Bus-Mounted Vision Sensing for Traffic Object Detection: BFTD and a Local-Global Attention Framework

Area of Science:

• Computer Vision
• Intelligent Transportation Systems
• Machine Learning

Background:

• Bus-mounted cameras offer unique perspectives for intelligent transportation systems.
• Detecting traffic objects from elevated bus viewpoints is challenging due to scale skewness, dense interactions, and occlusion.

Purpose of the Study:

• To introduce the Bus Front-view Traffic Dataset (BFTD) for robust traffic object detection.
• To propose YOLO-M2LA, an attention-based detection framework tailored for bus-mounted camera data.

Main Methods:

• Collected and annotated a high-resolution dataset (BFTD) with 8131 images and 56,137 instances.
• Developed YOLO-M2LA, a framework incorporating CBS-SPD for fine-grained information preservation and M2LA for local-global attention modeling.
• Conducted experiments on BFTD and public benchmarks to evaluate detection performance.

Main Results:

• The BFTD dataset effectively captures diverse real-world traffic scenarios, including challenging conditions like rain, fog, and occlusion.
• YOLO-M2LA significantly improves detection accuracy, especially for small and crowded traffic participants.
• The proposed framework achieves a practical balance between accuracy and efficiency.

Conclusions:

• BFTD and YOLO-M2LA provide valuable resources for advancing vision-based traffic sensing from bus-mounted cameras.
• The developed methods demonstrate effectiveness in overcoming common challenges in bus-view traffic object detection.
• The dataset and implementation are publicly available to support further research.