Temporal-Scale-Aware Logical Track Reassociation for Lightweight Personnel Flow Monitoring under Sparse Inference

This preprint presents a temporal-scale-aware logical track reassociation framework for lightweight personnel flow monitoring under sparse inference. The proposed method addresses trajectory fragmentation and raw ID changes caused by reduced detection frequency by combining elapsed-time-aware motion extrapolation, bounding-box-height-normalized spatial matching, direction-consistency gating, and persistent logical identity management. The framework is training-free and requires no additional detector calls, Re-ID models, optical flow, or retraining. Experiments on 120 industrial personnel-flow videos recorded at 15 FPS with detection performed at 1 Hz achieved 90.00% exact video-level IN/OUT counting accuracy, compared with 52.50% for sparse ByteTrack, 54.17% for OC-SORT, and 55.00% for Hybrid-SORT. The proposed reassociation layer introduces only 7.02% additional CPU processing time per frame, demonstrating improved counting reliability with limited computational overhead.