Queue Length Estimation from CCTV using CNN

Canteens, banks, ticket counters and clinics all suffer the same problem: queues build silently until customers are already angry, and staff only react when someone complains. The cameras are already on the ceiling — what is missing is software that watches them. Counting people in a queue is harder than it looks: perspective shrinks distant people, crowds occlude each other, and a naive detector misses half the line. This project uses the density-map approach from crowd-counting research — a CSRNet-style CNN regresses a per-pixel density map whose integral is the count, which handles occlusion far better than bounding-box detectors. Counts inside each configured ROI feed a wait-time estimator (Little's-law style, calibrated per counter), and the dashboard shows live counts, wait times, hourly trends and threshold alerts.

1. CCTV frames (or recorded footage) are ingested per camera and the operator draws ROI polygons over each service line once during setup.
2. Each frame's ROI crop is passed through the density-map CNN, which outputs a per-pixel density map; integrating it gives the people count.
3. Person bounding boxes from a lightweight detector are overlaid for visualization, while the density integral remains the authoritative count.
4. A per-counter service rate (people served per minute, calibrated from historical data) converts the count into an estimated wait time.
5. The dashboard polls the pipeline and renders live counts, wait times, alert banners and the hourly trend chart.
6. When a count crosses its threshold, an alert is raised and logged with timestamp, camera and count.
7. The model is evaluated on ShanghaiTech crowd-counting splits and on annotated site footage for count MAE and wait-time MAE.

Tech stack:

• Python 3, PyTorch (CSRNet density CNN)
• OpenCV (frame capture, ROI, overlays)
• ShanghaiTech crowd dataset (benchmark)
• NumPy, Matplotlib
• HTML/CSS/JavaScript dashboard (canvas)
• SQLite (alert log)
• Git

Dataset & model details

• Dataset: ShanghaiTech crowd-counting Parts A/B (1,198 images, ≈330k annotated heads) for density-map training; fine-tuned and evaluated on site-collected queue footage with ROI annotations gathered at build time.
• Task: Crowd density estimation; input = ROI frame crop, output = per-pixel density map whose integral is the people count; wait time regressed from count × calibrated service rate.
• Model: CSRNet-style CNN — VGG-16 frontend with dilated-convolution backend, trained with Euclidean loss on Gaussian density ground truth.
• Metrics: Count MAE ≤ 1.8 persons and wait-time MAE ≤ 1.5 min (design targets) on annotated site footage. No accuracy is claimed as measured until the build is completed for the order.

• [Density-map people counting] CSRNet-style CNN regresses crowd density maps — robust to occlusion and perspective, unlike box detectors in dense queues.
• [ROI queue definition] Click-to-draw polygons per camera define each service line; counts are computed only inside the ROI.
• [Wait-time estimation] Calibrated per-counter service rates convert queue counts into estimated wait times shown on the dashboard.
• [Multi-camera dashboard] Live grid of camera feeds with overlaid counts, ROI polygons, wait times and threshold alert banners.
• [Threshold alerts] Configurable per-counter limits (e.g. 10 people) trigger visual alerts suggesting staff action like opening another counter.
• [Hourly and weekly analytics] Queue-length time series and day×hour heatmaps reveal peak patterns for staffing decisions.
• [Alert log] Timestamped history of every threshold crossing for the operations report.

• Density-map CNN training and inference pipeline
• ROI drawing and per-counter calibration tool
• Multi-camera monitoring dashboard with alerts
• Hourly/weekly analytics and alert log
• Evaluation on ShanghaiTech splits + site footage
• Project report PDF (background, crowd-counting theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (density maps, CSRNet, perspective, Little's law)

• 1.8-person MAE is a design target, stated honestly — the report documents the actual achieved figure after the build.
• Needs a reasonably elevated camera angle; top-down or steep views count best, while head-on views suffer occlusion.
• Very dense crowds (50+ in one ROI) exceed the calibrated range and are reported with a low-confidence flag.
• The wait-time estimate assumes the calibrated service rate holds; a slow cashier or a complex transaction breaks the estimate.
• Privacy: the system counts people and stores no identities — face recognition is explicitly out of scope.