Abstract

Accurate crowd analysis in high-density environ-ments is critical for public safety. Conventional bounding-box-based detection methods often fail under severe occlusion and overlap because their predictions depend on box quality and post-processing such as Non-Maximum Suppression (NMS). This paper presents a transparent crowd analytics pipeline built around a YOLOv12 backbone, point regression, density modeling, and spatial-entropy-based explainability. Instead of predicting a bounding box for each person, the proposed frame-work predicts a single representative point per individual, which preserves one-to-one mapping in dense scenes and reduces missed or suppressed detections. The detected points are converted into density maps, and spatial entropy is computed to quantify disorder and identify potentially unsafe regions. The uploaded source manuscript reports a Mean Absolute Error (MAE) of 10.8, a counting accuracy of 92%, and a real-time throughput of 36 FPS, while also providing interpretable outputs through point overlays, heatmaps, and highlighted risk zones. The overall result is a practical and transparent framework for intelligent surveillance and public safety monitoring.