A Review : Human Activity Recognition

Abstract

A discipline called Human Activity Recognition (HAR) uses embedded sensors in cellphones and wearable technology to collect raw time-series information and infer human actions from them. It has become quite popular in a variety of smart home situations, particularly for continually tracking people's actions in ambient assisted living to offer geriatric care and rehabilitation. The system uses a variety of operating modules, including data collecting, noise and distortion removal during pre-processing, feature extraction, feature selection, and classification. Modern feature extraction and selection methods have recently been suggested, and they are categorized using conventional machine learning classifiers. However, the majority of the solutions make use of antiquated feature extraction methods that cannot distinguish between complicated activities. Deep Learning algorithms are widely employed in several HAR systems to efficiently recover features and classify data as high computational resources have emerged and advanced. As a result, the review paper's main objective is to provide a thorough summary of the deep learning approaches utilized in sensor-based identification systems for smartphones and wearable devices. The suggested methods are divided into traditional and hybrid deep learning models, each of which is presented along with its special qualities, benefits, and drawbacks. The study also goes through numerous benchmark datasets that are employed in current methods. The report concludes by listing a few difficulties and problems that need more study and development.