AI Analyzer
In Settings → AI Analyzer, choose the input to analyze, the AI model, and whether to use GPU acceleration. Media Type is either Video (Camera) (the default) or Audio (Mic/Speaker). For video, select a registered Camera; for audio, set Audio Source to Microphone (Input) (the default) or Speaker (Output). For video analysis, Region of Interest (ROI) restricts inference to a region defined on the camera; leave it empty to use the full frame.
The AI Model list depends on the media type. Video offers Object Detection (the default), Face Recognition, Pose Estimation, OCR (Text Recognition), License Plate Recognition, and more; audio offers Sound Classification (the default) and Edge Impulse Classification. See Part 4 for a description of each model. The selected model adds the following fields.
| AI model | Field | Description |
|---|---|---|
| Teachable Machine Classification | Model Path (.tflite) | Upload the TensorFlow Lite model exported from Teachable Machine. Only 224×224 RGB models are supported; choose the Floating Point or Quantized type when exporting. |
| Label Path | The labels.txt file generated with the model. It lists the class names in order, one per line. | |
| Edge Impulse Classification, Object Detection, and Visual Anomaly (video); Edge Impulse Classification (audio) | Model Path (.eim) | Upload the .eim model file exported from Edge Impulse Studio. |
| Face Recognition | DB Name | The database that stores and looks up recognized faces (default: default). A name that does not exist yet is created automatically, and analyzers that use the same name share face data. |
| Anti-Spoofing | Blocks recognition triggered by photos or screens (off by default). Recognition becomes slightly slower but resists spoofing attempts. | |
| OCR (Text Recognition) | OCR Language | English (the default), Korean, Japanese, Chinese (Simplified), German, French, Spanish, Portuguese, Hindi, Arabic, or Russian. |
| License Plate Recognition | Region | Latin (EU/US/SouthAmerica) (the default) is for Latin-script plates from Europe, the US, and South America; Korea is for Korean plates with Hangul and digits. |
Mode balances analysis speed against accuracy. Fast is recommended for lower-powered devices such as a Raspberry Pi, Accurate for high-performance devices such as PCs and Jetson, and Balanced in between. The available values depend on the AI model: Object Detection, Face Detection, Face Recognition, Fire Detection, and Line Crossing Counter offer all three; Pose Estimation, Hand Tracking, License Plate Recognition, and Custom Anomaly Detection offer only Fast and Accurate; the remaining models use a single model, so the Mode field is not shown.
Detection FPS caps how many times per second analysis runs. The default is 5; Pose Estimation, Hand Tracking, Action Recognition, Hand Gesture Recognition, and Line Crossing Counter default to 10, and the computation-heavy Custom Anomaly Detection defaults to 1. A value of 0 means no limit. If the controller is overloaded, lowering this value is more effective than changing the mode.
GPU Acceleration appears for Object Detection, Face Detection, Face Recognition, Pose Estimation, Hand Tracking, Line Crossing Counter, and OCR, and is off by default. When on, the model runs on the GPU (Vulkan), which is faster than the CPU on hardware that supports it; on integrated or older GPUs it may actually be slower, so leaving it off can be the better choice. Devices without GPU support, such as a Raspberry Pi, always run on the CPU regardless of this setting.