Supported devices
The devices below have been tested with the Embedded Stream Processor. The list is not exhaustive: other cameras from the same vendor built on the same chipset are usually compatible, and AI boxes are supported by accelerator rather than by model. If you are planning a purchase, confirm the exact model with Innovatrics first, because on-device support depends on the chip, the vendor SDK version and the camera firmware.
Smart cameras
Smart cameras must use an Ambarella System on a Chip (SoC) with Ambarella SDK 3.0. The Embedded Stream Processor is installed as a vendor application package on the camera.
| Vendor and series | SoC | Tested models | Notes |
|---|---|---|---|
| Hanwha P series | Ambarella CV2 | PND-A9081RV, PND-A6081RV, PNV-A6081R, PND-A6081RF | Installed through Hanwha Open Platform |
| AXIS M and P series | Ambarella CV25 | M4216-LV, M1055-L, M2035-LE, P1245 Mk II, P1265 Mk II, P1275 Mk II | AXIS OS 11.11.73 or later; AXIS OS 12 is not supported |
| Dynacolor U2-F and U2-B series | Ambarella CV22 | U2-6, U2-B | |
| Lilin Z7 series | Ambarella CV22 | Z7R8182X2-PAI | Ambarella 3.0.x firmware; installed as a Lilin plugin |
AI boxes and edge computers
AI boxes run the Embedded Stream Processor as a Linux application and take video from an attached USB camera, an RTSP camera or a file through a GStreamer pipeline. One box can process several streams by running several settings files.
| Accelerator | Tested devices | Notes |
|---|---|---|
| Hailo-8 | Axiomtek RSC101 | Any Hailo-8 device with HailoRT 4.17.0 drivers; Ubuntu 22.04 |
| NVIDIA Jetson (Xavier NX, AGX Xavier, Orin) | Axiomtek NVIDIA Jetson systems, Advantech MIC-710AIX (Xavier NX) | CUDA or TensorRT acceleration through ONNX Runtime |
| NXP i.MX 8M Plus | NXP evaluation and partner boards | TensorFlow Lite with the VX delegate on the NPU; USB camera input |
Supported chipsets and on-device features
The Embedded Stream Processor ships with a set of solver binaries per accelerator. Not every stage of the pipeline is available on every chip; where a stage is missing, leave it to the server (for example, run liveness as CPU on server in the Station edge-stream settings).
| Accelerator | Face detection | Template extraction | Passive liveness on device |
|---|---|---|---|
| Ambarella CVFlow (CV2, CV22, CV25, CV28) | accurate_mask | balanced, accurate_mask | Distant, Nearby |
| Hailo-8 | accurate_mask | accurate_mask | not available — use server-side liveness |
| NVIDIA Jetson (ONNX Runtime) | accurate_mask | fast, balanced, accurate_mask | Distant, Nearby |
| NXP i.MX 8M Plus (TensorFlow Lite) | accurate_mask | balanced, accurate_mask | not available — use server-side liveness |
Face templates are version-bound: a device must extract templates with the same algorithm as the Face Matcher server it reports to, otherwise the server re-extracts them from the crops. See Face processing pipeline for the server-side algorithms. The full accelerator matrix of the underlying SDK is on the Embedded Toolkit hardware acceleration page.