Multi-Granularity Attention Fusion for Video-Based Person Re-Identification in Dynamic Urban Scenes. JAAIR 2026, 1 (1).