Design And Implementation Of Energy-Efficient Scheduling For High Precision Activity Recognition Care System Using Smartphones

With the increasing competition and heavier pressure in modern society, people put more emphasis on personal health. In recent years, health care services based on user activity recognition technology continue to make new progress, leading to some commercial wearable products such as Fit Bit becoming increasingly popular and universal. However, due to certain restrictions such as extra data transmission condition, smart phones are still the most pervasive and effective platform for activity recognition, thus activity recognition applications based on smartphones emerge in endlessly. Nevertheless, the limited amount of power in smartphones becomes the bottleneck in the process of activity recognition and has been one of the primary problems which researchers manage to address. There are varied solutions to save energy consumption including the optimization of classifier algorithm, the rational management of sensors or other energy saving methods whereas the results of these studies still remain many problems in practical application. Therefore, the sampling, monitoring and reduction of energy consumption are yet the vital challenge for activity recognition and health services based on smartphones.We start by analyzing the current situation of researches in the field of activity recognition, followed by noting the problem of recognition accuracy and energy consumption in the process of activity recognition based on the smartphones. Based on these existing technologies and the latest research work, we propose an activity recognition health care system. First, we propose the overall design and function modules of our system, then we present details for the three sub-modules. In activity recognition sub-module, we explore the activity-dependent optimization parameter tuple choice in the data collection phase that different types of activities correspond to different sampling and classification feature sets. Besides, a Hybrid Classification Algorithm and Dynamic Period Identification and Calibration Algorithm are developed to distinguish nine kinds of periodic and non-periodic types of activity. After identifying user’s individual types of activity, we design a social network clustering module. In this sub-module, we establish a user’s social network diagram and compute the feature vectors of user’s community by adopting relational classification, and combine individual activity inference with Community inference using marking relaxation method. When the historical activity results are calculated, the community behavior influential factor is derived. The last sub-module is energy-saving scheduling module. In this module, we propose a learning based activity prediction model to abstract user’s daily activities of individuals and communities and produce activity time set and sequence of activities set. Based on these data sets, our system are able to predict the next activity categories of users and then dynamically adjust the sampling frequency of the sensors, classification feature sets as well as calculation cycle while the new identified categories of activities are supplemented to the original activity data sets at the same time. Finally, we design and do experiments to test and verify our system.Our system is designed and implemented on the Android mobile platform. During the experiment, we conclude that the activity recognition accuracy of our system has reached more than 97% in average, while the energy consumption is reduced by about 20% compared to conventional sampling patterns. In the long-term activity recognition process, due to continuously enriched learning model, the energy consumption can be further optimized and reduced by nearly 10%. The experimental results show that our paper can effectively provide a low-power, high-precision activity recognition health care service.