UAV Multi-object Tracking by Combining Two Deep Neural Architectures

Detecting and tracking multiple objects from unmanned aerial vehicle (UAV) videos is an high challenging task in a wide range of practical applications. Almost all traditional trackers meet some issues on UAV images due to camera movements causing view change in a 3D directions. In this work, we propose a Convolutional Neural Network specialized in multi-object tracking (MOT) for images captured from UAV. The architecture we introduced is composed by two main blocks: i) an object detection block based on YOLOv8 architecture; ii) an association block based on strongSORT architecture. We investigated different versions of YOLOv8 architectures with the strongSORT as association trackers. Experimental results on the VisDrone2019 dataset show that the proposed solution outperforms the up to date state-of-the-art tracking algorithms performance on UAV videos reaching the 42.03% in Multi-Object Tracking Accuracy (MOTA).