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

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Tactile pavement detection is crucial for guiding visually impaired individuals.
• Existing datasets lack diversity and sufficient sample sizes for robust model training.
• Semantic segmentation networks need to balance accuracy with computational efficiency for real-time applications.

Purpose of the Study:

• Introduce a novel semantic segmentation network, Multi-scale Residual Attention Convolution UNet (MSRAC-UNet), for tactile pavement detection.
• Develop a new, diverse dataset (BlindPath-SegDataset) to overcome limitations of existing tactile pavement datasets.
• Evaluate the performance and efficiency of MSRAC-UNet for real-time edge deployment.

Main Methods:

• Designed MSRAC-UNet incorporating multi-scale spatial feature fusion and a receptive field attention module.
• Utilized depthwise separable convolutions and varied receptive fields for computational efficiency.
• Created and utilized the BlindPath-SegDataset for comprehensive model evaluation.

Main Results:

• MSRAC-UNet achieved a competitive 95.12% Mean Intersection over Union (MIoU) on public benchmarks.
• The network outperformed transformer-based models like SegFormer in tactile pavement segmentation.
• On the BlindPath-SegDataset, MSRAC-UNet demonstrated real-time performance (45 FPS) with a small model size (21.80 MB).

Conclusions:

• MSRAC-UNet offers an optimal balance between accuracy and efficiency for tactile pavement detection.
• The proposed network and dataset significantly advance the capabilities for real-time edge deployment in assistive technologies.
• MSRAC-UNet shows superior performance compared to existing state-of-the-art models, particularly for practical, resource-constrained applications.