Research on the optimization of local lane segmentation in BEV based on dynamic serpentine convolution

BEV high-definition maps play a crucial role in autonomous driving and navigation systems, where their segmentation accuracy directly affects system performance and safety. Traditional feature extraction networks, when dealing with complex BEV maps, are often limited by their fixed kernel sizes and shapes, leading to insufficient accuracy in critical tasks such as lane segmentation. This paper improves and proposes an A-HDN framework for high-definition segmentation in BEV, observing the slender and continuous features of linear structures and introducing a Dynamic Serpentine Convolution network (DSConv). This network can flexibly conform to the lane structures in BEV and learn features, while also staying close to the target structure under constraints, thus better learning the features. Additionally, a Ghost module is introduced, which allows the learning network to better preserve features without affecting model performance. Finally, experiments show that this algorithm has increased the map segmentation precision by 1.6 IoU and improved directional detection by 0.8 mAP.