Research on Road Scene Understanding of Autonomous Vehicles Based on Multi-Task Learning.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Safe autonomous driving relies on accurate road scene understanding.
• Current visual perception systems require efficient models for simultaneous multi-task processing.
• Multi-task learning offers performance and computational advantages.

Purpose of the Study:

• To propose an efficient multi-task model for joint detection of traffic objects, drivable areas, and lane lines.
• To enhance the accuracy and computational efficiency of road scene understanding models.
• To address the need for compact, fast, and accurate perception models for autonomous vehicles.

Main Methods:

• Developed a multi-task learning model named YOLO-Object, Drivable Area, and Lane Line Detection (YOLO-ODL) using hard parameter sharing.
• Implemented a weight balancing strategy to automatically adjust model parameters during training.
• Utilized a Mosaic migration optimization scheme to improve model performance indicators.

Main Results:

• The YOLO-ODL model demonstrated strong performance on the BDD100K dataset.
• Achieved state-of-the-art results in terms of both accuracy and computational efficiency.
• Successfully integrated the detection of traffic objects, drivable areas, and lane lines into a single model.

Conclusions:

• The proposed YOLO-ODL model offers an effective solution for comprehensive road scene understanding.
• The model's efficiency and accuracy are well-suited for real-time applications in autonomous driving.
• The weight balancing strategy and optimization scheme contribute to the model's superior performance.