Features Extraction Methods For Daily Activity Recognition Based On Accelerometers

With the development of computer science and technology, it is turned people’s lives toward intellectualization and digitization gradually. This is an era where our work and entertainment was digitized, shopping was digitized, and even walking, running and sleeping could be digital express.In recent years,there have been many researchers dedicated to identified human daily activity, and these achievement are widely used in the study of health monitoring, daily activity tracking, home and work automation, interaction homme-machine and so on. The general process of daily activity recognition include information collection, feature extraction and activity analysis. There are shortcoming of traditional activity recognition based on image in information collection, due to the method is sensitive to lots of factor such as color,shade and activity sheltered by other object.So many study of activity recognition is based on acclerometer data collected by portable devices integrated the accelerometer. In feature extraction, the present study has not yet proposed uniform and effective extraction method because of behavior and application environment. In activity recognition, many studies prefer to adopt supervised classification method rather than clustering method. The running of classification method based on supervised learning need the label of activity. But it is so boring and time-consuming in the process of the label, and it’s such a burden to activity recognition.We have done the following work to focusing on the above problem. First of all, we propose a framework of activity recognition based on the cluster method of MCODE. Compared with other methods, the advantage of this framework is that we don’t need to label activities during the process of information collection. Next, three methods of feature extraction are proposed in the above framework. The feature values extracted by there methods are all simple time-domain features, and the running tine of extraction is short.The framework proposed operates in three stages:feature extraction, there are three method of feature extraction proposed in framework; activity model built based on similarity measurement of Euclidean distance, this process build network used discrete data; cluster the activity model by the clustering algorithm MCODE, and then evaluate the cluster result using Clustering evaluationalgorithm FM-index. In the experiment, three method of feature extraction is used to analysis two data sets include different activity data. Through cluster analysis, we came exactly to the conclusion that the cluster results are 0.97 and 0.87. It can prove the that the framework is effective to activity recognition. At the same time,we made a comparative analysis on the effect of the different method of feature extraction in the experience of activity recognition.