A Multi-Task Road Feature Extraction Network with Grouped Convolution and Attention Mechanisms.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Autonomous driving faces challenges in complex environments, necessitating collaborative multi-tasking solutions.
• Multi-task learning networks have demonstrated efficiency in various domains like NLP and recommendation systems.
• Recent advancements have extended multi-task learning to visual road feature extraction.

Purpose of the Study:

• To propose an advanced multi-task road feature extraction network for autonomous driving.
• To integrate group convolution, transformer, and squeeze-excitation attention mechanisms for enhanced visual perception.
• To simultaneously address drivable area segmentation, lane line segmentation, and traffic object detection.

Main Methods:

• Development of a novel multi-task road feature extraction network.
• Integration of group convolution for efficient feature extraction.
• Incorporation of transformer and squeeze-excitation attention mechanisms for improved feature representation.

Main Results:

• The proposed network successfully performs multiple road perception tasks simultaneously.
• Experiments on the BDD-100K dataset show superior performance compared to existing algorithms.
• Achieved higher accuracy in drivable area segmentation, lane line segmentation, and traffic object detection.

Conclusions:

• The developed multi-task network offers a promising approach for autonomous road perception.
• Provides a foundation for generating highly accurate maps in visual-based autonomous driving.
• Advances the field of intelligent vehicle perception systems.