Real-Time Multi-Object Tracking with YOLOv8

Durmus Berk, Mert Adak, Berk Adak, Cansu Celik, Hasan F. Ates

Abstract: In this paper, we used advanced deep multi-object trackers for real-time object tracking and models trained with various datasets in YOLOv8 environment. From a wide range of currently developed trackers, the trackers with the best tracking capabilities are used for evaluation in this paper. These are SMILETrack, ByteTrack and BoTSort. Since these object trackers perform tracking via object detection, object detection with YOLOv8, which has passed many performance audits, was used with these trackers to improve tracking performance. The aim of this paper is to obtain a detection model that can accurately detect objects and track at least 15 frames per second in real-time using the trackers described above. In this paper, it is shown that the trained model is able to track 10 different classes of small objects in aerial videos with high accuracy without interruption.