CameraStack now has a reference port for the Rockchip RV1126B, proven on the Luckfox Aura with a Waveshare IMX415 Camera (A). It joins the RV1106, the Sophgo SG2002, the Axera AX630C and the other reference ports, and it turns the board into a complete multi-profile network camera: 4K hardware video, microphone audio, JPEG snapshots, Imaging, OSD, motion and object analytics, RTSP and the shared ONVIF service set.
The most useful result is how little of CameraStack had to change.
A new producer, the same camera server
The portable Go core supplies ONVIF, RTSP, TLS and device management on the RV1126B exactly as it does everywhere else. A compact C producer handles the hardware side, driving Rockchip RKMPI, AIQ, VPSS, RGA, IVS and the RKNPU2 neural engine. Encoded video, audio samples, events and object records cross the same producer protocol into the same core used on the RV1106, SG2002, AX630C and Hi3519V101.
Even the native layer is mostly shared. The RV1126B producer reuses CameraStack’s common C libraries for wire framing, configuration transactions, pause state, polygon mapping, event formatting, OSD state and object-model decoding. Its own code is concentrated where it should be: the RV1126B media graph, the ISP controls, the RKNPU2 tensor boundary, the ALSA audio route and the board’s lifecycle rules.
That was the portability test. A more capable Rockchip platform arrived, and the generic product core behaved as mature infrastructure, with only bounded board-specific work on top.
4K video with independent profiles
The Sony IMX415 supplies a full 3840x2160 image at 30 fps. CameraStack keeps the sensor, ISP and VPSS pipeline resident and exposes two independent hardware encoders. The main profile runs at 1920x1080 or 3840x2160 at up to 30 fps; the sub profile runs at 640x360 at up to 10 fps. Either can use H.264 High or H.265 Main.
Frame rate, bitrate and GOP change on the running encoder. A codec change replaces only the selected VENC channel, and a main-resolution change moves only that branch between 1080p and 4K. The sibling stream, audio, snapshots and analysis path keep running, and every fresh reader begins with a complete keyframe.
Encoding is demand-driven. When a profile has no viewers, CameraStack pauses that encoder after a grace period and leaves the video pipeline resident, so snapshots, motion and object metadata stay available while both live video tracks are idle.
The microphone is part of the profile
The Aura’s onboard microphone is a native CameraStack audio source. The default main profile carries H.264 video, 8 kHz mono G.711 mu-law audio and ONVIF metadata; the sub profile carries video and metadata only.
The producer captures the microphone through ALSA, downsamples it to 8 kHz and emits one 160-byte packet every 20 ms, with RTP timestamps advancing by sample count so scheduling jitter cannot accumulate as clock drift. Acceptance testing covered packet cadence and combined audio and video decoding.
Object detection with selectable models
The 3 TOPS RKNPU2 runs five detector families at 1 or 5 fps. Three are portable CameraStack models: PicoDet-S, YOLOX-Nano and PP-YOLOE-S. Two are vendor examples supplied by the operator: Luckfox’s YOLOv10n and Rockchip’s YOLOv5s.
A standard ONVIF Analytics request selects the installed model, confidence threshold and detection rate while the camera stays live. RGA prepares the model input, RKNPU2 runs inference, and the common CameraStack tracker turns the result into stable ObjectIds and ONVIF scene metadata. All five models passed fixed-image tests and detected a person in live checks through CameraStack Viewer, with the expected differences in confidence and false positives. That establishes interoperability; a controlled accuracy comparison is separate work.
The vendor examples matter in practice. The open models show that a customer can deploy a fully redistributable stack, while a model supplied and tuned by the silicon or board vendor can be a stronger starting point for a commercial product. CameraStack supports both. It does not redistribute the Luckfox or Rockchip artifacts, and advertises them only when the operator has installed model files they are licensed to use.
Camera features stay close to the hardware
One resident 960x540 analysis branch feeds both Rockchip IVS and the NPU. IVS provides a 30x17 activity grid for up to four ONVIF MotionRegionDetector polygons. CameraStack maps each polygon onto that grid and applies the shared sensitivity and damping rules without scanning full video frames on the CPU.
The Imaging service reaches AIQ for brightness, contrast, colour saturation, sharpness, exposure, white balance and backlight compensation. An exposure-normalised ISP statistic drives the ImageTooDark event, so covering the lens still produces the correct transition after automatic exposure raises the gain.
Both profiles have independent JPEG snapshots and four text OSD slots. The OSD is applied before video compression and attached to the JPEG path too, so recorders, viewers and snapshots see the same text. OSD and Imaging state survive encoder idle periods, codec changes and the 1080p to 4K resolution change.
The result is a strong RV1126B starting point for camera manufacturers. The capture and hardware-control integration is already done, the vendor-specific code stays small, and the rest of the product comes from the same core CameraStack maintains across its reference ports.
Building an RV1126B camera? The port details are on the RV1126B chipset page, and evaluation builds are available: get in touch.