Abstract
Human activity recognition (HAR) has gained significant attention in recent years. Previous studies used various features from time, frequency, and wavelet domains to recognize the activities, but it is not clear how to determine the best features that can efficiently identify activities in less time. In this study, we aim to explore and elucidate the significance of the most relevant features in HAR, shedding light on their semantic meanings. We utilize different filter-based techniques like chi-square, correlation coefficient, and information gain, coupled with forward selection to find the top-ranked features. Additionally, we apply union and intersection operations on the resultant feature subsets of filtering techniques to obtain common and aggregated features. By considering these common relevant features, we observe a significant decrease in computational time by 97%, 95%, and 90%, with a reduction in memory usage by 10%, 25%, and 20% for benchmark datasets, i.e., UCI HAR and WISDM (phone accelerometer and gyroscope datasets) respectively, even though with a slight compromise in accuracy.
