Activity Recognition with Wristband Based on Histogram and Bayesian Classifiers

With the increasing popularity of wristbands and their applications on the health monitoring, the recognition of various activities has become a significant feature which helps the users to assess and review their daily practices [1]. As a result, it can improve their health via the recorded information. The core task of the activity recognition is to automatically classify a large amount of motion data recorded by the accelerometers to one of the predefined activity classes such as eating, walking, etc. In contrast to the conventional recognition approaches [2], the algorithms are required to be simple, efficient and accurate as the battery powers and the allowable computational powers of the wristbands are very limited. Some approaches directly transfer the motion data to the mobile phones by the bluetooth so that more sophisticated computations can be performed. Nevertheless, the transmission of the raw digital motion data can consume a lot of power and occupy a considerable bandwidth. This paper proposes an algorithm that highly reduces the communications between the wristbands and the mobile phones by firstly preprocessing and labeling the raw motion data based on their characteristics. Only the labels and some encoding values are transferred to the mobile phones for performing the classification. In order to achieve the high accuracy, the classification is performed by combining different approaches such as the histogram approach and the Bayesian classifier approach. The obtained results show that the proposed approach achieves a good accuracy, which just uses some labels and the encoding value information.