Add vision subpackage with MediaPipe HeadTracker

[alozowski]

Adds daemon-side head tracking, that can run out-of-process so it never stalls the 50 Hz control loop, and to coexist with the conversation app on the CPU-bound CM4 / Pi 4 (supersedes the original in-process MediaPipe migration this branch started as)

What changed

• new vision/subpackage: YuNet detector (cv2.FaceDetectorYN) + a single-target tracker, run in a spawned multiprocessing.Process (own GIL/core). Replaces the in-process MediaPipe tracker, which serialised on the GIL and starved the loop.
• daemon eases the tracking aim per 50 Hz tick -> motion stays smooth regardless of detection rate, so the detector can run slow (10 fps) and cheap
• dedicated valve-gated, downscaled + rate-limited camera branch feeds the worker over its own unixfdsink socket – converts frames only while tracking, nothing when idle.
• full-frame detection at 320 px / 10 fps (no ROI crop), YuNet model fetched from the HF hub
• multi-person: gaze follows the closest / most-prominent face, with hysteresis so it doesn't ping-pong
• enable_head_tracking(weight=0) pauses (frees the head + detector CPU) without tearing down the process -> cheap per-turn on/off for the conv-app speak/listen handoff
• decouples the local IPC framerate from capture (helps Configurable served/IPC camera framerate to reduce CPU for on-device vision (Reachy Mini Lite locked to 1080p@60) #1241): serves the IPC feed at a capped fps so on-device vision apps don't pay the full capture rate
• drops the vision / mediapipe==0.10.14 extra, vision installs via the existing opencv extra. Missing OpenCV raises ImportError instead of exit(), so it can't kill the robot process

Validated on-robot (Wireless/CM4):

• loop holds ~50 Hz while tracking; coexists with the conversation app (~1.5 of 4 cores free).
• collision-safe with speaking + emotions/dances; crash isolation – a worker crash doesn't take the daemon down and self-heals on the next turn

DoA-driven search isn't yet in this PR, could be a next PR, as @FabienDanieau suggested.

Closes #1175
Conv app PR: pollen-robotics/reachy_mini_conversation_app#381

[FabienDanieau]

Is it an issue to have reachy_mini_toolbox as a dep in the conv app?

[pierre-rouanet]

LGTM, indeed as @FabienDanieau points out. Maybe we can use toolbox as a dependency. Not strong opinion on this. Just if we do, we have to make sure that everything included runs on a RM wireless.

Maybe next step is ton add start/stop/getface API entry points to trigger the head tracking in background?

[FabienDanieau]

I made some tests. rf-detr-nano is too heavy for the rpi. Mediapipe seems to be the most suitable solution for edge devices. There is a lighter model that we can use instead of the facemesh of the tool box: see face_track.py.
We need to mind the numpy 1.x dependency though.

Maybe next step is ton add start/stop/getface API entry points to trigger the head tracking in background?

I agree with @pierre-rouanet, the right call would be to enable head tracking on the daemon side (in a way similar we did for the wobbling see #1001). Basically we need to add a appsink to the video pipeline to feed the head tracker.

Also the tracking could be extended with the DoA. But let's keep that for another issue :)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[alozowski]

next checkpoint, almost the final: 5b42110

• we now resample instead of crop: detect on the whole frame downscaled to 320 px (gstreamer videoscale in the daemon tracking branch) every tick, without ROI window, so full FOV and robust to fast motion
• detection rate is 10 fps, but the head stays smooth because the daemon eases the aim at 50 Hz independent of the detection rate
• YuNet score threshold is 0.6 so it holds the face through head turns and at ~1 m
• YuNet model fetched from HF hub instead of the vendored ONNX (first enable needs network once, then it's cached)
• weight 0 now pauses tracking without tearing down the detector; weight > 0 resumes instantly
• local IPC camera feed capped at 15 fps (also touches the client camera_gstreamer.py) so on-device apps don't pay the full 30 fps convert

[alozowski]

another checkpoint, f710992

we don't yet have DoA tracking but I added a simple multi-person one:

• gaze follows the most prominent face (largest ≈ closest, weighted toward center) instead of sticking to the first one acquired
• switches only after a challenger leads by a margin for a dwell window, so no ping-pong between comparable faces as it used to be

[alozowski]

and another big checkpoint (5db721e) that dropped opencv from the tracking runtime, YuNet now runs on onnxruntime with a numpy decode:

• same detector and same weights: the model is the official dynamic-input YuNet export (face_detection_yunet_2026may, byte-identical weights to 2023mar), hosted on our HF and pinned by revision
• the decode reproduces cv2.FaceDetectorYN bit-for-bit, there's a parity test running both on the same model and frame
• frames are padded up to /32 (320x180 -> 320x192) because YuNet's feature pyramid needs stride-aligned dims, opencv did this silently, raw ORT crashes without it
• ORT session is pinned to 1 thread so the detector can't spread across cores and fight the control loop
• the robot installs [vision] (onnxruntime, ~16 MB) instead of [opencv] (~50 MB wheel / ~119 MB unpacked), so opencv is now dev/calibration-only
• onnxruntime bumped to 1.27.0 for both vision and nn_kinematics (verified FK/IK bit-identical), which needs python >=3.11

[alozowski]

another checkpoint, a5018b0: the tracker is now just a camera client

• dedicated tracking branch/socket/valve deleted: the tracker reads the same camera IPC as any local app, pause = the worker just disconnects
• one videorate total: IPC_FPS=10 on the daemon paces everyone, detector included (measured with a fakesink client: 123 frames / 12 s)
• aim at the nose now (eye midpoint made the robot look slightly above), single-face only
• explicit gst elements, direct try_pull_sample(), v4l2convert on RPi, no more ctrl+c traceback
• example + SDK docs added

next to try: drop the worker process and use a plain thread, and it would save ~200 MB of duplicated python/ORT/GStreamer stack

[FabienDanieau]

Yeah thanks!