MSRAC-UNet: a multi-scale receptive field attention-enhanced network for tactile pavement segmentation

This study introduces Multi-scale Residual Attention Convolution UNet (MSRAC-UNet), a novel semantic segmentation network, to our knowledge, tailored for tactile pavement detection. The network incorporates multi-scale spatial feature fusion and a receptive field attention module to improve segmentation accuracy, especially in complex environments. By employing a combination of different-scale receptive fields and depthwise separable convolutions, MSRAC-UNet achieves computationally efficient performance while maintaining high accuracy. Furthermore, we introduce BlindPath-SegDataset (a novel dataset, to our knowledge, for tactile pavement detection) to address the limitations of small sample sizes and the lack of diversity in existing datasets. Extensive experiments demonstrate that MSRAC-UNet achieves an optimal balance between accuracy and efficiency. It yields a competitive 95.12% MIoU on public benchmarks, outperforming recent transformer-based models like SegFormer. On the BlindPath-SegDataset, it maintains 45 FPS with only 21.80 MB of parameters, demonstrating its superiority for real-time edge deployment.