the ground-truth in different scenarios, respectively.
Online Perception Performance Estimation for Autonomous Driving Vehicles
Lect. Notes Electrical Eng.
Society of Automotive Engineers (SAE)-China Congress ; 2022 ; Shanghai, China November 22, 2022 - November 24, 2022
2023-04-29
10 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
ONLINE PERCEPTION PERFORMANCE EVALUATION FOR AUTONOMOUS AND SEMI-AUTONOMOUS VEHICLES
Europäisches Patentamt | 2021
|PERCEPTION ASSISTANT FOR AUTONOMOUS DRIVING VEHICLES (ADVS)
Europäisches Patentamt | 2019
|Perception assistant for autonomous driving vehicles (ADVs)
Europäisches Patentamt | 2021
|PLANNING DRIVEN PERCEPTION SYSTEM FOR AUTONOMOUS DRIVING VEHICLES
Europäisches Patentamt | 2021
|Obstacle perception calibration system for autonomous driving vehicles
Europäisches Patentamt | 2022
|