An improved YOLOv5 for real-time human detection in infrared images

Abstract

The recent prominence of using infrared technology for human detection has garnered attention, particularly for applications in intelligent video surveillance and self-driving systems. This technology offers advantages in adverse weather conditions and night vision. However, within deep learning, the challenge of variable illumination during human detection persists. This study presents a novel method for identifying individuals in thermal images by enhancing the you only look once (YOLOv5s) algorithm. The approach incorporates the Bi-directional feature pyramid network (BiFPN) and the convolutional block attention module (CBAM) to improve the model’s feature integration and extraction capabilities. Evaluation on established thermal imaging datasets confirms the method’s superiority over state-of-the-art convolutional neural networks (CNN) based techniques, achieving remarkable precision 99.1% and recall 96.9%.