Impact of Time Domain Features & Inertial Sensors on Activity Recognition using Randomized Selection

Activity performed by the user is one of the major components of context sensing. Now a day's users are carrying Smartphones or wearable devices with them always. The device is fully equipped with the latest sensors, thus smart devices are prominently used in activity detection. The detection of activity is mainly dependent on three things-1- the sensors used for data collection, 2- the various features extracted from the raw data and 3-Machine Learning model used for training and testing. Researchers are using different sensors and extracting more numbers of features for getting better accuracy. However, feature dimensions are dependent on time of execution. Thus, an optimization is required between number of features used and its execution time. It is also required to find out the impact of different sensors on its accuracy and execution time. In this paper we have tried to discover the trade-off between number of features & sensor used with its accuracy and execution time. The evaluation of proposed work has been done by using publicly available dataset on UCI machine learning repository. Random selection methodology is used for selecting features and 5 popular machine learning algorithms is used to compare the results. The evaluation result shows that gyroscope helps in increasing accuracy if it is used along with accelerometer. We also conclude that features have significant effect on accuracy and execution time, and from various ML models Random forest & K nearest neighbor classifiers are providing better accuracy in most of the cases of Activity Recognition.