Smart Surveillance and Crowd Management System Using a Custom-Trained YOLO11n Person Detector

Noticing that a crowd is building up is hard for a human operator on a long shift, and it is also hard for offthe-shelf detectors when the footage is aerial or camera-mounted high, because people are only a few pixels tall.We built an offline tool to help with this.It pairs JATAYU v2, a person detector that we trained ourselves on our own dataset (YOLO11n architecture), with a deliberately simple analysis layer: a user-defined grid of zones, LOW, MODERATE and HIGH labels based on count thresholds that the operator sets, a heatmap of where people were detected most often, a log of threshold crossings, and a Streamlit dashboard that runs on an ordinary laptop CPU.On its validation split the detector reached precision 0.724, recall 0.583, mAP50 0.638 and mAP50-95 0.312.On an 8-second aerial clip (240 frames, 960 × 540) the full pipeline analysed 120 frames in 11.7 s and found 154.8 people per frame on average, with a maximum of 176, whereas COCO-pretrained YOLOv8n and YOLOv8s found 0.41 and 0.11.The clip has no groundtruth counts, so this comparison is qualitative only.Additional tests show that the count depends heavily on the confidence threshold and the inference size, hardly at all on frame skipping, and that the zone grid and the zone threshold have to be chosen together.All numbers are detections per frame rather than unique people, and the heatmap and density labels are indicators, not calibrated measurements.