Object Detection on Thermal Images for Unmanned Aerial Vehicles Using Domain Adaption Through Fine-Tuning

This work addresses state-of-the-art object detection methods using deep learning on thermal images for application on Unmanned Aerial Vehicles (UAVs). For this purpose, fine-tuning is performed using a custom dataset. Special focus is given to the generation of this dataset, as the annotations for the thermal images are automatically generated from simultaneously acquired visual images. The bounding boxes found on visual images using state-of-the-art object detection methods are applied as annotations to the thermal images. Furthermore, it is shown how the fine-tuned models can be executed in real-time on the drone's embedded PC, which is limited in its computing power, by using additional accelerator hardware.