Overview

This paper introduces a method for action recognition using 3D point clouds generated from Frequency Modulated Continuous Wave (FMCW) radar signals. The approach applies Support Tensor Machine (STM) to classify various actions based on the 3D spatial and temporal information contained in the point clouds.

Key Contributions

3D Point Cloud Generation:
- Applied 3D Capon Beamforming to generate 3D point clouds from FMCW radar data.
- Utilized subsampling to reduce computational complexity while maintaining accuracy.

Crouching

Hands Up

Sitting

Standing Up

Walking

Action Classification:
- Used Support Tensor Machine (STM) for classification, achieving an accuracy of 82% across five distinct actions.
- Analyzed misclassification cases and identified potential improvements.

Experimental Setup

Radar Sensor: Texas Instruments IWR6843-ODS
Tested Actions: Sitting, standing, walking, raising hands, squatting.
Participants: 11 individuals performing each action within a 0.5m–2m range from the radar.

Results

Achieved 82% accuracy in classifying five different actions.
Highlighted key areas for further improvement, such as handling overlapping gesture patterns.

Poster (in Korean)

Conclusion

This research demonstrates the feasibility of using FMCW radar for robust action recognition, providing a cost-effective and privacy-preserving alternative to camera-based systems. (Yoon et al., 2023)

Additionally…

I also worked for fall detection with 3D point cloud with LSTM and RNN structures!
The most difficult part in radar projects is always getting large amount of dataset😇

Action Recognition from 3D Point Clouds