Learning Dynamic View Reliability for Multi-Camera Multi-Object Tracking

Abstract

Multi-Camera Multi-Object Tracking (MC-MOT) aims to maintain consistent identities across views, yet it remains challenging when visual evidence is corrupted by view-dependent occlusions and varying visibility. While recent graph-based methods offer flexible spatio-temporal modeling, most rely on view-invariant fusion strategies that treat all camera inputs as equally reliable. This “blind” aggregation allows noise from occluded views to propagate into the target representation, leading to frequent identity switches. To address this limitation, we propose VistaTrack, a hierarchical graph-based framework that explicitly models the time-varying reliability of camera views. Central to our approach is a novel Dynamic Feature Fusion (DFF) module inserted between spatial clustering and temporal linking. By employing a dual-stream robust consensus mechanism, DFF dynamically estimates instance-level reliability scores, effectively suppressing noisy observations while emphasizing informative viewpoints. The resulting reliability-weighted identity embeddings are then associated across time to form coherent trajectories. Extensive experiments on the WildTrack, CAMPUS, and EPFL benchmarks demonstrate that VistaTrack achieves new state-of-the-art performance, showing significant improvements in IDF1 and tracking stability under severe occlusions.