Non-Invasive Human Activity Recognition using Radar dataset

Abstract

Human Activity Recognition has enabled state of art applications in medical healthcare, surveillance systems, digital entertainment and various other sectors. Therefore, prediction of such kind of movements remained an interesting aspect in the field of research. Wearable sensor and vision- based systems have been utilized for the detection of Activities of Daily Life (ADL), however suffer from various limitations including intrusiveness, lighting conditions and privacy issues. This study proposes a transfer learned, deep learning model for Frequency Modulated Continuous Wave (FMCW) radar-based system operating at a frequency of 5.8 GHz with 400 MHz bandwidth for classification of human activities. Radars are highly sensitive to human body movements and can capture small variations as Doppler shift. I have designed and demonstrated a generalized deep learning classification system for the detection of ADL irrespective of the geographical location and the ages of the subjects based on data from FMCW radar-based system. I have utilized experimental public data from 99 participants consisting of 1453 micro-Doppler signatures, at nine different locations including five retirement homes and four laboratories. This study propose micro-Doppler images normalized for speed profile of individuals to obtain age group independent feature maps of human activities and engineer a deep convolutional neural network architecture with a high classification accuracy, sensitivity and specificity of 99.1%, 99.2% and 100%, respectively.