What it does
Classifies what a person is doing from the radar point cloud alone. Fine-tunes our base model on your own labels, so a few hundred sessions is usually enough.
What it needs from your data
Label · Yperson-scoped · single-label · 2–40 classes
Minimum20 sessions per class
Input · Xradar point cloud · fixed point count
Known limitations
Small movements are hard for radar to see. Someone waving a hand shows up far more faintly than someone jumping, so actions that differ only in small movements can be confused. If two of your labels differ that way, expect them to mix up.
Read the white paper →Baseline results
not yet benchmarked
Changelog
1.2.0Adds the 512-point input shape. Raises the per-class minimum from 12 to 20
1.1.0OpenPCDet adapter updated to v0.6
1.0.0First release