A Robust Thermal Infrared Vehicle and Pedestrian Detection Method in Complex Scenes

Area of Science:

• Computer Vision
• Machine Learning
• Infrared Imaging

Background:

• Detecting motion-blurred, tiny, and dense objects in thermal infrared images presents significant challenges.
• Existing methods struggle with the complexities of thermal imaging, especially in dynamic scenes.

Purpose of the Study:

• To propose a robust method for thermal infrared vehicle and pedestrian detection.
• To enhance the detection of motion-blurred, tiny, and dense objects in complex thermal scenes.

Main Methods:

• Introduced a weight parameter (β) to optimize the Feature Selective Anchor-Free (FSAF) module's loss function for motion-blurred object detection.
• Integrated optimized anchor-free FSAF branches with the YOLOv3 detector, combining them with anchor-based branches for improved precision.

Main Results:

• Achieved a mean average precision (mAP) of 72.2%, outperforming other typical thermal infrared detection algorithms.
• Demonstrated superior performance in detecting challenging objects like motion-blurred, tiny, and dense targets.

Conclusions:

• The proposed method offers an effective solution for motion-blurred object detection in thermal infrared imagery.
• The integration of FSAF and YOLOv3 significantly boosts detection precision for small and dense objects, advancing the field of thermal object detection.