PortLaneNet: enhancing port RTG lane detection with explicit feature learning network

Lane detection plays a crucial role in automated driving technology. However, current lane detection algorithms primarily focus on daily scenes, such as cars on urban roads and highways, while limited research addresses closed scenes, such as rubber tire gantries (RTGs) in ports. In this study, we propose a novel lane detection method based on explicit feature learning (EFL) to address this gap. Compared to general scenes, RTG lane lines in port exhibit distinctive characteristics, including regular straight-shaped features, relative spatial stability, and severe wear from heavy RTGs. There is also a higher requirement on algorithm precision while with a limited dataset size. To effectively address these challenges, we proposed PortLaneNet, which introduces an EFL network for better feature extraction, coupled with a shape loss function for improved evaluation and supervision. The multiscale feature extraction of the EFL network improves the model's global perception ability to cope with the loss of features caused by the severe wear of the port lane lines. The improved shape loss function strengthens the linearity of the model's inference results to adapt to the shape characteristics of port lane lines. To improve the model's spatial feature extraction capability, an adaptive channel attention mechanism is also introduced, which embeds spatial positional information of rows. The proposed method is evaluated on a self-collected dataset under various challenging environments, where extensively comparative experiments against six state-of-the-art algorithms with various backbones show that the proposed PortLaneNet demonstrates better detection performance in terms of F1 score, precision, and recall while with computation efficiency suitable for real-time applications.