Mining Image Semantics via Deep Learning: A Robust Lane Detection Approach for Autonomous Driving

Abstract

Autonomous driving has attracted huge research interest from both academia and industry. As one of the key components for the safe driving of autonomous vehicles, lane detection allows vehicles to correctly locate itself in the lane and follow the traffic rules. Unlike traditional detection methods that rely on the extraction of professional and hand-designed features, this paper proposes a robust lane detection method by mining semantic information via the deep learning model LaneNet, which can cope with more complex road scenes, and relieve the restriction of deep learning models that detect a fixed number of lanes. Specifically, we first utilize the LaneNet to segment lane pixels from the road scene. Then we distinguish different lane instances by using a clustering loss function based on the distance vector of lane pixels. Finally, we verify our method on two datasets, Tusimple and CULane. The results show that the detection accuracy of Tusimple is up to 94.3%, CULane’s normal level is 90.4% and other more complex levels can reach up to 70%. Furthermore, by comparing with existing approaches, simulation results confirm the robustness of the proposed method in lane detection. In addition, we combine lane detection with driving decision based on intelligent driving simulation platform PanoSim5, which illustrates the effectiveness of our proposed lane detection method for autonomous driving.