Illumination-adaptive Road Surface Detection System Considering Small Target Recognition

Vision-based road surface detection plays a crucial role in ensuring the driving safety of autonomous vehicles. However, most existing methods primarily focus on daytime driving scenarios, with little attention given to nighttime road surface detection. Additionally, these methods often assume that the road surface in the image is clear and complete, which does not align with real-world applications. This paper proposes a detection system that automatically adapts to different illumination conditions and considers scenarios where the road surface area is small and challenging to recognize. First, we use a novel horizon detection method to separate the environmental area from the road surface area in the image. The environmental information is then used to determine whether the driving scene is during the day or night, allowing the system to automatically load the appropriate model parameters based on the illumination conditions to capture road surface characteristics under varying lighting conditions. We focus on the actual driving road surface area by extracting patches and feeding them into a multi-patch fusion decision network (MFDNet) proposed in this study, where the prediction results are fused at the decision level based on the confidence of each patch. This method ensures that small road surfaces in the images are accurately detected while providing sufficient information for recognizing larger road surface areas. Results on public datasets demonstrate that this system has better overall performance and practical application value compared to other methods.