论文地址:Keep your Eyes on the Lane: Real-time Attention-guided Lane Detection
解决real-time efficiency. LaneATT是一个anchor based的车道线检测模型。基于全局信息来推理车道线信息,以此来解决车道线被占用,缺少markers等。
Introduction
- 车道线检测的挑战:extreme light and weather, lane markings occluded by objects, data imbalance, real-time.
- 利用anchor-based feature pooling 实现高准度和实时性
How
LaneATT是单阶段anchor-based模型。从前试图提取特征图,然后利用anchor-based feature pooling提取每个anchor的特征。这些特征和global infromation结合利用attention机制(可以解决occlusion),在经过全连接生成最后的输出车道线。
- anchor-based feature pooling。anchor的表示: origin point O = ( x o r i , y o r i ) O = (x_{ori}, y_{ori}) O=(xori,yori),处在图像边缘; direction θ \theta θ.
- global information attention