Enhancing Long-Duration Multi-Person Tracking in Hospitality Settings Through Random-Skip Sub-Track Correction

Long-duration customer monitoring with overhead cameras is essential for service analytics in restaurants and cafés, yet state-of-the-art multi-object trackers still suffer from track fragmentation whenever individuals are briefly missed or occluded during multi-hour recordings. We introduce Random-Skip Sub-Track Correction (RSSC) – a tracker-agnostic, post-hoc module that retrospectively joins fragmented tracklets without altering the real-time pipeline. RSSC generates R auxiliary frame streams by randomly skipping up to S consecutive frames, re-runs the baseline tracker on each, and validates candidate linkages wherever any sub-track bridges a temporal gap in the original trajectories. Because it operates solely on sparse frame indices and trajectory meta-data, the extra cost is negligible on low-power edge devices. Field experiments on five one-hour videos captured in a live doughnut shop show that RSSC reduces mean-absolute error in stay-time estimation by up to 23.5% and root-mean-square error by up to 18.2% across four representative trackers (SORT, OC-SORT, DeepSORT, Norfair) while adding virtually no runtime overhead. Although systematic detector failures at extreme viewing angles remain challenging, the results confirm that RSSC is an effective, lightweight remedy for sporadic fragmentation and a practical step toward robust, long-term identity maintenance in resource-constrained hospitality environments.