A Comparative Study of the Performance of Personnel Detection based on Visible and Infrared Images
Futao Liu, Zhenling Li, Yuezhong Wan, Weida Cao, Duanwei Ma, Xiaolei Li, Yan Peng · 2022
Personnel security has always been a hot topic of research especially in power systems. Effective detection of people in power scenes has become a high priority. Currently, algorithms for personnel detection have been developed considerably, and they are mainly used for visible images. Since infrared images imaged by thermal radiation can detect living blood creatures easily, they can be used in personnel detection at night. To investigate the effect of visible and infrared images on the performance of personnel detection algorithms in different environments, this thesis uses two object detection algorithms, YOLOv5 and Faster RCNN, to train personnel detection models for visible daytime, visible nighttime, infrared daytime and infrared nighttime datasets respectively. Comparative experiments are conducted to quantitatively and qualitatively analyze the performance advantages and disadvantages of the personnel detection algorithms for visible images and infrared images during the day and night. The experimental results for both models consistently show that the visible images are similar to the infrared images during daylight hours with a good target detection model, the visible image is slightly better in performance. While the infrared image is significantly better than the visible image at night, and the [email protected] can be improved by 0.2. For small human target detection, the infrared image has a significant advantage over the visible image at night, with an average detection size reduction of 48.0% and a minimum detection size reduction of 64.6%.