Recognize Action Skeleton (STGCN PP)

@kennymckormick1 VersionClassify

Buffers a sliding window of per-frame Landmarks2d.Human22 skeletons and runs an STGCN++ graph-convolution network over the raw joint sequence to emit a ranked [Classification] or a single top-1 Classification of NTU RGB+D 60 actions.

How it fits

Typical backends

Live action telemetry. Label a body-landmark stream into the expected skeleton layout, recognise the action, and publish each result over HTTP.

Action-gated alert. Recognise actions from a landmark stream and let only the actions of interest pass downstream.

Caveats

State lifetimeRecognition is temporal — it consumes one person's skeleton per fire and holds an internal sliding window that must fill before any real action is produced; the buffer persists across fires until the container restarts.
Hard constraintJoint coordinates are normalised by the configured source frame width and height, so both must be set to the upstream image size or every joint is offset and accuracy collapses.
Parameter interactionThe window length, stride, and clip length together set the recognition latency and cadence; a longer window improves accuracy on slow actions but delays the first result.
AccuracyAction labels are drawn from the NTU RGB+D 60 taxonomy, so an action outside those classes is reported as the closest in-vocabulary label; this graph model is lighter than a heatmap-volume recogniser but slightly more sensitive to keypoint noise.

Versions

a5340da1latestdefault
Automated release
6/25/2026