A Thermal Infrared Pedestrian-Detection Method for Edge Computing Devices.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Infrared Imaging

Background:

• Thermal imaging excels in varied lighting but struggles with weak textures, occlusion, and small objects.
• High-performance models face latency issues on edge devices with limited computational power.

Purpose of the Study:

• To propose a real-time thermal imaging pedestrian detection method optimized for edge computing.
• To address challenges of weak texture, object occlusion, and small object detection in thermal imagery.
• To achieve efficient and low-latency pedestrian detection on resource-constrained edge devices.

Main Methods:

• Multi-scale mosaic data augmentation to improve object diversity and texture.
• Integration of a parameter-free attention mechanism to enhance feature representation with minimal computational overhead.
• Quantization and multi-threading techniques for accelerating multi-channel video detection on edge devices.
• Development of a high-quality thermal infrared dataset (YDTIP) for research and evaluation.

Main Results:

• The proposed method demonstrates improved performance in complex environments by enhancing object diversity and texture.
• The parameter-free attention mechanism effectively boosts feature enhancement without significant computational cost.
• Quantization and multi-threading accelerate video detection on edge devices, enabling real-time performance.
• Experimental validation on multiple datasets, including the self-built YDTIP, shows competitive advantages over existing methods.

Conclusions:

• The developed real-time thermal imaging pedestrian detection method is effective for edge computing applications.
• The combination of data augmentation, attention mechanisms, and optimization techniques successfully tackles limitations of existing systems.
• The new dataset facilitates further research and development in thermal infrared pedestrian detection.