Adding Gaussian Noise to Sparse Frames from a mmWave Radar for Improving HAR

Taku Watanabe, Runhe Huang

Abstract: Millimeter wave (MmWave) sensors are used to recognize human movements. However, when there is little or no movement, the number of point clouds in a frame decreases or disappears, leading to decreased recognition accuracy. Addressing this challenge to improve accuracy is crucial research. This study addresses the challenge of improving accuracy in human activity recognition (HAR) using mmWave sensors. Sparse frames obtained from the radar often suffer from decreased point cloud density, leading to reduced recognition accuracy, especially in scenarios with minimal or no movement. To tackle this issue, we propose an algorithm that augments points in sparse frames by adding Gaussian noise. Leveraging the signal-to-noise ratio (SNR) of the point cloud, we generate points following a Gaussian distribution at appropriate coordinates. By introducing four-dimensional Gaussian noise based on SNR for x, y, z, and Doppler velocity dimensions, we achieve higher accuracy compared to using only spatial dimensions as predictors. We evaluate the efficacy of preprocessing on the point cloud using a simplified model consisting of LSTM, FC, and MaxPooling layers. Our results demonstrate that preprocessing improves accuracy, with an increase from 95.4% to 96.1 %.