Adversarial Stress-Testing and Failure Diagnosis in Multi-Object Trackers
Investigating deterministic failure modes and edge-case fragility across 16 multi-object tracking algorithms under synthetic and real-world adversarial stress.
A research exploration into why and when visual multi-object trackers fail. By synthesizing fine-grained perturbation axes (severe occlusions, trajectory crossings, detector dropouts, camera shake, and spatial noise), this study isolates algorithmic breakdowns across Kalman filter approximations, appearance re-identification embeddings, and association algorithms (ByteTrack, BoT-SORT, OC-SORT, DeepOCSORT, etc.).
2025-11-20 · confidence high
Pure motion-based association algorithms (e.g., standard Kalman filters) undergo rapid track fragmentation under high-frequency camera shake, whereas observation-centric trackers (OC-SORT) preserve trajectory consistency by recalculating virtual momentum over occlusion intervals.
2026-03-15 · confidence high
Re-ID appearance embeddings reduce ID switches during linear crossing trajectories, but exhibit catastrophic association latency and false swaps under sudden illumination changes or severe sensor noise compared to second-stage low-confidence IoU association (ByteTrack).
