A Comprehensive Analysis of Feature Extraction Techniques for Human Activity Recognition using Deep Learning

A comprehensive analysis of feature extraction techniques used in deep learning(DL) based Human Activity Recognition (HAR) across multiple domains such as healthcare, security, and sports were discussed in this article. DL techniques have demonstrated significant improvements in accuracy and resilience over traditional machine learning (ML) methods. Feature extraction is a crucial component of any activity recognition system, essential for determining the final classification outcome. Its primary goal is to capture shared attributes within images from the same class of activities. The term “feature” refers to information derived from raw images using various techniques. The complexity of feature extraction and the quantity of features needed directly affect the performance of activity recognition systems. This paper reviews various methods for feature extraction in HAR systems using DL. HAR systems are vital in several domains, such as healthcare, security surveillance, and human-computer interaction. The process of HAR involves several steps: first, data preprocessing is performed, followed by the extraction of spatial and temporal features (STF). These features are then merged and input into a classifier for activity recognition. According to the review, using CNN for feature extraction yields an accuracy of 98.74%.