CBAM meets DropBlock: enhancing robot steering-angle prediction with hybrid attention and structured dropout

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Accurate steering angle prediction is vital for intelligent auto-navigation in complex scenarios.
• Traditional end-to-end systems struggle with prediction accuracy and feature interference.

Purpose of the Study:

• To propose a lightweight steering angle prediction network.
• To enhance prediction accuracy by addressing interference from complex scene features.

Main Methods:

• A three-stage progressive framework with multi-scale feature pyramid, channel/spatial attention decoupling, and dynamic region feature dropout.
• Bi-domain attention fused region mask and de-fuzzy network for cooperative optimization.
• Dual-channel feature compression, multilayer perceptron for channel weight generation, and gated convolutional kernel for spatial attention.

Main Results:

• The proposed method effectively strengthens key semantic features like lane lines and obstacles.
• Interfering information from non-critical regions (road texture, sky) is suppressed.
• Experimental results demonstrate strong performance in steering angle prediction.

Conclusions:

• The developed lightweight network shows significant improvements in steering angle prediction accuracy.
• The cooperative optimization strategy effectively handles complex scenarios.
• The method holds promising applications for intelligent auto-navigation systems.