Comparative performance analysis of U-Net and DeepLabV3+ for semantic segmentation in traffic environments

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Intelligent traffic systems rely on computer vision for real-time scene understanding.
• Current semantic segmentation models struggle with low-quality traffic images (low light, blur, noise).
• This limitation hinders the development of robust autonomous vehicle perception systems.

Purpose of the Study:

• To develop a robust method for processing degraded traffic images.
• To enhance semantic segmentation and object detection accuracy in challenging visual conditions.
• To improve the reliability of perception systems for autonomous driving.

Main Methods:

• A sequential pipeline involving super-resolution (SR), semantic segmentation (SS), and object detection using YOLOv8x.
• Utilized U-Net and DeepLabV3+ for pixel-level semantic segmentation.
• Employed YOLOv8x for precise object detection and validation of segmentation masks.

Main Results:

• U-Net achieved PSNR of 41.93 dB, SSIM of 0.997, mIoU of 0.750, and mAP of 0.950.
• DeepLabV3+ resulted in PSNR of 46.03 dB, SSIM of 0.938, mIoU of 0.819, and mAP of 0.937.
• The integrated approach demonstrated improved performance on degraded traffic image data.

Conclusions:

• The proposed method effectively processes low-quality traffic images, enhancing scene understanding.
• Sequential SR, SS, and object detection improve the robustness of autonomous vehicle perception.
• This approach addresses critical limitations in current traffic scene analysis for safer autonomous driving.