Metamorphic Testing for Traffic Light Recognition in Autonomous Driving Systems

Abstract

Traffic light recognition is significant for environ-ment perception and behavior planning in autonomous driving systems. Erroneous traffic light recognition results can lead to serious traffic accidents. Thus, it is necessary to ensure the correctness of traffic light recognition in autonomous driving systems. However, the existing approaches to testing traffic light recognition need manual collection and labeling test data which become prohibitively expensive as the number of test situations increases. In this paper, we propose an approach that uses metamorphic testing to test traffic light recognition. The approach does not need labels to verify traffic light recognition results and can automatically generate traffic lights of different colors of the same scene. In the final experimental stage, we implement our approach to test traffic light recognition models and end-to-end autonomous driving systems. The result of the experiment demonstrates that our approach verifies traffic light recognition results independent of labels and can automatically generate natural traffic lights. Moreover, our approach also discovers traffic light recognition errors.