Fast Lane Detection Flexibly Adapting to Road Structure Information

Modern commonly used lane detection algorithms usually transform the problem into an instance segmentation problem, and the speed and scalability of the algorithm need to be improved. At the same time, such a method ignores the structural information of the lane itself, which is precisely the important feature information in complex situations such as lane occlusion. In this regard, we propose a new method for the expansion of high -speed, challenging scenarios and unstructured roads. We adopt a row-based selection method, which can significantly reduce the computational cost. In addition, we propose a new structural loss function for the lane structure information to model the continuity and curvature of the lane. Extensive experiments on the CULane dataset show that our method has superior speed and accuracy. At the same time, we test it in the actual unstructured scenario, which proves that the method has certain scalability and flexibility.