Multi-Lane Detection via Multi-Task Network in Various Road Scenes

摘要

Lane detection is one of the most important technologies in the autonomous driving system. Conventional road segmentation-based or lane marking-based lane detection methods may fail in various complex scenes. To address the problem of lane detection in various road scenes, this paper proposes a multitask based approach. The approach simultaneously segments road and lane marking instances. The branches of the two tasks share the same encoder but employ different decoder. In order to train the network with different datasets, this paper devises a training strategy to make sure only one branch can be trained per iteration. Finally, a fusion algorithm is presented to fuse the results from road segmentation and lane marking instances segmentation. The multi-lane (ego-lane, left and right adjacent lanes) detection results are obtained. Tested with Cityscapes and TuSimple dataset, the accuracy of lane detection is satisfied for the road scenes with the lanes number from zero up to four. The experiments demonstrate that the proposed approach is accurate and robust for various road scenes.