DSC-Net: Enhancing Blind Road Semantic Segmentation with Visual Sensor Using a Dual-Branch Swin-CNN Architecture

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Visual sensors are vital for urban navigation systems, especially for the visually impaired.
• Semantic segmentation of blind roads faces challenges in extracting global context and edge features.
• Existing Convolutional Neural Networks (CNNs) struggle with global context and discontinuous feature detection.

Purpose of the Study:

• To develop an advanced method for accurate semantic segmentation of blind roads.
• To enhance the understanding of complex urban environments for navigation systems.
• To improve the detection of road boundaries and occluded objects.

Main Methods:

• Introduced Dual-Branch Swin-CNN Net (DSC-Net), integrating Swin-Transformer and U-Net architectures.
• Employed Spatial Blending Module (SBM) to reduce blurring from object occlusion.
• Utilized hybrid attention module (HAM) within Inverted Residual Module (IRM) for boundary sharpening and processing speed.

Main Results:

• Achieved a mean Intersection over Union (mIoU) of 97.72% on a specialized blind road dataset.
• Demonstrated exceptional performance on additional public datasets.
• Successfully enhanced the extraction of global context and edge features for blind roads.

Conclusions:

• DSC-Net effectively overcomes limitations of traditional CNNs in blind road semantic segmentation.
• The proposed method significantly improves accuracy and boundary detection in challenging road scenarios.
• DSC-Net offers a promising solution for enhancing navigation assistance technologies.