BRNet: a Bilateral Lane Detection Framework Based on Recurrent Feature-Shift Aggregator

The lane is an important representation of the normal driving of the intelligent vehicle. Due to the complexity of the detection environment (e.g., extreme illumination, vehicle occlusion, etc.) and the slender nature of lanes, Effective collection and fusion of features from different levels of the lane is the key to the task. But research in this area is not deep enough. In this work, a Bilateral Lane Detection Framework (BRNet) is proposed to acquire spatial features and rich contextual features by designing Spatial Path and Context Path. In addition, we propose a strategy to fuse low-level features with deep-level features and utilize a feature fusion module to achieve information integration for different-level features. Experiments on two different lane detection benchmark sets demonstrate the state-of-the-art performance of our method.