EA-UNET: An Enhanced and Efficient Model for Left-Turn Lane

Area of Science:

• Computer Vision
• Deep Learning
• Autonomous Systems

Background:

• Left-turn lanes are crucial for urban intersection safety.
• Current lane detection methods struggle with accuracy, computational cost, and environmental factors.
• Autonomous vehicle navigation demands precise and efficient lane detection.

Purpose of the Study:

• To develop a lightweight deep convolutional neural network for accurate left-turn lane detection.
• To overcome the limitations of existing semantic segmentation algorithms.
• To enhance the safety and efficiency of autonomous vehicle navigation.

Main Methods:

• Proposed EA-UNet, a lightweight deep convolutional neural network.
• Replaced the standard U-Net encoder with EfficientNet-B0 for improved feature extraction.
• Introduced a novel MP-ASPP module with CBAM for refined attention mechanisms.
• Created a comprehensive real-world dataset for left-turn lane segmentation.

Main Results:

• EA-UNet demonstrated superior performance compared to baseline U-Net and other state-of-the-art models.
• Achieved accurate and efficient segmentation of left-turn lanes.
• Showcased robustness in complex urban intersection scenes.

Conclusions:

• EA-UNet offers a significant advancement in left-turn lane detection for autonomous vehicles.
• The proposed model provides a lightweight, accurate, and efficient solution.
• EA-UNet enhances the reliability of autonomous navigation systems in challenging conditions.