AI Models

AI models

The camera runs AI on its built-in NPU. The AI itself comes as model files (.rknn). PicoCam includes five ready-made models so you can try everything straight away, and you can add your own.

The ready-made models

They are in picocam/models/ and installed with the full package (not with --lite).

Model	What it does	Used by template
Objects (COCO 80, YOLOv5n, 320)	Finds 80 kinds of objects: people, cars, buses, trucks, bikes, animals ...	People & vehicles, ANPR, People counting
Number plate detector (YOLO11n, 320)	Finds the plate on a vehicle	ANPR
Indian plate reader (PlateNet OCR)	Reads single and two-row Indian plates	ANPR
Face detector (RetinaFace, 320)	Finds faces and five landmarks	Face capture
Face embedding (MobileFaceNet)	Turns a face into numbers so people can be recognised	Face capture

To use them: open AI & Analytics, click a template (for example People & vehicle detection) and Apply. That is all.

Important: models are built for one chip. Rockchip compiles a model for a specific chip. The ready-made models are built for RV1103. On an RV1106 board they may refuse to load; the page then tells you which chip your board is, and you convert the model again for it (below). System > Overview shows your chip. This is the single most common reason for "model not valid".

Getting the models onto the camera

Situation	Do this
Normal install	Already there
You installed the lite package	picocam/tools/deploy.sh --models (Windows: deploy.ps1 -Models). It adds them to /userdata/picocam/models
Only one model	AI & Analytics > Models > Upload, choose the .rknn file (and a class list), or a model package
Offline	Download picocam-models.tar.gz from the release page and use deploy.sh -p picocam-models.tar.gz --models

Models take space: about 12 MB for all five. On SPI NAND boards with a small user partition the script checks first and tells you if it will not fit. A card is not needed for models, but an SD-booted board has plenty of room.

Use your own model

1. Train with Ultralytics YOLOv5, YOLOv8 or YOLO11, PyTorch or Paddle, and export to ONNX. For YOLOv8/11 use models/export_yolov8_rknn.py: it produces the layout the NPU likes.

2. Convert to RKNN with the project's tools (the Docker image in docker/ contains RKNN-Toolkit2):

   python3 models/convert.py model.onnx calib.txt model.rknn --size 320 --preset yolo --platform rv1103   # or rv1106

   calib.txt lists 20-100 pictures from your own camera, one path per line. They are what makes the 8-bit conversion accurate. Do not publish them: real footage shows people and number plates.

3. Package (optional but tidy): python3 models/package.py model.rknn --decoder yolox --labels labels.txt makes one .tar.gz that carries the settings the camera needs.

4. Upload under AI & Analytics > Models. The task (detector, classifier, text reader, face embedding) and the tensor shapes are recognised automatically. Test on live picture shows the result at once, and Benchmark measures speed on your NPU.

5. Use it in Pipeline as the detector, or as a stage that runs on each found object (for example vehicle, then plate, then text).

Rules for models: 8-bit input, normalisation built in (--preset), no more than 16 MB on 64 MB boards and 40 MB on larger ones.

Supported outputs: YOLOv5 (also custom anchors), YOLOv8/YOLO11 (Rockchip layout), YOLOX, YOLOv10, RetinaFace, classifiers, CTC text readers, embeddings. Anything else can be added with a small decoder plugin (picocam/sdk/), administrators only.

Licences

The ready-made models are conveniences with their own licences: some are derived from Ultralytics YOLO (AGPL-3.0) and public face-model architectures. They are not covered by the project's MIT licence. If you build a product, check NOTICE or use the lite package with models you have the rights to.

How well does it work?

An honest rule of thumb on these small chips: people and vehicles at a normal gate distance work well; plates need a plate that is roughly 100 pixels wide or more in the analysed frame (raise the sub-stream resolution or use the detection area to zoom into the road) and enough shutter speed at night (Image > Scene > Road / ANPR does this). Faces work best in even light at close range.