Deep analysis of YOLO model and/or correlation tracking algorithm

This video was developed in the context of the Pervasive AI Developer Contest with AMD (link: https://www.hackster.io/contests/amd2023). Starting with the 58th second of the video, as the UAV progresses, the original YOLOv3 model (YOLOv3 DarkNet) can follow the UAV without any problem. But the YOLOv3 model quantized and compiled for Kria KR260 (YOLOv3*) cannot follow the UAV up to the moment with the time mark 1:06 – this is the result of quantization that generates some accuracy losses of the YOLO model. But, based on the correlation tracking algorithm (CTA), the combined model (YOLOv3* + CTA) can perfectly follow the drone on 21 more frames. Please see the rest of the video in which the combined algorithm (YOLOv3* in blue + CTA in red) can track the UAV without missing any frames.