Research on Human Activity Recognition Based on Smartphone Sensors

Abstract

To address the problem that current smart bands and fitness applications require users to manually select the exercise mode, which often leads to incorrect activity records, this paper studies an automatic human activity recognition method based on the built-in sensors of smartphones. The method is evaluated on the public UCI HAR dataset, which contains sensor records of six daily activities (walking, walking upstairs, walking downstairs, sitting, standing, and laying) collected from 30 volunteers, with 561 time-domain and frequency-domain features extracted from the signals of a three-axis accelerometer and a three-axis gyroscope for each record. Three classification models are compared: the decision tree, the K-Nearest Neighbor based on Euclidean distance, and KNN after PCA dimensionality reduction. The experimental results show that KNN achieves the best performance with an accuracy of 90.16%; the decision tree reaches 86.29%; after the features are compressed from 561 to 30 dimensions by PCA, KNN still maintains an accuracy of 87.41%, which significantly reduces the computational cost. The confusion matrix analysis shows that sitting and standing are the main sources of confusion (136 misclassifications in total), because the sensor signals of these two low-intensity static activities have extremely low amplitude and highly similar statistical features. In addition, a signal-level simulation platform is built to verify the learning curves, the hyperparameter sensitivity, and the noise robustness. The results prove the feasibility of high-accuracy activity recognition locally on mobile devices and provide a reference for the automatic activity recording of smart wearable devices.