Harnessing Keypoints with Deep Metric Learning for Lane Line Instance Segmentation

Abstract

The paper mainly studies methods on lane line detection in autonomous driving applications. The specific aim of paper is to experiment on the instance-based segmentation technique for a lane line detection in images. We incorporate the process of selecting portions or patches of lane line pixels as keypoints. We then improve the proposed method with deep metric learning techniques, more specifically, with an associative embedding having an angular loss function, to ensure that, keypoints with similar features belongs to the same instance of lane line. Next, the instance detection of the lane line is completed by cluster fitting using Kmeans. We also experiment our modified lane line detection model on TuSimple and CULane datasets. To confirm the important applications of our proposed methods, we evaluated it and achieved some competitive accuracies. The experimental results show the feasibility of the proposed method