A Cross-Layer Feature Aggregation and Distribution Method for Lane Detection

Lane detection is crucial for achieving environmental perception in autonomous driving. Unlike ordinary objects, lane lines are usually longer, narrower, and more easily obstructed. Therefore, lane detection networks need to utilize features at different levels to accurately locate lanes fully. Most existing methods use Feature Pyramid Network (FPN) to fuse information from different levels. However, FPN structures only fuse features from adjacent layers, indicating that the fusion among cross-layer features is indirect. This method leads to information loss during the fusion process, making it difficult to locate lanes accurately. Hence, we propose a cross-layer feature aggregation and distribution method that enables direct communication among different feature layers to enhance progressive fusion in lane detection. In addition, we introduce a context augmentation module to enrich contextual information and boost the detection performance of the network in scenarios such as lane occlusion and extreme lighting. Numerous tests conducted on the CULane and TuSimple datasets illustrate that our proposed approach surpasses various representative lane detection networks, highlighting its effectiveness in managing complex conditions.