Person Detection in Thermal Images using Deep Neural Networks

Abstract

The efficient management and enhanced security of sports fields can be achieved through an automated occupancy analysis system. Thermal imaging technology provides unique advantages for occupancy analysis, making it an attractive option for sports field management and security. In this study, our objective is to present a deep neural network-based methodology for detecting individuals in thermal images, with the aim of enhancing sports field management and improving security measures. We use the Thermal Soccer dataset having 3000 images, which includes annotations for soccer players in the images, divided into three sets: training, validation, and testing sets. This research investigates the performance of three state-of-the-art deep learning models, namely YOLO, DETR, and EfficientDet, for the purpose of person detection in thermal images. The objective is to assess their effectiveness in accurately detecting humans in sports field environments. The performance of different versions of yolo, including Yolov5s, Yolov5m, Yolov5l, and Yolov5x is compared, on the Thermal Soccer dataset. Our results show that Yolov5x outperforms the other versions of Yolo, with a precision of 0.994, recall of 0.987, mAP@.5 of 0.995, and F1 score of 0.989. The performance of the yolov5 model with the highest score i.e. yolov5x is then compared to two other state-of-the-art models, DETR and EfficientDet. Our results show that Yolov5x outperforms DETR and Efficientdet with a map@.5 score of 0.994, followed by DETR with 0.962, and EffDet with 0.856. YOLOv5 X has the highest F1 score of 0.989, followed by DETR with 0.953 and EffDet with 0.841. Overall, our study demonstrates the effectiveness of deep neural network-based approaches for person detection in thermal images for sports field management and security. Our results show that YOLOX is the most effective model for this task, offering promising avenues for further research and application.