Multi-Task Environmental Perception Methods for Autonomous Driving.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving systems face challenges in environmental perception, including inaccurate detection of small objects and complex scenes.
• Current algorithms struggle with feature redundancy, limited contextual interaction, and poor information fusion, hindering multi-task efficiency.

Purpose of the Study:

• To develop an end-to-end multi-task environmental perception model for autonomous driving.
• To simultaneously perform traffic object detection, lane line detection, and drivable area segmentation.

Main Methods:

• Implemented a multi-stage gated aggregation network (MogaNet) for enhanced channel-wise contextual interaction.
• Introduced a restructured weighted bidirectional feature pyramid network (BiFPN) for optimized multi-scale object detection.
• Utilized a unified model with one detection head and two segmentation heads for efficient multi-task processing.

Main Results:

• Achieved 81.4% mAP50 for object detection on the BDD100K dataset.
• Obtained 28.9% IoU for lane detection and 92.6% mIoU for drivable area segmentation.
• Demonstrated effective real-world performance, significantly improving environmental perception.

Conclusions:

• YOLO-Mg effectively addresses limitations in current autonomous driving perception models.
• The model provides a robust foundation for safer and more reliable autonomous driving systems.
• Simultaneous multi-task learning enhances overall environmental perception capabilities.