Oblique Convolution: A Novel Convolution Idea for Redefining Lane Detection

Lane detection plays an important role in the field of automatic driving. Conventional convolutional operations typically focus on local block-like region, while lane often span across multiple image regions as strip lines. How to make up for the limitations of ordinary convolution in extracting lane information and make the network focus on the strip lanes is a challenging problem. For this purpose, we propose an oblique convolution idea to effectively extract the information of the whole lane. Specifically, an Oblique Rotation Module (ORM) is proposed to rotate the lane feature map into a vertical orientation, enabling better lane positioning and recognition by using row anchors. Additionally, we introduce a Strip Spatial Attention Module (SSAM) to extract global features of the entire lane and employ deformable convolutions to adapt to curved lane lines. Finally, we restore the feature map using Anti-oblique Rotation Module (AORM) and obtain lane detection results through classification-based predictions. During training, we incorporate a multi-scale auxiliary classification loss to predict the presence of lane lines at different network levels, aiding the network in learning the structural information of lane lines. A large number of experiments on the lane detection benchmark dataset show that our method can achieve advanced performance, and we have achieved good results in public datasets. The accuracy on the Tusimple dataset is 96.50%, the F1 on the CULane dataset is 79.33%. Both results rank in the top 10 on the leaderboard, and the inference speed reaches 35 FPS. Code and data are available on https://github.com/gongyan1/Oblique-Convolution