Human Action Recognition (HAR) is frequently used to detect an individual’s motion patterns based on a series of observations of the individual’s body and environment. This technique is essential to many fields such as stroke rehabilitation and driver monitoring. There are two main approaches for a HAR system: vision-based and sensor-based. As more sensors (e.g., GPS and accelerometers) are integrated into modern smartphones, interest in action recognition using mobile devices has increased in recent years, where machine learning plays a central role in the action detection.
The performance of a HAR system depends on data quality and sensing and processing parameters to control data processing. However, to the best of our knowledge, the user-provided data is prone to human error, and the impact of each input parameter on classification performance is not fully understood. In this dissertation, we proposed three novel works to improve HAR performance. First, we summarized a set of typical sensing and processing parameters and evaluate their impacts on classification performance using regression analysis. Second, we applied and adapted the concept of intraclass correlation coefficient to measure the consistency of time-series data (such as accelerometer readings). Third, to detect and remove smartphone label error, we proposed a scheme for four stratified- trained classifiers and an ensemble technique for them. These works provide a guideline to improve HAR performance from both training data and parameter optimization.