A Rain Image Generation Approach for Autonomous Driving Perception Sensors

Proper rain image generation is important to assess the safety of autonomous perception system in rainy environment and improve perception performance. Rain image generation methods for autonomous driving vision sensors have become a research hotspot. The diversity of data can be extended and the performance of vision-based perception can be improved by effective real rain image generation. Although recent deep learning-based image generation of real rain has good performance, it is still an open problem for autonomous driving scenarios. Firstly, the existing rain images dataset only considers the streak characteristics produced by rain in modeling real rain features, and does not consider the interference of raindrops and fog on the image. Secondly, the distribution of the generated rain streaks, raindrops and fog lacks the influence of physical characteristics brought about by vehicle speed, rainfall and gravity during vehicle operation. Moreover, the lack of a real-time and efficient method for generating realistic raindrop images makes data enhancement and safety testing of sensing systems challenging. To solve the problem of real rain images generation for autonomous driving scenarios, we propose an efficient data generation method that combines physical properties and robot operating system to realize real-time generation of real rain for dynamic scenes and real-time data publish.