Human Activity Recognition via Smartphone Embedded Sensor using Multi-Class SVM

Human Activity tracking is the process of detection and understanding of the human activity. It can be done by analyzing human motion behavior data extracted from different smartphone-embedded sensors. Recognizing human activity has become widely popular and particularly attracted many researchers in different industries. Activity recognition has become increasingly important in many areas, especially for the recognition of fitness, sports, and health monitoring. This paper propose a robust model that is trained and tested on remotely extracted data from the smartphone-embedded inertial sensor. Initially, the system clean the input data and then performs windowing and segmentation. After pre-processing, a number of features are extracted. Further, the Lukasiewicz similarity measure (LS) based features selection is used to reduce the features set by removing the least important features. In the next step, the Yeo-Johnson power transformation method is utilized to optimize the selected features. The optimized features set is then forwarded to the multi-class support vector machines (SVM) classifier. The system was designed and experimented with over a well-known dataset named WISDM. The presented model performed well by achieving a mean accuracy rate of 94%.