MLM-Net: Streamlined Multi-Lane Detection Network with Spatio-temporal Memory for Video Instance Lane Detection

Lane detection is a crucial task in autonomous driving as it helps vehicles maintain stable and efficient driving in complex traffic environments. Image-based and video-based lane detection are two common methods, with video-based lane detection being more effective in handling complex scenarios such as lane occlusion and wear. However, most existing video-based lane detection methods suffer from information redundancy, where the same key frames of the same lane are repeatedly calculated during the detection of multiple lanes. Additionally, the spatio-temporal sequence information of video-based lanes is not fully utilized. To address these challenges, MLM-Net is proposed as a streamlined multilane network that facilitates the sharing of spatiotemporal information. The Global Memory Feature (GMF) module is redesigned to mitigate information redundancy across multiple lanes, while the Local Spatio-Temporal Feature (LSTF) module is introduced to enhance the utilization of spatiotemporal consistency information. Additionally, the Enhanced Global Memory Feature (EGMF) module is introduced to extract enhanced features and handle challenging detection scenarios. Experimental results demonstrate that our method achieves state-of-the-art performance for video instance lane detection on the VIL-100 dataset.

I. Introduction

Deep learning has greatly advanced lane detection, but challenges remain due to complex traffic conditions and road damages [10]. Most methods focus on static scenes, lacking adaptability to dynamic driving scenarios. Video lane datasets like VIL-100 [1] improve robustness and accuracy. Video-level detection with temporal consistency tackles occlusion and lane wear, enhancing performance.