MiBee Eye makes a spare Linux board announce itself to an ONVIF recorder as an ordinary camera

The design decision worth copying is the frame hub. Capture and H.264 encode run once, and RTSP on port 8554, the local recorder and the web UI on 8088 all subscribe to the same encoded output ^[5]. Detection is another subscriber on that hub, and the project says a slow model, or a model that will not load, leaves streaming unaffected ^[6]. The two builds reach the model by different routes. The Rust build reads the shared pre-encode YUV buffer, so inference never touches capture or encoding; the Go build subscribes to the encoded hub and decodes keyframes to feed the model ^[8].

The integration claim is that against an ONVIF NVR the board looks like an ordinary camera ^[4]. Three SOAP services carry it, Device, Media and Imaging, alongside WS-Discovery multicast announcements that spare an operator typing an address ^[4]. For that to hold on a given network, the recorder has to receive those announcements. The project also says NVRs cannot tell the Go and Rust builds apart, and the stated basis is the same protocols, the same web API and the same SPEC v1 JSON envelope in both codebases ^[16]. That transfers if your recorder exercises nothing outside that surface. The writeup leaves out the recorders it was tested against and any frame rate or CPU figures for the software encode path ^[5].

Detection has three preconditions before it returns a single box: a build made with the compile-time flag, libonnxruntime.so present on the board, and the model file ^[21]. The release tarball includes neither the library nor the model, and the project calls detection "a setup step, not a checkbox" ^[12]. The model is NanoDet-Plus-m 320 in ONNX form under Apache-2.0, with a small registry that hot-swaps between 320 and 416 variants and a YOLOX-nano decoder for a different family ^[9].

Failure behaviour is stated plainly. With the runtime library or the model missing, the capability reports ai:false, the API says so, and it never returns fabricated boxes ^[11]. A client that reads the capability field can distinguish a broken detector from an empty scene; a client that reads only the box array sees the same empty list either way.

Runtime model upload exists and is off by default behind an explicit config gate, because a model file is untrusted input to an inference engine ^[10]. A closed default is the right one here, and the gate keeps an upload endpoint from being an arbitrary-file path into the inference runtime.

For installations that integrate with a Chinese standards-based video platform, the optional GB/T 28181 path does SIP registration over UDP or TCP with digest auth, catalog and device-info queries, PS streams for live, playback and download, and SIP INFO playback control ^[14]. That playback reads the same local recordings the device already writes: H.264 segments on disk with an index.jsonl index, a retention setting in days and a storage cap ^[13]. GB 35114 A-level, with SM2 certificate registration and keyed-SM3 integrity, is a compile-time opt-in ^[15].