An Enhanced Multi-Object Tracking Model Used in Low-Frame-Rate Video

Abstract

Surveillance and remote monitoring systems operating under bandwidth and storage constraints commonly record at extremely low frame rates, often as low as 1 frame per second (fps). At this temporal resolution, the core assumptions underlying conventional multi-object tracking (MOT) break down. Motion prediction based on the Kalman filter becomes unreliable because interframe displacements exceed its predictive range. At the same time, spatial overlap between consecutive frames approaches zero, rendering Intersection over Union (IoU)-based matching uninformative. Under these conditions, tracking becomes primarily appearance-driven. This work examines MOT behavior under extreme temporal sparsity and proposes Low-Frame-Rate Multi-Object Tracking (LFR-MOT), a purely appearance-based tracker that removes motion prediction entirely and relies on re-identification (ReID)-based appearance features with a two-stage matching strategy to handle detection uncertainty. Experimental results at 1 fps on UA-DETRAC and VisDrone show substantial improvements over motion-based and hybrid baselines. On VisDrone, existing motion-based methods yield Identity F1 (IDF1) as low as 17.3%, whereas LFR-MOT achieves 75.0%. On UA-DETRAC, LFR-MOT reaches 63.6% IDF1 under the same sparse temporal sampling. A separate within-dataset analysis on CityFlowV2 examines sensitivity to frame-rate reduction, and additional evaluations on real-world surveillance footage show that the method preserves identity consistency in challenging low-quality scenarios. Feature-based matching therefore provides a practical solution for surveillance systems operating under severe resource constraints and long interframe intervals.