Project at a Glance
- Domain: Computer Vision, Public Safety, Smart Cities, Event Ops
- What we built: A privacy-first, real-time crowd counting system that turns live CCTV/IP camera feeds into accurate counts and density heatmaps to help operators spot congestion early and act fast.
- Runs on: Edge devices (Jetson-class) or cloud VMs; ONVIF/RTSP compatible.
- Highlights: Sub-second alerts, adaptive to indoor/outdoor scenes, dashboard with historical analytics, no face IDs stored.
The Challenge
Large venues, transit hubs, universities, and street festivals in dense cities experience sudden surges that are hard to catch by eye. Traditional people-counters struggle in:
- High density (severe occlusion)
- Mixed scenes (camera angles, lighting, weather)
- Real-time needs (latency budgets well under a second)
Our Approach
We designed a two-path pipeline that adapts per scene:
- Density-Estimation Path
Generates a pixel-wise density map to stay robust in high-density areas with heavy occlusion (inspired by recent “fuss-free” efficient networks and real-time frameworks). - Detection Path (Heads/Persons)
Activates in low–medium density regions for precise counting and region-of-interest analytics; a routing strategy picks the right path per tile when scene density changes.
Key Features
- Live Count & Heatmaps: Per-camera totals, per-zone density, and configurable thresholds for “warning” and “critical”.
- Operator Alerts: Instant notifications when zones exceed safe occupancy for N seconds.
- Smart Zoning: Drag-and-drop polygons for gates, corridors, platforms, prayer halls, etc.
- Scene Adaptation: Auto-calibration for new cameras; learns background and scale.
- Historical Analytics: Hourly/day-of-week patterns, peak analysis, exportable CSVs.
- Edge-First & Privacy-First: Runs locally if desired; streams can stay on-prem; no face recognition or identity tracking by default.
- Open Protocols: RTSP/ONVIF ingest; REST/WebSocket APIs for control room software.
Results & Impact
- Faster decisions: Operators see congestion before it’s obvious on raw video.
- Better safety & flow: Early rerouting reduces pinch points at entries, stairs, and ticket gates.
- Operational planning: Historical trends help allocate staff and signage for busy periods.
(Exact throughput and accuracy depend on camera resolution, crowd density, and hardware. On modern GPUs or embedded accelerators, we target real-time processing with practical latency budgets for control rooms.)
Data & Evaluation
- Training/Validation: Curated public datasets (e.g., ShanghaiTech, UCF-QNRF) plus synthetic augmentation for lighting and perspective; optional client-scene fine-tuning.
- Benchmarks: Report mean absolute error (MAE), mean squared error (MSE), and alert-time latency under representative densities—consistent with modern literature on hybrid detection/regression and multi-scene real-time methods.
Privacy & Ethics
- No PII by default: We do not store identities; outputs are counts and density maps.
- On-prem options: Full processing on edge devices; optional redaction at source.
- Configurable retention: Video frames never leave your premises unless you choose.
- Transparent operations: Clear signage and opt-in policies where required.
Where It Fits
- Event venues & festivals for live capacity management
- Transit hubs (stations, platforms, concourses)
- Universities, hospitals, malls, stadiums
- Smart-city control rooms and lawful crowd management
Deliverables
- Deployed service (edge or cloud)
- Web dashboard with live overlays & analytics
- API/SDK for integration with VMS/PSIM
- Admin guide + MLOps handover
- Optional fine-tuning on your cameras
