CameraStack on the
Allwinner A733.

How this port works

The A733’s camera and encoders are reached through Allwinner’s native libraries. CameraStack handles platforms like this with a C producer: a compact native adapter, written in C straight against Allwinner’s VIN, ISP, G2D, Cedar and VIPLite interfaces, that owns the vendor media pipeline from sensor to hardware encoder and hands encoded video to the portable core. The Go server runs unchanged above the producer and does not link the vendor SDK.

Control flows the other way too. Because the producer owns the encoder, settings changed through ONVIF are applied live: frame rate, bitrate and GOP change mid-stream, each change publishes a fresh IDR keyframe first, and synchronisation-point requests map to on-demand keyframes. Encoding is demand-driven per profile: each hardware encoder runs, and draws power, only while clients consume its stream, while the resident pipeline keeps snapshots, motion events and object metadata flowing with both video encoders idle. The proven profiles are 1080p H.264 High at up to 30 fps and a 640x360 H.264 High sub-profile.

Analytics, Imaging and OSD

The Analytics service runs ONVIF MotionRegionDetector rules: up to four named polygons, each with its own sensitivity and damping. The producer uses the A733’s G2D engine to reduce a borrowed full-resolution frame to a 320x180 luma plane at 5 fps, CameraStack averages that into a 32x18 activity grid and rasterises arbitrary ONVIF polygons onto it, so motion detection never scans full video frames on the CPU and keeps running while both video encoders are idle.

Object detection runs on the A733’s NPU. The ObjectDetector analytics module advertises the models present on the device: the royalty-free portable PicoDet-S and PP-YOLOE-S alongside Allwinner’s own YOLOv5s. A standard ONVIF client can switch the model, confidence threshold and 1 or 5 fps detection rate live. A model change replaces only the resident VIPLite network and its buffers; the sensor, ISP, both encoders and the daemon stay resident. Detected objects are tracked with stable IDs and published as ONVIF scene metadata, and metadata-only clients can subscribe without waking a paused video encoder.

The Imaging service is wired to the running ISP: colour saturation, sharpness, automatic or manual exposure with manual exposure time and gain, and automatic or manual white-balance mode. ImageTooDark is derived from the same analysis plane, normalised by live ISO and exposure time, so a covered lens stays detected when automatic exposure raises the gain to compensate.

JPEG snapshots borrow the same fresh full-resolution frame as video, at both profile resolutions, including while the video encoders are paused. Four text OSD slots per profile provide black or white palettes, live clock text, standard and custom positions and the full transparency range; video uses Cedar’s hardware overlays while the JPEG path composites the same OSD state in software, so snapshots and streams always match.

Built on the Allwinner SDK

The producer builds against Allwinner’s camera and media libraries. Capture runs through VIN, the video input module that owns the sensor, with the vendor ISP and its tuning; Cedar drives the hardware H.264 and JPEG encoders; G2D performs the colour conversion and resize that feed motion analysis and the object models; and VIPLite runs those models on the NPU. The producer builds on the same shared common core as CameraStack’s other custom-SDK platforms, so framing, configuration, motion-grid mapping, event formatting, OSD rendering, pause state and the object-model decoders are well-tested portable code, with only small A733-specific customisations on top.