FlipNet: An Attention-Enhanced Hierarchical Feature Flip Fusion Network for Lane Detection

Lane detection is a vital task in the field of autonomous driving for it provides valuable information on drivable locations. However, complex scenarios like severe occlusion, discontinuous lane appearance, and illumination variation still hinder the accurate detection of lanes. This paper presents FlipNet, a novel and efficient neural network that detects lanes in complex environments by taking advantage of feature flip fusion and attention mechanism. First, a hierarchical feature flip fusion module (HFFF) is developed to utilize spatial information and aggregate global content. HFFF constructs a hierarchical structure consisting of multiple scales of sub-feature maps and uses flip fusion to pass spatial information in a two-way manner. Then, a double-layer attention enhancement mechanism (DAEM) and a dual-pooling coordinate attention (DCA) are proposed to enhance the features extracted by the encoder backbone. DAEM highlights valuable features and reduces background noise, which helps the network better capture the long-range dependent lane structure and be more robust in challenging scenarios. Experiments show our method achieves state-of-the-art performance and obtains new best results among segmentation-based methods in three popular lane detection benchmarks: CULane, Tusimple, and LLAMAS.