YOLOv5 Integrated With Recurrent Network for Object Tracking: Experimental Results From a Hardware Platform

Abstract

Advances in object tracking integrate feature-based methods with modern deep learning. Initial object detection is crucial for continuous tracking. Still, challenges include establishing reliable data associations across frames amidst issues like motion blur, lighting variations, occlusions, and differing object sizes, particularly notable in surveillance and autonomous navigation. This study investigates achieving real-time object tracking by integrating YOLOv5 with recurrent networks. A significant breakthrough in real-time implementation involved introducing multithreading into YOLOv5s detector input, trained on the VisDrone2019 dataset, which notably reduced latency. The testing video utilized in the experiment contained eight specific issues about aerial objects, namely: identical targets, multi-scaling, partial occlusions, full occlusions, changeable illuminations, fast target movement, out-of-view tracking, and low-resolution adaptation. Implemented on Nvidia’s Edge AI platform, this approach demonstrated superior performance with an F1 score of 93.31 for Intersection over Union values exceeding 0.5, achieving an impressive frame rate of 167.147 frames per second in just 0.0024 seconds. Comparative experimental findings strongly support the real-time tracking of objects in different scenic situations from an aerial platform.

Keywords

• Real-time object tracking
• recurrent network
• you look only once (YOLO)v5