Abstract:

For the problem of lack of lane line quantity recognition and insufficient accuracy of lane line segmentation in complex environment for self-driving cars， a lane line detection method with a two-branch instance segmentation network structure is proposed. Firstly， the method uses an encoding-decoding framework to improve the detail recovery ability and support multilevel feature fusion. Secondly， the fusion of feature pyramid network and advanced residual network improves the model's understanding of contextual information and deep semantics， which efficiently extracts the semantic features in complex lane lines. Then， the introduction of the SE module and the weighted least-squares fitting module strengthens the model's overall feature expression and generalization ability so as to improve the flexibility and accuracy of the model， and enhance the geometric shape prediction of lane lines without losing the computational performance of the model. Finally， the F1 of the algorithm experimentally tested on CULane and TuSimple datasets reaches 76.0% and 96.9%， respectively， and the experimental results show that the method can obtain good detection performance in complex environment such as light change， occlusion and road damage， and multiple lanes， and effectively improves the lane line detection accuracy.

Key words: instance segmentation, lane detection, feature pyramid network, autonomous driving