COMPARATIVE ANALYSIS OF LIGHTWEIGHT TRACKERS UNDER CORRELATED AND UNCORRELATED NOISE

Abstract

The research focuses on the robustness of lightweight single-object trackers that can operate in (near) real time on weak CPUs such as Raspberry Pi 4 and 5 when the input video is corrupted by noise. The goal is to compare how different families of embedded-friendly trackers – correlation-filter, Siamese and transformer-based – degrade under additive white Gaussian noise (AWGN) and salt-and-pepper noise (S&P), each applied in four combinations of temporal and channel-wise correlation, and to identify algorithms that offer the best robustness–efficiency trade-off for embedded vision systems. The results show that all trackers exhibit monotonic AUC degradation as noise intensity increases, with Gaussian noise causing only moderate accuracy loss and salt-and-pepper noise leading to much steeper performance drops. HiT, ViTTrack and NanoTrack v3 achieve the highest accuracy on clean or mildly degraded data; however, HiT and NanoTrack v3 degrade faster than NanoTrack v2 and are overtaken by NanoTrack v2 at medium and high noise levels in all settings except temporally correlated AWGN for NanoTrack v3. ViTTrack is the only tracker that consistently maintains higher AUC than NanoTrack v2 across all noise intensities and correlation regimes, providing the most stable performance. ECO and DaSiamRPN lose accuracy much faster, especially under salt-and-pepper noise. Salt-and-pepper noise is generally more destructive than AWGN, while the temporal and channel-wise correlation of noise causes noticeable but moderate shifts in the degradation profiles without changing the overall ranking of trackers.