Exploiting Image Enhancement and Edge Detection for Low‐Light Road Segmentation

ABSTRACT Low‐light road segmentation is a challenging dense prediction task, which is very important for the safety monitoring of intelligent sea ports at night and the autonomous vehicles of shipping logistics. Most current research focuses on scenes with sufficient light. At the same time, there are few datasets for low‐light scenes, making research on low‐light perception very difficult and seriously restricting the shipping logistics industry's ability to operate safely at night. Directly applying segmentation methods developed for well‐lit scenes to low‐light road segmentation is unsatisfactory. To solve this problem, we propose a new approach by building an image enhancement module and an edge detection module, and integrating them into existing well‐lit segmentation models as a plugin to meet the road segmentation requirements for low‐light scenes. Specifically, to compensate for the lack of low‐light image detail, we design an image enhancement module that achieves end‐to‐end pixel‐level image enhancement by connecting four image processing filters in series and using convolutional neural network to predict hyperparameters. Additionally, to address the problem that road edges become blurred and difficult to extract in low‐light images, we design an edge detection module to maximize its ability to extract road edges by selecting differential pixel pairs using different strategies and efficient combinations. We conduct comprehensive experiments on our newly released dataset, LoRD, demonstrating that our method significantly outperforms previous state‐of‐the‐art models with relatively few parameters and computational cost. Our method achieves new SOTA performance in terms of accuracy and computational efficiency, achieving 93.29 at 78.44 FPS in DDRNet‐23‐slim. The source code and dataset will be publicly available.