CameraStack Server now has a reference port for the Ingenic T31. The T31 is a compact camera SoC used in a large number of low-cost IP cameras, and on this port it runs the same CameraStack Server as our larger platforms: ONVIF Media and Media2, Imaging, events, polygon motion detection, person detection, multiple video profiles, and RTSP streaming over every transport including TLS.
We built and tested it on a Vanhua VH-NM-Z55I camera module: a T31X with a GalaxyCore GC4653 sensor delivering 2496x1440 at 25 fps. Nothing in the port depends on that module beyond its sensor and memory layout; it builds against Ingenic’s IMP media SDK.
A full camera feature set in 64 MB
The T31X has 128 MB of memory in the package. Half of it is reserved for the media pipeline, where the ISP and the hardware encoders keep their picture buffers. Linux gets the other 64 MB, of which about 58 MB is usable once the kernel has loaded.
That 58 MB accommodates the whole camera: the ONVIF services, the RTSP server and its transports, TLS, the event and analytics engines, the object tracker and the configuration store. In our test runs the CameraStack server process used 11 to 14 MB of resident memory, and the native producer that drives the vendor SDK used 2 to 7 MB, depending on which features were active. The rest is left for the operating system and for the product’s own software.
There is no reduced edition of CameraStack for small cameras. The server on the T31 is built from the same source as the server on an NVIDIA Jetson, with the same ONVIF services, the same RTSP transports and the same configuration and security model. A client sees the same camera behaviour on both.
Keeping the work off the CPU
Most of the camera’s work never reaches the CPU. A small C producer drives Ingenic’s sensor, ISP and encoders directly through the IMP SDK. Video stays in the media pipeline until it is encoded, and only the encoded H.264 or H.265 pictures, motion events and detections pass to the server. There is no GStreamer runtime on the device.
The ISP and the encoders produce the video. Motion detection reads the ISP’s own statistics grid, so up to four ONVIF MotionRegionDetector polygons and the ImageTooDark tampering event never scan video frames on the CPU. The single MIPS core handles control, networking and packetising. In samples taken during four-hour runs streaming both profiles, the core was about 40 percent idle, and about 30 percent idle with person detection turned on.
Person detection is optional. It runs on the core’s SIMD instructions through Ingenic’s JZDL runtime, at one or two inferences per second, on images taken from the substream. Detections go into the shared tracker and come out as ONVIF scene metadata, the same as on the NPU platforms. A product that does not need it can leave the model package out and keep video and motion events.
Video profiles and camera controls
The T31’s two hardware encoders serve up to four ONVIF profiles, so clients can create extra profiles without needing more encoders. Each encoder can use H.264 Main or H.265 Main. Bitrate changes apply to the running encoder, and a codec or resolution change rebuilds only the selected channel while the sensor, ISP and the other stream keep running.
The Imaging service operates the running ISP: brightness, contrast, saturation, sharpness, exposure, gain, white balance and backlight compensation. Hardware JPEG snapshots and a four-slot text and date/time overlay complete the set, with the overlay composed into both video streams and the snapshots.
Building a T31 camera
The T31 sits in a very large number of camera designs. For teams building on it, the camera software is now an integration on top of the Ingenic SDK rather than a new project. If you buy T31 camera modules, from Vanhua or another module maker, and need custom firmware with a professional ONVIF and RTSP stack, we can build it.
The port details are on the Ingenic T31 chipset page, and evaluation builds are available: get in touch.