Log-Likelihood Clustering-Enabled Passive RF Sensing for Residential Activity Recognition

Abstract

Physical activity recognition is an important research area in pervasive computing because of its importance for in e-healthcare, security and human-machine interaction. Among various approaches, passive RF sensing on the basis of well-tried radar principle has potential to provides unique non-invasive human activity detection and recognition solution, and draws more attention. However, this technology is far from mature. This paper presents a novel HMM-log-likelihood matrix based feature characterizing of the Doppler shifts to break the fixed sliding window limitation in traditional feature extraction approaches. We prove the effectiveness of proposed feature extraction method in K-means&K-medoids clustering algorithms with experimental Doppler data gathered collected from a passive radar system. The time adaptive log-likelihood matrix-based approach outperforms the traditional SVD, PCA and physical feature based approaches, and reaches 80% in activity recognizing rate.

A simplified machine learning approach for recognizing human activity

“…Khalifa et al [19] have introduced a unique study where recognition is carried out from energy harvesting data. Similar line of research is also carried out by Hsu et al [20] where principal component analysis is used for dimensional reduction Usage of Hidden Markov Model has been used for constructing a framework for activity recognition as seen in work of Li et al [21] along with usage of k-means clustering for feature extraction. Hidden Markov model and supervised classification technique are also reported to be used in the activity recognition system (Sok et al [22]).…”

Unsupervised Doppler Radar Based Activity Recognition for e-Healthcare

“…From each range-Doppler plot, the range column which contains the detected subject is extracted to form up the Doppler spectrogram. More details can be found in Section 3 in [20] or Section III in [23].…”

“…Micro-Doppler radars have been extensively used for activity recognition with a focus on healthcare purposes [20], [23], [24]. Currently, majority of studies are based on pipelines, which are totally supervised or consist of a combination of unsupervised and supervised approaches.…”

“…In a pilot study, the Doppler-Radar-2018 dataset was used. The work employed Hidden Markov Models (HMM) in order to extract activity information from each Doppler sequence [23]. The output of the HMM training was clustered using K-Means and K-Medoids.…”

“…In this paper, the Doppler-Radar-2018 dataset, that was used previously in [23], is considered. An unsupervised framework is developed despite labels availability.…”