EnRDeA U-Net Deep Learning of Semantic Segmentation on Intricate Noise Roads

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Road segmentation is crucial for autonomous systems like the Self-Driving Sweeping Bot (SDSB).
• Real-world road conditions present significant challenges due to noise from weather, lighting, obstacles, and road degradation.
• Existing U-Net architectures struggle with these complex environmental factors.

Purpose of the Study:

• To develop and validate an enhanced U-Net model, EnRDeA U-Net, for improved road segmentation.
• To evaluate the performance of EnRDeA U-Net against other U-Net extensions on a dataset with intricate road noises.
• To analyze the features and segmentation capabilities of different U-Net models in adverse conditions.

Main Methods:

• Proposed the EnRDeA U-Net, incorporating a Residual U-Net block as an encoder and an attention gate as a decoder.
• Utilized a dataset featuring complex road noise conditions (sunshine, shadows, obstacles, cracks, varied materials).
• Compared EnRDeA U-Net with Primordial U-Net and Residual U-Net, analyzing their structures, parameters, and performance.

Main Results:

• The EnRDeA U-Net demonstrated effective validation on the intricate road noises dataset.
• Detailed analysis of network features and segmentation performance was conducted for all three U-Net extensions.
• Experimental results presented comprehensive data on network structures, training losses, and performance metrics.

Conclusions:

• The EnRDeA U-Net shows promise for enhancing road segmentation accuracy in challenging real-world scenarios.
• The study provides valuable insights into the performance of different U-Net architectures under noisy conditions.
• Findings contribute to the development of more reliable autonomous driving systems.