Point Instance Segmentation Considering Feature Enhancement for Lane Detection

Accurate feature extraction is essential for lane detection using the deep learning method. This paper proposes an endto-end lane detection framework for feature-enhanced point instance segmentation. The framework consists of two networks. First, a feature enhancement network that expands receptive field by retaining more lane semantic information provides local and global multi-directional hierarchical features for the point instance segmentation of Stacked Hourglass Networks. Then, a point instance segmentation of Stacked Hourglass Networks for the lane classification. Thus, the lane classification of clustering feature points is transformed into problems of point instance segmentation, which makes the detection network not limited by the number of fixed lanes, reduces the estimation of non-essential prediction points, and increases the accuracy of lane identification. Finally, the classic TuSimple dataset verifies the proposed network, and the accuracy precision is up to 93.6%, with false positive (FP) 0.1 and false negative (FN) 0.07.