Hi Simon et al
I've done a few projects with pose/hand/face tracking running (on mobile), then communicating via OSC messages with a Pd process (in one case on mobile device, in other case over local network to a laptop). For these I've used MediaPipe, this is the pose detection 'solution':
which has a python example app (I've only used the mobile example apps, which actually built and ran on the first try...I can't always say the same for python environments)
Note that the implementation (e.g. the raw graphs and models, etc) are openly available as difficult-to-build targets (using the Bazel build system), or wrapped into prebuilt libraries (via existing dependency managers); it's not an ideal open-source setup (which to me entails easy use/modification of lower levels on canonical build systems). But it should certainly be hackable enough for your purposes.