Improving Radar Human Activity Classification Using Synthetic Data with Image Transformation.

Rodrigo Hernangómez,Tristan Visentin,Sławomir Stańczak,Hamid Khodabakhshandeh,Lorenzo Servadei

doi:10.3390/s22041519

Abstract

Machine Learning (ML) methods have become state of the art in radar signal processing, particularly for classification tasks (e.g., of different human activities). Radar classification can be tedious to implement, though, due to the limited size and diversity of the source dataset, i.e., the data measured once for initial training of the Machine Learning algorithms. In this work, we introduce the algorithm Radar Activity Classification with Perceptual Image Transformation (RACPIT), which increases the accuracy of human activity classification while lowering the dependency on limited source data. In doing so, we focus on the augmentation of the dataset by synthetic data. We use a human radar reflection model based on the captured motion of the test subjects performing activities in the source dataset, which we recorded with a video camera. As the synthetic data generated by this model still deviates too much from the original radar data, we implement an image transformation network to bring real data close to their synthetic counterpart. We leverage these artificially generated data to train a Convolutional Neural Network for activity classification. We found that by using our approach, the classification accuracy could be increased by up to 20%, without the need of collecting more real data.

Highlights

We focus on human activity classification with radar due to the research interest that it has currently awakened [5,6,7,8], for which we count on the help of millimeter-wave radar sensors operating at 60 GHz
We have focused on improvements of deep learning methods for human activity classification on radar images
We have presented our own architecture, Radar Activity Classification with Perceptual Image Transformation (RACPIT), to mitigate the dependency of classification accuracy of Convolutional Neural Networks (CNNs) on the dataset recorded for the training process

Summary

Introduction

The fast advances in Machine Learning (ML) research help to solve more and more important problems. Generating high-quality datasets that adequately represent reality to its full extent is still highly demanding. Transfer learning and Generative Adversarial Networks (GANs) assist the improvement of classification or regression tasks from incomplete datasets [4]. We focus on human activity classification with radar due to the research interest that it has currently awakened [5,6,7,8], for which we count on the help of millimeter-wave radar sensors operating at 60 GHz. Problems with datasets in this context can originate from several reasons: 1. Incomplete dataset, meaning that a limited number of subjects executed the activities Problems with datasets in this context can originate from several reasons: 1. Incomplete dataset, meaning that a limited number of subjects executed the activities

Methods

Results

Conclusion