HUMAN ACTIVITY RECOGNITION USING MACHINE LEARNING ALGORITHMS

Abstract

The problem of recognizing human activities is used to define patterns of human behavior. Signals from built-in sensors of smartphones or wearable devices are used for recognition. The raw data must be preprocessed using a noise filter and sampled using fixed sliding windows. A feature vector is obtained from each window by computing statistical variables from the time and frequency domain. In this paper, traditional machine learning and convolution neural network models are trained on the preprocessed data. The best results were achieved using the Random Forest classifier (0.93 F-measure) and the convolutional neural network (0.94 F-measure). Based on the obtained results and the surveyed literature, the efficiency of different classifiers largely depends on the used data set, which is why many different techniques yield similar results.