Action Recognition from 3D Point Clouds
Recognizing human actions from 3D point clouds built from FMCW radar signals (IPIU 2023).
Overview
In this paper, we recognize human actions from 3D point clouds built from frequency modulated continuous wave (FMCW) radar signals. A support tensor machine (STM) classifies each action from the spatial and temporal information in the point clouds.
What I did
- Built 3D point clouds from FMCW radar data with 3D Capon beamforming, and subsampled them to cut computation without losing much accuracy.
Crouching
Hands Up
Sitting
Standing Up
Walking
- Classified the actions with an STM and looked through the misclassified cases to see where the model went wrong.
Setup
- Radar: Texas Instruments IWR6843-ODS
- Actions: sitting, standing up, walking, raising hands, crouching
- Participants: 11 people, each performing every action 0.5 to 2 m from the radar
Results
The STM reached 82% accuracy on the five actions. Most errors came from actions whose movement patterns overlap, which is the first thing I’d work on next.
Poster (in Korean)
Takeaways
FMCW radar can recognize actions reasonably well, and unlike a camera it doesn’t record anyone’s face. It’s also cheap. (Yoon et al., 2023)
I also tried fall detection on the same 3D point clouds with RNN and LSTM models. The hardest part of every radar project was the same: collecting enough data. 😇
