Enhanced Detection and Recognition of Road Objects in Infrared Imaging Using Multi-Scale Self-Attention.

Area of Science:

• Computer Vision
• Machine Learning
• Infrared Imaging

Background:

• Detecting small, low-contrast targets in infrared imagery, especially in traffic, is challenging.
• Traditional methods struggle with feature extraction and suffer feature loss during transmission.
• Existing algorithms like YOLOv8 have limitations in processing infrared small targets effectively.

Purpose of the Study:

• To develop a fast and accurate detection and recognition model for small road targets in infrared scenarios.
• To enhance feature extraction and prevent information loss in infrared small target detection.
• To improve the real-time processing capabilities of infrared target detection systems.

Main Methods:

• Proposed a novel model based on YOLOv8, integrating an improved DyHead structure with multi-scale self-attention.
• Introduced and enhanced the Gather-and-Distribute Mechanism to mitigate feature loss in the Feature Pyramid Network (FPN).
• Optimized the network architecture through pruning to reduce computational complexity and enhance processing speed.

Main Results:

• Achieved a 3.1% improvement in Average Precision (AP) and a 2.5% increase in mean Average Precision (mAP) on the VisDrone2019 dataset.
• Demonstrated significant enhancements in detection accuracy and processing speed for small infrared targets.
• Showcased improved robustness and adaptability in complex road traffic environments.

Conclusions:

• The proposed multi-scale self-attention model effectively addresses the challenges of small infrared target detection.
• The integration of DyHead and Gather-and-Distribute Mechanism significantly boosts feature perception and reduces information loss.
• The pruned, fast detection model is suitable for real-time applications in infrared road traffic monitoring.