Queue Length Estimation using YOLO

Queues at banks, canteens and ticket counters are usually measured by someone standing and counting — tedious, inconsistent and impossible to keep up all day. Without continuous data, staffing cannot be matched to real demand: counters sit idle or get overwhelmed. A camera-based estimator changes that by watching the counter continuously, reporting how many people are waiting right now and logging the count over time. The catch is that naive person counting also counts everyone walking past. This project solves the real problem with three layers: YOLOv8 for person detection, centroid tracking to keep identities stable across frames, and a configurable region of interest so that only people actually inside the queue zone contribute to the count. The detector is trained with CrowdHuman (Shao et al., 2018), the public benchmark of thousands of images with dense person annotations built specifically for crowded scenes, and the evaluation notebook measures precision and recall plus an ROI-sensitivity check so the region dependence is documented, not hidden.

Dataset & model:
Dataset name: CrowdHuman (Shao et al., "CrowdHuman: A Benchmark for Detecting Human in a Crowd", arXiv:1805.00123, 2018).
Source: Official academic release of the CrowdHuman benchmark.
Task: Person detection in video frames, composed with tracking and ROI counting.
Classes: Person (human) — dense person annotations built for crowded scenes, per the official dataset.
Model: YOLOv8 (Ultralytics) person detector trained with CrowdHuman annotations; centroid tracking with persistent IDs across frames.
Input: Video files or live camera streams of service counters.
Prediction: Person bounding boxes with confidence scores per frame, tracked across frames.
Output: Live queue-length count overlaid on video, plus a timestamped queue-length timeline log.
Evaluation metrics: Detection precision and recall on the validation split, plus an ROI-sensitivity check — computed by the Jupyter notebook during the buyer's build. No metrics claimed.

Working:

1. Video frames are captured from a file or camera stream with OpenCV.
2. Training phase: YOLOv8 is trained with CrowdHuman dense person annotations.
3. Inference phase: YOLOv8 detects persons in each frame, returning bounding boxes with confidence scores.
4. Box centroids are matched across consecutive frames to assign persistent tracking IDs.
5. Tracked centroids falling inside the configured queue ROI polygon are counted as queued; people outside the polygon never enter the number.
6. Evaluation phase: the notebook measures detection precision and recall on the validation split and runs an ROI-sensitivity check. The count, tracking IDs and timestamps are logged into a queue-length timeline, and the demo app overlays the live count with boxes and IDs on the video.

Specifications:
Model | YOLOv8 person detector (Ultralytics) + centroid tracking with persistent IDs
Dataset | CrowdHuman — dense person annotations (Shao et al. 2018)
Task | Person detection, tracking and ROI-restricted queue counting
Input | Video files or live camera streams of service counters
Counting rule | Only tracked centroids inside the configured ROI polygon are counted
Output | Live count overlay with boxes and IDs + timestamped queue-length timeline
Evaluation | Precision, recall on validation split + ROI-sensitivity check — computed during the buyer's build
Demo app | Processes uploaded videos or live streams
Scope | Single-counter feeds with one configured ROI

[YOLOv8 Person Detector] (implemented) — YOLOv8 trained with CrowdHuman dense annotations for person detection on video frames.
[Centroid Tracker] (implemented) — Box centroids are matched across consecutive frames to keep persistent tracking IDs stable.
[Configurable Queue ROI] (implemented) — A queue ROI polygon restricts counting to the queue zone; an ROI editor adapts it to any camera view.
[Live Count Overlay] (implemented) — The demo app overlays the live queue-length count with per-person bounding boxes and IDs on the video.
[Queue-Length Timeline Log] (implemented) — Counts are logged with timestamps for staffing and service-time analysis.
[Evaluation Notebook] (implemented) — Computes precision and recall on the validation split plus an ROI-sensitivity check.
[Video and Live-Stream Demo] (implemented) — The demo app processes uploaded video files or a live camera stream.
[Wait-Time Estimation] (future-scope) — Estimating wait times from tracked dwell times is not included; it is future work, not a current feature.

Complete source code (detection, tracking, ROI counting, logging, demo app)
Jupyter training and evaluation notebook (precision, recall, ROI-sensitivity plots)
Project report PDF (background, CrowdHuman, methodology, evaluation, error analysis)
PPT presentation for final review
Viva Q&A preparation document (YOLO, CrowdHuman, tracking, ROI counting, evaluation)
Setup guide (environment, dataset download, camera configuration, running on your own footage)

CrowdHuman is the evaluated dataset: viewpoints, lighting and crowd densities far from it classify worse, and the report documents this coverage gap openly.
Heavy occlusion in dense crowds breaks centroid tracks and undercounts; the system is an estimator, not an exact census, and the report shows where it underperforms.
Scope covers single-counter video feeds with one configured ROI; the ROI must be defined per camera view.
This is an educational prototype for queue analytics, not a certified occupancy or safety system — counts are advisory and must not drive capacity or safety decisions.
Low-resolution or poorly positioned cameras degrade both detection and tracking stability.