Non-local Spatial Information Based Lane Detection

Lane detection mainly obtains road images through vehicle cameras, and then detects lanes on the road and provides the exact location and shape of each lane. It can be widely used in autonomous driving and advanced driving assistance system. In order to improve the adaptability of lane detection model to urban and poor-lighting environments, a new lane detection method is proposed, which uses a non-local spatial information module to capture and fuse spatial relationships with long-range dependencies, that increases the poor environmental adaptability of the lane detector. We validated our methods on the lane detection benchmark CULane. It demonstrates adaptability under the condition of blocked and poor-light conditions, and robustness in complex scenarios.