Edge-AI People Counter on Jetson Nano

Malls, libraries, labs and campuses all need footfall numbers — for staffing, safety limits and space planning — but manual clickers don't scale and cloud video analytics streams identifiable footage off-site with bandwidth and privacy costs. Edge AI resolves the tension: detect and count on the device, and transmit only anonymous numbers. The Jetson Nano is the classic student edge-AI board — a real CUDA-capable GPU in a classroom budget — and person detection is the canonical workload. This project builds the complete counter: a YOLOv8n model exported to TensorRT FP16 running on the Nano, centroid-based tracking across a virtual counting line for IN/OUT direction, a dashboard with live occupancy, hourly charts and overcrowding alerts, and a buyer-run counting test procedure so the accuracy in the report is measured on the student's own doorway, not claimed.

1. The camera captures frames on the Jetson Nano; each frame is preprocessed to the model's input size.
2. YOLOv8n, accelerated with TensorRT FP16, detects persons and outputs bounding boxes with confidence scores each frame.
3. A centroid tracker assigns persistent track IDs across frames, following each person through the scene.
4. When a track's centroid crosses the virtual counting line, the crossing direction decides IN vs OUT and the counters update.
5. Counts and timestamps are published over MQTT/Wi-Fi to the dashboard; video frames stay on the device.
6. The dashboard renders live occupancy, the hourly footfall chart and the event feed, and checks the overcrowding limit.
7. The student measures accuracy with the test sheet: timed intervals of manual clicker counts compared against the system count, recorded per lighting and crowd condition.

Tech stack:

• NVIDIA Jetson Nano 4GB with camera (CSI or USB)
• YOLOv8n person detector exported to TensorRT FP16
• Python with OpenCV (capture, preprocessing, centroid tracking)
• MQTT telemetry plus web dashboard (occupancy, charts, alerts)
• COCO pre-trained weights (person class) — no training required

Dataset & Model Details

• Dataset: COCO (Common Objects in Context), official release — the detector uses the pre-trained person class only; no custom training dataset is required.
• Source: COCO official dataset; weights from the model vendor's COCO-trained release.
• Task: Object detection (person class) + multi-object tracking + line-crossing counting.
• Model: YOLOv8n (nano variant) exported to TensorRT FP16 for Jetson Nano inference.
• Input: Camera frames resized to the model input resolution. Prediction: person bounding boxes with confidence scores. Output: track IDs from the centroid tracker and IN/OUT crossing events.
• Evaluation: published COCO detection metrics belong to the vendor's release and are cited as such; the project's own counting accuracy is measured by the student with the included test-sheet procedure (manual clicker vs system count over timed intervals, across lighting and crowd conditions). Final performance depends on camera placement, mounting height, lighting, crowd density and occlusion — the report documents the measured numbers with their test conditions, never a pre-claimed figure. Design target: approximately 85–90% counting accuracy in controlled doorway conditions.

• [On-device person detection] YOLOv8n (pre-trained on the COCO person class) exported to TensorRT FP16, running entirely on the Jetson Nano — no cloud inference.
• [Virtual line counting] Each tracked person crossing the configurable counting line increments IN or OUT by direction, using centroid tracking with track IDs.
• [Live occupancy dashboard] Current occupancy, today's IN/OUT totals, an hourly footfall bar chart and a scrolling event feed.
• [Overcrowding alerts] A configurable occupancy limit raises a dashboard and optional buzzer/Telegram alert when the space exceeds it.
• [Privacy by design] Only anonymous counts and timestamps leave the device; the video frames never leave the Nano and no identities are stored.
• [Buyer-run accuracy procedure] A counting test sheet (manual clicker vs system count over timed intervals) so the report documents measured accuracy from the student's own setup.
• [Configurable zones] Counting line position, direction and dwell filters are adjustable for doorways, corridors and entrances of different widths.

• Complete Jetson Nano Python code (capture, TensorRT inference, centroid tracking, line-crossing logic, MQTT publishing)
• TensorRT export and optimization guide for YOLOv8n FP16
• Camera mounting and counting-line configuration guide
• Counting test-sheet template (manual-vs-system accuracy procedure)
• Web dashboard app (live occupancy, hourly chart, event feed, overcrowding alerts)
• Project report PDF (background, model choice, tracking design, evaluation method)
• PPT presentation for final review
• Viva Q&A preparation document (YOLO, TensorRT, tracking, edge AI, COCO, evaluation)
• Setup guide (JetPack setup, dependencies, first detection run)