Abstract
Real-time multi-object tracking (MOT) requires detector-tracker configurations that balance detection accuracy, association reliability, identity preservation and computational efficiency, particularly in central processing unit (CPU)- only and resource-constrained inference settings. However, previous You Only Look Once (YOLO)-based MOT studies have often evaluated isolated detector-tracker pairs, domain-specific scenarios, or heterogeneous hardware settings, which limits direct comparability across configurations. This study evaluates fourteen detector-tracker configurations using seven lightweight YOLO-based object detectors, namely YOLOv3-tiny, YOLOv5n, YOLOv8n, YOLOv9t, YOLOv10n, YOLO11n and YOLO12n and two online trackers, BoT-SORT and ByteTrack. Experiments were conducted on MOT17 and MOT20 under a unified CPU-only private-detection protocol. Each configuration was assessed using metrics covering detection accuracy, tracking performance, identity preservation and computational efficiency. No single configuration dominated all evaluation dimensions. BoT-SORT generally showed stronger tracking performance and identity preservation, as indicated by higher tracking and identity-related metrics and fewer identity switches, whereas ByteTrack showed higher throughput and shorter runtime. Matched-detector comparisons supported this trade-off, indicating that BoT-SORT was more favorable for association and identity continuity, whereas ByteTrack was more favorable for computational efficiency. Among the evaluated detectors, YOLOv8n, YOLOv10n and YOLO11n showed relatively balanced accuracy-efficiency profiles across datasets and metrics. These findings support the interpretation that CPU-only MOT performance is shaped by detector-tracker interaction rather than detector accuracy or runtime efficiency alone. Therefore, configuration selection should be aligned with deployment priorities, particularly the tradeoff between identity continuity and throughput.
Keywords: BoT-SORT, ByteTrack, CPU-only Inference, Detector-tracker Configuration, Lightweight YOLO Detectors, Multi-object Tracking.