Human activity recognition (HAR) has been actively researched in fields such as healthcare to understand and analyze human behavior in human-robot interaction. However, most studies have struggled to recognize activities like turning and motion transitions, which are often associated with dynamic balance. Therefore, we propose a novel HAR approach using a single sensor to collect and early fuse motion and position data. The aim is to enhance the accuracy of motion classification for daily activities and those that cause imbalance, which have traditionally been difficult to recognize. We constructed a quarantine room environment for data collection and to evaluate the impact of the suggested features on behavior. Five deep learning models were trained and evaluated to identify the optimal model. The collected data was classified and analyzed by the selected model, which demonstrated an average accuracy of 98.96%.