A Raspberry Pi camera, captured from libcamera through rpicam-vid.
Declared like any other sensor:
topology do
link :head do
sensor :eye, {BB.Camera.Libcamera, camera: 0, width: 1280, height: 720}
end
endConsumers never talk to this module. They go through BB.Camera, which
refcounts them and asks the driver to start capturing when the first one
arrives and to stop when the last one leaves. Two browser tabs on the same
camera produce one rpicam-vid, not two.
More than one camera
A Pi 5 has two camera connectors, and a Compute Module carrier may have more. Declare one sensor per camera, each naming its own:
link :head do
sensor :left_eye, {BB.Camera.Libcamera,
camera: "/base/axi/pcie@1000120000/rp1/i2c@88000/imx219@10"}
sensor :right_eye, {BB.Camera.Libcamera,
camera: "/base/axi/pcie@1000120000/rp1/i2c@80000/imx219@10"}
end:camera is an index or a libcamera camera id, both as
BB.Camera.Libcamera.Devices.list/1 reports them. Prefer the id when there
is more than one camera: it names the connector, while an index is a
position in a list that shifts when a camera fails to probe. An id is
resolved to an index by listing the cameras afresh at the start of each
capture, which takes about a tenth of a second.
Each camera is captured by its own rpicam-vid, and each allocates its frame
buffers from the kernel's contiguous memory (CMA). A board with little of it
- 64 MiB is common - can run out with two cameras at 720p, and the second
capture then fails with
dmaHeap allocation failureon the console. Either give the kernel more CMA, or lower:buffer_count; seeoptions_schema/0.
What it captures
JPEG, encoded in software by rpicam-vid. A Pi camera produces raw Bayer
data, the Pi's ISP turns it into YUV, and rpicam-vid encodes that with
libjpeg, so every frame costs CPU in proportion to its size. :quality
trades size for fidelity.
What it reports
Capture is confirmed to the hub on the first whole image, not when
rpicam-vid starts. A camera that is missing, busy, or cannot get buffers
would otherwise announce BeginCapture a moment before failing.
Dimensions come from the SOF marker of each frame, which is what was
actually encoded, and BB.Camera.stream_info/2 reports those once capture
has produced something.
The frame rate in stream_info is the one that was requested. libcamera
is asked for it, and chooses a sensor mode that can deliver it, but this
driver does not read back what it got. What the frames actually did is
measured by BB.Camera.Hub and reported separately in
BB.Message.Camera.CaptureState.
Frames do not go on pubsub
A robot's pubsub carries capture lifecycle only:
BB.Message.Camera.BeginCapture and EndCapture, published by the hub.
Frame bytes travel over BB.Camera's bounded media interface, and never
appear in an event or an error.
Options
See options_schema/0. Any of them can be filled from a parameter with
param/1, and changing that parameter while the robot runs reconfigures the
camera - see handle_options/2.
Limitations
- Raspberry Pi only, in practice:
rpicam-vidis built for the Pi's camera stack. - Timestamps are arrival times. The MJPEG stream on
rpicam-vid's stdout carries no timestamps, so a frame is stamped when the read completing it reached the BEAM. SeeBB.Camera.FFmpeg.Capturefor what that does and does not mean. - A failed capture is not retried. The failure is reported and capture ends; reopening the camera is the consumer's decision, taken by subscribing again.
- Changing an option that reaches
rpicam-vidrestarts it. Capture ends and begins again around the change, with the events to match. rpicam-vid's diagnostics go to the VM's console, not into the error.
Summary
Functions
Take up options that changed while the robot was running.
The format this camera is declared to produce.
Functions
Take up options that changed while the robot was running.
Reached when a param/1 reference in the camera's DSL options resolves to a
new value. Only a change that alters the rpicam-vid invocation restarts
capture: libcamera configures the camera once, when it starts, so a new size
or rate means a new rpicam-vid. Capture stops and starts again around the
change; it is not a seamless transition and does not pretend to be. An
option that doesn't reach the command line - a framer bound, say - takes
effect without disturbing the stream.
New options this driver cannot use stop the camera. handle_options/2
has no way to refuse a change and stay running, so a value that satisfies
the parameter's declared type but not this driver - a width of -1 against
a :pos_integer - takes the sensor down, and down again on each restart,
because the parameter was stored before anyone was asked about it. Declare
parameters with bounds that match the options they feed.
The format this camera is declared to produce.
Pure: it is called during discovery against robots that are not running, so it reads the options and nothing else. Options that still hold unresolved parameter references cannot imply a format, and produce an error rather than a guess.