Modification of YOLOv4 for Human Detection in Elevated Thermal Images

Several deep learning networks have already been developed to detect objects given an input image. However, most of these networks are usually designed around detecting objects using a standard set of images or datasets such as MS COCO. Using these networks on a custom dataset already performs well but lacks context on the detected object. Furthermore, they can be improved upon by modifying the network depending on the simplicity or complexity of the dataset being used. With this, the paper presents two ways of modifying and improving a base network for detecting humans on the custom thermal elevated objects (TEO) human thermal dataset. The results show that TrimmedYOLO, the network created using the layer removal method, produced comparable results as the base YOLOv4 while being more lightweight regarding FPS and BFLOPS needed. On the other hand, TEO-EfficientYOLO-Mish and TEO-YOLO-ReT-Mish, the networks created using the backbone replacement method, can perform inference the fastest while still producing fairly high mAP results.