Calibration-Free Passenger Re-ID in Mixed-Modality, Crowded Buses in East Africa

Public transport in sub-Saharan Africa (SSA) often operates under extreme crowding, with standing passengers filling aisles and riders visible only as heads or shoulders. This makes conventional Re-ID systems, trained on uncrowded, full-body outdoor images, fail when applied to on-board CCTV, which itself mixes low-resolution monochrome and colour streams. To address this, we are developing a calibration-free, tracklet-level framework that links passengers across two to three uncalibrated cameras (front boarding, mid-aisle movement, rear-door exits). Preliminary findings from multi-hour RGB+IR bus footage recorded in Nairobi and Kigali indicate that a baseline YOLOv12–BoTSORT–OSNet pipeline reliably achieves Re-Identification (Re-ID) under low crowd conditions. However, performance rapidly declines as crowding increases, highlighting the challenges of this task. We are pursuing optimization plans that combine motion-aware gating, dual-branch embeddings resilient to head-only views, and assignment strategies robust to mixed-modality CCTV.