Robot steering-angle prediction lightweight network based non-local attention and lane guidance

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Accurate steering angle prediction is crucial for autonomous robot navigation.
• Existing methods often face challenges with computational efficiency and prediction accuracy in complex environments.

Purpose of the Study:

• To develop a novel end-to-end prediction network for robot steering angle prediction.
• To improve computational efficiency and prediction accuracy using integrated attention and lane line guidance mechanisms.

Main Methods:

• A ResNet-based network incorporating Non-Local Block and Ghost Module for enhanced feature extraction and global context modeling.
• A lane line annotation method combining Canny edge detection and Hough transform for semantic guidance.
• Utilized ReduceLROnPlateau scheduler for adaptive learning rate adjustment to mitigate overfitting.

Main Results:

• Achieved a 54.88% increase in inference speed compared to baseline models.
• Reduced Mean Absolute Error (MAE) by 8.47% and Root Mean Square Error (RMSE) by 18.23%.
• Ablation studies confirmed the effectiveness of Non-Local Block and Ghost Module in improving model performance and efficiency.

Conclusions:

• The proposed network offers a high-precision, efficient, and low-latency visual perception solution for real-time autonomous navigation.
• The integration of non-local attention and lane line guidance significantly enhances steering angle prediction.
• The method demonstrates strong generalizability and potential for deployment in complex robotic navigation scenarios.