Domain Adaptive Road Perception Network of Autonomous Vehicles

Road conditions can significantly influence the safety of autonomous vehicles (AV). Existing sensors of AV are usually less effective in recognizing road conditions during night and inclement weather. While intelligent tire systems, can identify different road conditions accurately and hardly be affected by weather or illumination conditions. However, data collected under different working conditions differ from each other with significant shifts and are hard to collect due to the tremendous field experiments. To this end, we propose a domain-adaptive model that can extract invariant features to different driving speeds. The proposed method contains two modules: a) a data pre-processing module to extract variable period signals and b) an adversarial transfer learning module for learning invariant features cross working conditions. The field tests have demonstrated that the proposed method performs better on road condition perception than other transfer learning methods.