Dataset for Real-World Human Action Detection Using FMCW mmWave Radar

TL;DR

This work addresses real-world, privacy-preserving human action detection using mmWave radar by introducing a dataset collected in 28 private homes, focusing on sit-down and stand-up transfers annotated with a thermal sensor to preserve privacy. It leverages a baseline CNN architecture and analyzes input representations built from radar feature images $($$DT$$, $XT$, $YT$, $ZT$$)$ and their combinations, using a sliding-window detection framework. The study reveals substantial gaps between validation and test performance due to real-world variability, limited spatial diversity, and action-location distribution, with the best results achieved using $DT$+$ZT$. The dataset provides a realistic benchmark for HAR in in-home settings and highlights the practical challenges of deploying privacy-preserving radar-based action detection in everyday environments.

Abstract

Human action detection using privacy-preserving mmWave radar sensors is studied for its applications in healthcare and home automation. Unlike existing research, limited to simulations in controlled environments, we present a real-world mmWave radar dataset with baseline results for human action detection.

Figures (6)

• Figure 1: Chirp Smart Home Sensor (left) is installed in the homes at 122 cm (48") height (right).
• Figure 2: Radar signal processing to obtain 3D point cloud.
• Figure 3: Low resolution thermopile data used for annotation.
• Figure 4: Distribution of sit-down and stand-up action locations in the dataset. Each arc is one meter apart.
• Figure 5: Baseline CNN model architecture from wuDTXTYTZTPaperjin2019multiple.