Fast Multi-Lane Detection Based on CNN Differentiation for ADAS/AD.

Lane detection is a challenging task in Advanced Driving Assistance System / Autonomous Driving (ADAS/AD) with strict requirements on real-time and accuracy. Inspired by the differential calculus, we propose a fast multi-lane detection method to achieve high accuracy while satisfying the real-time requirement of ADAS/AD. We formulate the lane lines into a series of endpoints of differential line segments. A convolutional neural network is then presented to simultaneously predict the heatmap and the embedding vectors, whose loss function is designed to detect the semantic lane lines instead of focusing only on the visible lane lines. The heatmap is used to represent the endpoints in lane differentiation while the embedding vectors are used to group the endpoints into different lane lines. Finally we utilize a non-maximum suppression algorithm to combine the heatmap with the embedding vectors to obtain the final results of multiple lane lines. We conduct extensive experiments on TuSimple and CULane datasets to evaluate the proposed method. The experimental results demonstrate the efficiency of our approach, which can achieve the highest accuracy with 46.0 times speedup compared with eight state-of-the-art methods.