Searching Efficient Models for Human Activity Recognition
Shamisa Kaspour, Nikhil Raj, Alankrit Mishra, Abdulsalam Yassine, Thiago Eustaquio Alves de Oliveira · 2021
Human Activity Recognition (HAR) can be measured in various ways in a new era of growing technologies. This paper studies different classification and data processing tasks. This paper proposes using HAR to monitor the elderly while being power-efficient and respecting an individual’s privacy, allowing it to be run on mobile devices like smartphones or smartwatches. Upon reviewing other methods of HAR by sensor data, we realized that they severely lacked in the areas mentioned earlier. Moreover, they either used older classification techniques or made too complex and over-the-top models for the same. We tested a total of nine methods to find the best model/method, from simple support vector machines (SVM) and convolutional neural networks (CNN) to hybrid models. The best results were produced by a simple, fully connected network (multi-layer perceptron) with the data condensed using Fisher’s linear discriminant analysis (FLDA) that gave us 98.6% accuracy. Our final model satisfies both the requirements we had set; it is simplified and produces benchmark results.