Underground environmental perception: deep learning-based track segmentation for mine-use suspended monorails

Abstract Robust visual perception is a foundational challenge for automating transportation in the adverse conditions of underground mining. In this paper, we present a comprehensive case study of applying deep learning-based segmentation methods to detect and segment suspended monorail tracks. The key contributions of this work are twofold. First, we captured images of suspended monorail tracks in a workshop scenario by using an on-board camera mounted on a mine vehicle, with all images annotated by using specialized tools to construct a semantic segmentation dataset for mine use suspended monorails. Second, we systematically compared ten state-of-the-art (SOTA), up-to-date, and classic segmentation models to evaluate their segmentation accuracy, robustness, and computational efficiency for detecting monorail tracks in the workshop environment. Experimental results demonstrate that the U-Net architecture offers a favorable balance between performance and complexity compared to traditional methods. In terms of computational demand, U-Net requires only 8.25 GFLOPs and contains 7.44 M parameters for processing a 3-channel 224 × 224 image with a batch size of 1. Regarding the segmentation accuracy, the U-Net model can achieve a mean Intersection over Union (mIoU) of 87.91% and an average recall rate of 92.90% on our annotated dataset, outperforming all other baseline methods.