Temperature Effect on Thermal Imaging and Deep Learning Detection Models

Abstract

Infrared cameras can be a great supplement to the environmental perception systems for autonomous driving. Compared to optical cameras, radars, or Lidars, infrared cameras exceed in detecting heat-radiating objects, such as humans and animals, potentially improving the safety of autonomous cars. The underlying detection algorithms for infrared images are generally the same deep learning models applied for optical cameras. However, as the working principles of infrared and optical cameras are different, so are the images they produce. This paper presents the visual differences in infrared images caused by ambient temperature changes and examines their effect on deep learning detectors. Specifically, this study investigates two infrared datasets, one from McMaster University CMHT group and the other from the FLIR company. They represent a northern cold climate and a tropical climate, respectively. Two YOLO-based object detection models are trained on the two datasets separately. The evaluation results show that a colder temperature results in a better performance. In the meantime, cross-evaluation shows a sharp decrease in performance when evaluating the model against the opposite dataset. Furthermore, a third model trained using both datasets outperform the previous two models in all metrics. This study highlights the importance of ambient temperature in training infrared image detectors and provides a feasible solution to performance mismatch issues.