A Method for Evaluating Key Elements Affecting Visual Perception Capabilities of Unmanned Autonomous Systems in Virtual Scenarios
Junchao Du, Haitian Liu, Binglin Liu, Qianchao Hu, Feng Wang · 2022
In order to reduce field testing expenses and construct rare extreme scenarios in the real world, it is convenient to evaluate the visual perception capability of unmanned autonomous systems more effectively. In this paper, a virtual base scenario based on a real test site is constructed using tilt photography, and then three target detection models, SSD300, Faster R-CNN and RetinaNet, which are fully trained using the COCO dataset, are used instead of the unmanned autonomous system visual perception module. After that, the dataset was built for testing with different values of influencing elements. The experimental results show that changing the quantified values of the influencing elements can effectively affect the detection effect of the target detection models, i.e., the unmanned autonomous system perception capability can be effectively tested even in virtual scenarios, which can save cost and shorten the cycle time for the development of unmanned autonomous systems.