Solution Description
Line Crossing realtime page: a target crosses the counting line → IN/OUT counts update instantly → the crossing event is logged. Everything runs on the camera itself.
Background
Demand for line-crossing counting spans sites of every kind: stores track entries and exits, corridors reveal flow direction, and facility gates need peak-hour analysis. Traditional approaches are all costly in their own way: gates and infrared beams require hardware plus construction work and still miss targets on wide passages; cloud video analytics charges per-camera subscriptions while footage leaves the local network; a self-hosted platform means servers, licenses and weeks of deployment.
Most sites don't need platform-grade precision. They need a solution that is fast to deploy, low in cost, and working the same day.
Solution Overview
The idea: one affordable camera plus one counting line turns "crossings" into structured data.
Deploy a CamThink NeoEyes NE301 edge AI camera (STM32N6, 0.6 TOPS, modular design) at the site and install the line-crossing app. All configuration happens in the web page: first deploy the detection model for your target and set the target class, then two clicks draw the counting line on the live view, drag the endpoints to fine-tune, and swap IN/OUT direction with one tap. The app runs object detection, multi-object tracking and line-crossing judgement on the camera, outputting bidirectional counts in real time; window reports are generated at the configured statistics period and pushed straight to your business system over MQTT / Webhook — no extra platform software.
What gets counted is decided by the on-device detection model: the camera ships with a person-detection model for footfall counting out of the box; for custom targets, train a dedicated detection model for your scenario and deploy it to the camera — training, quantization and deployment are fully open source, with zero changes to the solution code.
The whole loop closes on the camera: detection, counting and report generation in one device. Footage never leaves the local network — your business system only receives lightweight counting results.
Solution Components (BOM)
| # | Component | Model / Spec | Qty | Role |
|---|---|---|---|---|
| 1 | Edge AI camera | NeoEyes NE301 (comm module per site: Wi-Fi 6 / Cat-1 / PoE) | 1 per site | On-device detection, tracking and line-crossing counting |
| 2 | line-crossing app | .neapp app package, delivered with the solution | 1 set | Line configuration, realtime statistics, window report generation |
| 3 | Power & network accessories | Type-C power / PoE; Wi-Fi AP / PoE switch / SIM card | As needed | Camera deployment kit |
| 4 | Business system receiver | MQTT broker (incl. MQTTS) or Webhook-capable HTTP service | 1 set | Receives window reports into business flows |
Low-cost starting point: 1× NE301 (Type-C power + Wi-Fi) + the app package + any existing MQTT/Webhook receiver completes a site — no gate construction, no server.
Workflow
- Deploy the model and set the target — upload the specified detection model for what you want to count (a Model Zoo example or your own trained model) and select the matching target class; the bundled person model counts footfall out of the box
- Define the counting line — two clicks place the line with its outside direction, drag endpoints to fine-tune, save to apply
- On-device detection & tracking — the camera detects and tracks targets continuously, producing stable trajectories
- Crossing judgement & counting — a trajectory crossing the line counts as IN or OUT by direction; totals survive power loss
- Window statistics — IN/OUT per period is summarized (configurable 1–1440 min) into window reports
- Report delivery — reports are pushed via MQTT / Webhook; during outages they buffer within capacity and are re-sent best-effort after recovery
Key Features
Flexible deployment
- Draw to configure: two clicks define the counting criteria — no calibration, no training, edits take effect immediately
- Simple power and networking: direct Type-C power supply, or PoE single-cable power plus Ethernet; comm module (Wi-Fi 6 / Cat-1 / PoE) selected per site
- Self-contained on the camera: the app runs on the camera itself — no server, no platform software; relocating a site is uninstall-and-reinstall
Customizable detection models
- Self-trained model deployment: train a detection model for your scenario and deploy it to the camera to count your own targets; training and deployment are fully open source
- Model Zoo examples: the official model library offers tuned vision models (PT / ONNX / TFLite) ready to download and use
- Progressively improvable accuracy: the bundled person model provides a usable baseline; keep training on field samples, optimizing and redeploying to raise accuracy iteration by iteration
- On-device inference: detection and counting run locally on the 0.6 TOPS NPU — no cloud dependency
Low cost
- Starts at one camera: no gates, no construction; deployed the same day
- No cloud subscription: counting results go straight into your existing systems — no per-camera fees, no video egress bandwidth costs
- Data locality: footage and counts remain on the local network, fully under your control
Outcome
- Per-site cost is a single camera, producing data the day it is deployed — replaces manual tallies
- Switching scenarios requires only another detection model and a redrawn line; the statistics framework never changes
- In/out traffic lands in your system by window, giving scheduling, merchandising and efficiency analysis continuous data
- Accuracy is affected by mounting angle, lighting and target density; intended for traffic-trend analysis. Improve it iteratively by training on field samples and redeploying — always validate on site