Research On Sensor-based Activity Recognition

With the rapid development of microelectronics, sensor technology and the pattern recognition theory, human activity recognition based on MEMS sensors has gained wide attention of the researchers in recent years. The advantages of activity recognition based on sensors includes high portability, low power consumption and robustness to the surroundings. It has wide application prospects on motion tracking, human-computer interaction, augmented reality and other aspects. Most of the early studies designed wearable devices with sensors inside themselves. But with the rapid improvement of universal mobile devices including smart phones, activity recognition using mobile devices has become a research focus recently. However, comparing to the wearable devices, smart phones and other mobile devices don’t have a fixed orientation and position, and the computing power compared to the desktop platform is also a big gap. So there are still many problems to be solved.This thesis centers on human activity recognition using general mobile device. In order to achieve real-time and high accuracy activity recognition using mobile devices, this thesis has collected large amounts of sensors’data from different human activity under different position of a smart phone. A hierarchical recognition approach has been designed. This approach includes two kind of models, a device positons classification model and activity classification models for different device positions. These two models are connected. By get the position result from positions classification model, we can choose the specific activity classification model for the right position and get the final results. Therefore, a device positons classification model and activity classification models for different device positions have been trained, evaluated and analyzed. Meanwhile, the code implementation for this hierarchical approach on Android platform has been completed.The results show that the recognition accuracy of the hierarchical approach has improved 8.2%, comparing to regular approach (without position classification), and get up to 93.3%. Comparing to the recent researches, this approach keeps about the average recognition accuracy and improve the ability of mobile devices to adapt to different positions, and make the mobile devices be able to adapt to 8 different positions.