Testing Autonomous Driving System based on Scenic
Zheng Li, Zhanqi Cui, Huanhuan Wu, Yating Zheng · 2021 IEEE 21st International Conference on Software Quality, Reliability and Security Companion (QRS-C) · 2021
Autonomous driving system develops rapidly in recent years, which also brings many critical issues. Object detector is one of the most important modules in autonomous driving system, how to ensure its quality is an urgent problem to be solved. Many studies have been carried out to perturb the input images by mutation to test object detectors. But digital adversarial perturbations, e.g., changing image pixels, may never happen in the physical world. Therefore, a metamorphic testing approach for object detectors based on Scenic is proposed in this paper. An abstract scene in Github repository of Scenic and object detector Yolact were used in the experiments. The experimental result shows that the proposed approach can obtain a minimum of 43% and a maximum of 96% defect detection rates under different confidence probabilities.