Two phase ensemble classifier for smartphone based human activity recognition independent of hardware configuration and usage behaviour

Human activity recognition is an important technology in pervasive computing as it provides valuable information for smart healthcare and assisted living applications. Use of smartphones for activity recognition poses new challenges due to variation in hardware configuration and usage behaviour like how the smartphone is kept. Only a few recent works address one or more of these challenges. Consequently, in this paper we present a two phase activity recognition framework for identifying both static and dynamic activities addressing above mentioned challenges using smart handhelds. The framework through feature selection and ensemble classifier, address the variance due to different hardware configuration and usage behaviour. The two-phase framework is implemented and tested on real dataset collected from ten users with six different device configurations. Data is collected from accelerometer only as this sensor is available in any kind of smart handheld devices. In the first phase, the best training set is identified that is fed to the ensemble as input. In the next phase, the classifier based ensemble gives the final output through majority voting. It is observed that, with our proposed two phase classification, the accuracy level of 98% can be achieved for activity recognition while maintaining energy efficiency as only time domain features are considered.