Autotrinet YOLO triple attention framework for robust traffic sign detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Traffic sign detection (TSD) is crucial for intelligent transportation systems but faces challenges like small targets and environmental variability.
• Existing deep learning methods, including YOLO, struggle to balance accuracy, efficiency, and robustness in diverse scenarios.

Purpose of the Study:

• To propose AutoTriNet-YOLO, a novel framework enhancing TSD by integrating triple-attention, dynamic feature fusion, and adaptive computation.
• To address limitations in current TSD methods regarding accuracy, computational efficiency, and robustness.

Main Methods:

• Introduced the TriplePathBlock module, parallelizing Convolutional Block Attention (CBAM), Non-local Blocks, and Lite Transformer for multi-scale contextual dependencies.
• Implemented a Dynamic Fusion Gate for adaptive attention path weighting and a Selective Insert mechanism for pruning redundant operations.

Main Results:

• Achieved state-of-the-art performance on a traffic sign dataset with 86.6% mAP@50 and 65.3% mAP@50-95.
• Outperformed existing methods like TSD-YOLO and EDN-YOLO, with ablation studies confirming component contributions, especially CBAM for local features.

Conclusions:

• AutoTriNet-YOLO offers a scalable, computationally optimized architecture for robust TSD by unifying diverse attention mechanisms.
• The framework's real-time efficiency makes it suitable for edge deployment in autonomous driving systems.