우천 환경에서 다차선 검출을 위한 사전 지식기반 접근법

This study proposes a rule-based algorithm that leverages prior knowledge to detect multiple lane-lines in images captured under adverse weather conditions. In particular, rainy weather poses significant challenges for lane detection, as road images captured under such conditions often contain random raindrops, wet surfaces, and puddles. These elements cause light reflections and either blur or obscure the lane boundaries. Despite these disturbances, the actual lanes on the road remain unchanged. Based on this observation, the algorithm searches for regions that locally preserve lane-line characteristics by placing multiple small windows along the previously detected lane lines. Within each window, it detects valid segments using engineered features, the Hough transform, and sequential filtering. The algorithm demonstrates that these features outperform those learned by recent deep neural networks in capturing the properties of lane lines under rain-induced noise, achieving about 10% higher F1 scores than recent approaches.