Research on Methods of Improving Robustness of Deep Learning Algorithms in Autonomous Driving
Juan Xiao, Haojian Long, Renfa Li, Faying Li · 2022 IEEE International Conference on Advances in Electrical Engineering and Computer Applications (AEECA) · 2022
One of the main reasons why automatic driving cannot be industrialized today is that the target detection model used cannot have good detection results under various severe weather conditions, that is, the robustness is poor. If a model with good robustness to all possible image damage can be constructed, then interference similar to weather changes will not be a problem. This paper introduces a method of simulating image damage, and proposes a robust detection method to detect the robustness of the algorithm, and proves that two widely used target detection models have serious performance losses for damaged images. Provides a simple data enhancement technology to stylize training images. After training the model using this method, the robustness of the model can be significantly improved.