Human Detection in Infrared Image Using Daytime Model-Based Transfer Learning for Military Surveillance System
Eun Seop Kim, Whui Kim, Ju-Derk Park, Kunmin Yeo · 2023
In the field of civil and military surveillance, there has been a gradual increase in the application of deep learning-based image object detection systems to enhance accuracy. During daylight hours, the desired objects can easily be detected using cameras combined with image processing or deep learning techniques. However, during nighttime or when weather conditions prevent daylight, detecting objects through regular RGB cameras becomes challenging [1]. This paper proposes a method for detecting humans using images captured from infrared cameras, aiming to facilitate whole day surveillance systems for military bases. Typically, public image datasets almost consist of RGB images taken during the day [2]. However, gathering infrared image datasets is both essential and labor-intensive. Therefore, to minimize the collection of such infrared datasets, this study suggests using transfer learning based on daytime models to detect humans in infrared images. With a limited amount of collected infrared data, this transfer learning approach can enhance the accuracy of detection.