An Improved YOLOv8s Algorithm for Lane Detection in Road Scenes

Accurate lane detection from road scene images is crucial for environment perception in autonomous driving and intelligent transportation systems. However, in complex traffic scenarios, lane markings are slender, variable in shape, and easily affected by illumination changes, occlusion, and adverse weather, which restricts the feature representation and structural modeling capability of the conventional YOLOv8s lane detection model, resulting in suboptimal detection accuracy and robustness. To overcome these limitations, this study proposes an improved YOLOv8s-based lane detection algorithm, YOLOv8s-CLW. A composite attention mechanism is integrated into the feature extraction stage to enhance the network's focus on critical semantic features and spatial information of slender lane markings. Lightweight deformable convolution is introduced to adaptively model curved and non-rigid lane structures, and an improved bounding box regression loss is employed to enhance localization stability and detection precision. Experimental results demonstrate that, on the original test set, YOLOv8s-CLW achieves a Recall of 50.0% and a mAP of 59.8%, representing improvements of 0.6% and 8.9% over the baseline YOLOv8s. On an external validation set, Precision, Recall, and mAP reach 93.0%, 92.0%, and 96.1%, demonstrating significant performance gains. Qualitative visualization results under complex weather conditions (fog, rain, dust, and snow) indicate that the model can still detect partial lane lines, with the number of detected lane lines ranging from approximately 30% to 85% of that under normal conditions—80%~85% under mild interference and 30%~50% under extreme weather, while maintaining continuity and completeness of predictions, reflecting strong environmental adaptability and robustness.