CA-CRS+: A Zone-Aware, Resource-Informed Crowd Risk Scoring Framework with Cause-Specific Gate Intervention Recommendation for Stampede Prevention

CA-CRS+ is a zone-aware, resource-informed crowd-risk scoring framework operating on standard RGB video without thermal cameras or cloud dependency. It addresses four critical limitations of existing crowd-safety systems: generic alerts without cause, absence of spatial resolution, lack of gate-specific actionable recommendations, and lack of dynamic emergency resource validation. Key Contributions & Findings: • Benchmark-Validated Crowd Counting: Replaces whole-body detection with density-map regression (DM-Count VGG-19; validated at MAE 48.7 on ShanghaiTech Part A, 14.4 on Part B, and MAE 0.55 against multi-object-tracking ground truth on PETS2009 S2.L1). • Circular-Variance Conflict Index: Normalized directional conflict metric via magnitude-weighted circular variance of optical flow that does not degenerate under balanced counter-flow. • Non-Linear Crush Scoring: Solves the zero-velocity gridlock paradox where dense stationary crowds are scored as safe by linear models. On dense imagery, CA-CRS+ flags 55% of scenes as DANGER where a matched linear baseline flags 0%. • Cause-to-Gate Mapping with Ripple-Effect Downgrade: Emits named gate commands (Density → OPEN, Speed → CLOSE, Conflict → REDIRECT, MIXED → HOLD) with a safety check preventing crowd redirection from overwhelming downstream zones. • Dynamic Resource-Aware Alerting: Sizes marshal demand logarithmically with crowd population (D_mar = ceil(alpha * log10(1 + N))). • Multi-Camera Fusion: Fuses multiple real CCTV camera streams into an urgency-weighted (2:1) Global Risk Score (GRS). • Edge Performance: Runs at 4.0 frames/s on a single commodity NVIDIA Tesla T4 GPU, supporting real-time 1–2 Hz risk-update cadences. Provenance note: This research paper was initially prepared for IEEE Region 10 Conference (TENCON 2026), thoroughly upgraded after peer review to include ground-truth quantitative benchmarking, sensitivity analyses, and multi-camera fusion, and subsequently withdrawn from TENCON by the authors. It is released here as a permanent, open-access research preprint and reproducible artifact archive.