Comparison of Real-Time Multi-object Tracking with Limited Hardware Resources

Abstract

Real-time Multi-Object Tracking (MOT) on local devices is critical when privacy, latency, and connectivity limit cloud solutions. Deploying deep models on resource-constrained hardware remains challenging, so this paper investigates the trade-offs between accuracy, latency, and computational resources across smartphones, embedded GPUs, and desktop GPUs. A review of detection and tracking architectures highlights the evolution of You Only Look Once (YOLO) models and trackers such as SORT, ByteTrack, and BoT-SORT. We propose YOLO Floridsdorf (v1210), a novel detector with attention-based components and an NMS-free head to improve accuracy and reduce latency. Two domain-specific datasets simulate search scenarios for target object localisation. Experiments show that convolution-based detectors like YOLO v11 outperform attention-based models on devices with limited compute; for example, YOLOv11-s achieves 68.04% mAP at 12 Frames Per Second (FPS) on a smartphone, while YOLOv12-s achieves 67.32% mAP at 2 FPS. For tracking, ByteTrack offers the best balance of speed and robustness.