SOLOLaneNet: Instance Segmentation-Based Lane Detection Method using Locations

Abstract

Dealing with complex and variable road scenes is a challenge for autonomous driving, so lane detection with good performance is crucial. Currently, mainstream lane detection methods are divided into semantic segmentation-based and instance segmentation-based methods. The former is limited to detect a pre-defined, fixed number of lanes and unable to cope with lane changes, while the latter requires post-clustering processing despite a variable number of lanes detected. In this paper, we propose a novel instance segmentation-based method using lane locations to get an arbitrary number of lane instances directly. In addition, in order to improve the speed of lane detection, we employee key points to represent lanes instead of pixels to reduce the granularity of segmentation. In conclusion, we propose a real-time lane detection method which can predict an arbitrary number of lanes directly. We validate our method on two public datasets and the competitive results are obtained by comparing with existing methods.