Object and human detection system with termal camera images

Nowadays, computer vision and deep learning-based object detection technologies are widely used in critical domains such as security, search and rescue, and defense. However, traditional systems relying on RGB images suffer significant performance loss under adverse environmental conditions such as low light, fog, smoke, or dust. In this context, thermal imaging systems eliminate dependency on ambient light, enabling object detection even in low-visibility scenarios. In this thesis, a deep learning-based object detection system has been developed for detecting humans and body parts (head and arm) using thermal camera data. A total of nine models were trained using different variants (m, b, l) of the YOLOv10 architecture, and a comparative analysis was conducted with the region proposal-based Faster R-CNN architecture. A custom-labeled thermal image dataset was created on the Roboflow platform, and various data augmentation and preprocessing techniques were applied. The results demonstrated that the YOLOv10 architecture performs remarkably well in terms of speed and accuracy, achieving high precision on thermal image data when optimized configurations are used. Moreover, the proposed system's suitability for realtime applications and its capability to detect body parts at a fine level are among the unique contributions of this study.