Autonomous Driving Software Testing Based on Interpretation Analysis
Songcheng Xie, Shengpeng Zhu, Qifan He, Zhanqi Cui · 2024
With the development of artificial intelligence, autonomous driving has become a typical use of AI technology, while its safety has received widespread attention. However, current mainstream data-driven automated driving test methods lack interpretability, automation features, and a high degree of realism in generating test images. To solve the above problems, this paper proposes a method ATIA (Autonomous Test based on Interpretation Analysis), which generates gradient heatmaps based on Grad-CAM and uses deep image synthesis network to generate test images by replacing key objects according to the heatmap, and also designs a corresponding automated testing tool. The experimental results show that ATIA has the ability to detect defects, while generating images that are closer to real images.