RailDepth: A Self-Supervised Network for Railway Depth Completion Based on a Pooling-Guidance Mechanism

Depth completion of forward objects is critical to safe and effective autonomous driving as it can estimate dense depth from sparse light detection and ranging (LiDAR) data and RGB images. The study proposes a self-supervised depth completion method to address the deficiency in depth sensing images caused by the uneven distribution of point clouds generated by long-range LiDAR. The proposed method uses a stacked-hourglass-network structure as the backbone, which allows to perform multilevel deep fusion to extract features from uneven point clouds. In addition, a fusion module is introduced to fuse global-local features from different stages of the network output, and a pooling-guidance mechanism is designed to guide network convergence through adaptive pooling-guidance depth construction. Field experiments were conducted on the Tsuen Wan Line in Hong Kong to verify the performance of the proposed network. The results indicate that the proposed network outperforms the benchmark self-supervised depth completion methods, with a root-mean-squared error (RMSE) of 8165.673 mm and a mean absolute error (MAE) of 2596.418 mm. Specifically, the RMSE and MAE were improved by 4.7% and 2.4%, respectively, compared to the benchmark methods.