Assessing Human Activity Recognition Performances of Different Machine Learning Algorithms Using Sensor Data

With how extensively various kinds of sensors have become a part of modern life, a vastness of application and research areas has sprung up in its wake. One such use case is the budding field of Human Activity Recognition (HAR) which powers medical care, security purposes, assistive technology in homes, navigation systems, etc. In this work, the focus is to assess the performance of different machine learning principles using sensor inputs to form an understanding of their suitability in the context of HAR. Random forest, SVM, neural network and kNN algorithms are chosen for training and testing experiments conducted on a benchmark dataset, the UCI-HAR. At the outset, base models are developed and their accomplishments are observed to highlight those performing encouragingly, after which randomized type of hyperparameter tuning is utilized to further upscale the selected classifiers’ effectiveness in activity detection. This tuning method is packed with some useful advantages like lessening the computational burden while maintaining the quality. The results obtained show that out of all the models presented, the best accuracy of 96.54% is achieved by the tuned SVM model on the aforementioned dataset. Confusion matrix analysis suggested that standing and sitting activities see confusion to some extent across most models. Moreover, tweaking with the hyperparameters proved most fruitful for the Neural Network model.