An Approach to Classify Human Activities in Real-time from Smartphone Sensor Data

According to the scenario of the last few decades, sensor-based human activity recognition has gained considerable research attention due to its novel applications in health care, machine learning, human-computer interaction, etc. In this research, we have investigated several data pre-processing methods to recognize static and dynamic activities more accurately based on smartphone sensor data. The addition of magnitude and jerk-based features ensure the recognition of activities independent of the orientation of the smartphone in daily life. Besides, we have proposed a new feature namely Average peak-trough-distance (APTD), which ensures better recognition accuracy. We have also discussed a sliding window technique without overfitting, which utilizes observation of test data from previous windows with 50% overlapping for the detection of activity in real-time. Moreover, we have evaluated our proposed method showing a comparison between four classifiers. We have found the best accuracy for HASC2010corpus dataset using our proposed feature along with the multi-class Support Vector Machine, where we have utilized the Grid Search Cross Validation optimization technique to tune the hyper-parameters. We have also shown that the proposed approach outperforms the naive methods for the detection of activity using smartphones more precisely in real-time applications.