Road surface semantic segmentation for autonomous driving.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Semantic segmentation is crucial for autonomous driving but struggles with complex road environments.
• Existing methods often exhibit limitations in accurately segmenting road surfaces, especially with overlapping objects and boundary details.

Purpose of the Study:

• To propose a novel frequency-based semantic segmentation approach using a transformer architecture (FSSFormer).
• To enhance the performance of road surface segmentation in challenging traffic conditions.
• To improve the handling of overlapping targets and boundary information loss.

Main Methods:

• Developed a frequency-based semantic segmentation with a transformer (FSSFormer) model.
• Introduced a weight-sharing factorized attention mechanism to select critical frequency features.
• Employed a cross-attention method combining spatial and frequency features to refine boundary details.
• Utilized a parallel-gated feedforward network for position information encoding.

Main Results:

• The proposed FSSFormer demonstrated improved segmentation performance on complex road scenarios.
• Achieved a 2% increase in mean Intersection over Union (mIoU) compared to existing methods on the Cityscapes dataset.
• Successfully addressed challenges related to overlapping targets and boundary information loss.

Conclusions:

• FSSFormer offers a significant advancement in semantic segmentation for autonomous driving.
• The frequency-based approach effectively leverages frequency information for enhanced road surface segmentation.
• The method shows strong potential for real-world autonomous driving applications requiring high segmentation accuracy.