Multi Object Tracking with UAVs using Deep SORT and YOLOv3 RetinaNet Detection Framework

Abstract

Over the years, object tracking and detection has emerged as one of the most important aspects of UAV applications such as surveillance, reconnaissance, etc. In our paper, we present a tracking-by-detection approach for real-time Multiple Object Tracking (MOT) of footage from a drone-mounted camera. Tracking-by-detection is the leading paradigm considering its computational effectiveness and improved detection algorithms. Our algorithm builds on the baseline Deep SORT algorithm implemented for MOT benchmarks. However, to circumvent the challenges posed by videos captured from a significant height we use a combination of YOLOv3 and RetinaNet for generating detections in each frame. The results of our experiment on the VisDrone 2018 dataset exhibit competitive performance in comparison to the existing trackers.