Simultaneous 2D and 3D Lane Detection Algorithm Based on Multi-Granularity Lane Attention

Lane detection is a crucial step in the perception task of autonomous vehicle driving. However, to the best of our knowledge, there is little literature to simultaneously obtain the two-dimensional coordinates of lane in the image coordinate system and the three-dimensional coordinates in the vehicle coordinate system from the front view road image. In this paper, we propose Lane-MGA: an end-to-end 2D and 3D lane joint detection network model, which aimes to promote the effective detection of both through mutual supervision of 2D lane and 3D lane. In addition, we propose a multi-granularity lane attention mechanism, which aggregates the spatial features and channel features of lane at different granularities. The experimental results verify the effectiveness of Lane-MGA.