Learning Time Synchronization in Wearable Sensor Fusion for Human Activity Recognition

A wide variety of embedded sensors broaden the versatile functions of the smart mobile devices. Activity recognition is one of them that could be harnessed to track movement for athletes, monitor daily activities for patients and help control a healthy diet. Prior study was performed by comparing the activity recognition performance between the smart-phones and the smart-watches based on accelerometer or gyroscope sensors. The smart-watch sensors were observed to be better than the smart-phone sensors in recognizing the hand-based activities with a higher classification accuracy, especially when eating activities were involved. However, the implementation of sensor fusion can further improve the activity recognition performance. In this paper, a pairwise index selection algorithm is proposed to implement sensors' time synchronization that paces the way for processing sensor fusion. Our empirical results indicate that the sensor fusion can improve the overall classification accuracy achieved by the single sensor in personal and impersonal models, respectively. The sensor fusion of a phone accelerometer and a phone gyroscope can partially overcome the location disadvantage of the smart-phones in recognizing each activity through feature concatenation. Furthermore, sensor fusion can better generalize the characteristics of activities across people in impersonal models.