Human Activity Recognition Based on FMCW Radar Using CNN and Transfer Learning

Abstract

Radar-based human activity recognition (HAR) is a fascinating research topic as it can be applied in various fields such as healthcare, security, and smart homes. As a non-invasive or non-contact technology that uses electromagnetic waves to detect human activities, radar offers the advantage of protecting visual privacy and remaining resilient against environmental conditions. Common machine learning (ML) methods used for HAR, such as Support Vector Machine (SVM) and k-Nearest Neighbor (KNN), depend on complicated parameter tuning, and high-dimensional data can degrade their efficacy. We propose a unified approach based on a convolutional neural network (CNN), specifically the VGG19 architecture with transfer learning, to extract features and classify various human activities from radar data. The advantage of using CNN is the ability to integrate feature extraction and classification in one learning phase. Transfer learning accelerates learning by utilizing pre-trained models' knowledge, enabling swift adaptation to new problem domains in machine learning. We compare our proposed method against traditional ML classification methods (SVM, KNN) and the combination of CNN-based feature extraction and ML-based classification methods. Our experimental results show that our proposed method performs better than the others with an F1-score of 93%.