ubiMonitor: Intelligent Fusion of Body-worn Sensors for Real-time Human Activity Recognition

Real-time monitoring of the human daily activities using wearable sensors has attracted lots of attention from researchers due to its crucial role in emerging fields such as pervasive healthcare and ambient assisted living. However, despite being an active field, recent benchmarks show a difficulty in providing robust monitoring of the different various activities under realistic everyday life conditions. In this paper, based on an in-depth understanding of the nature of 3D accelerometers and the physical properties of the different daily activities, we propose ubiMonitor as an accurate real-time activity monitor using low-cost off-the-shelf three body-worn 3D accelerometers. For a robust monitoring, ubiMonitor (1) intelligently fuses the accelerometers and the acceleration in each of the three axes to provide key discriminative features for the different activities, (2) employs a novel hierarchical activity recognition scheme, and (3) applies an effective postprocessing stage to remove falsely detected activities and enhance the overall system accuracy. Experimental results using real traces from different eight subjects show that ubiMonitor can achieve an overall accuracy more than 95% with a median latency less than 3 msec. This is better than state-of-the-art by 23.4% in the recognition accuracy with a reduction of 70% in the sensors used.