Forward long-distance 3D reconstruction in rail transit scenarios based on occupancy networks

In rail transit autonomous driving scenarios, real-time three-dimensional (3D) reconstruction is crucial for understanding scenes and ensuring the safety of the driving environment. Urban rail trains, with their substantial weight and long braking distances, necessitate an extended forward perception range in 3D space. To address this challenge, this paper proposes a method for forward long-distance 3D scene reconstruction tailored for rail transit scenarios based on occupancy networks. Firstly, a 3D feature representation method using three mutually perpendicular spatial planes is proposed to mitigate the high computational complexity of spatial voxel features. Secondly, considering the characteristics of forward binocular vision in rail transit scenarios, we employ self-attention and cross-attention mechanisms to fuse features between different images. Thirdly, due to the limitations in projection distance of the Light Detection and Ranging (LiDAR) point cloud used for supervision, we introduce a method to generate long-distance dense ground truth during the training stage. By pioneering the application of occupancy networks in rail transit scenarios, this approach significantly extends the forward perception range of autonomous driving trains, achieving an impressive 77.79% mean Intersection over Union (mIoU) accuracy.