Human Action Recognition Using Triaxial Accelerometer Data: Selective Approach

With the deployment of kinematic sensors, the recognition of human activities using triaxial accelerometer has become unavoidable in various fields. In this work, we propose an approach based on the selection of the most expressive signals describing the action. This selection is based on an entropy calculation as it presents the amount of information contained in the signal. The extracted descriptors are relative to the time-frequency domain. For classification, we used support vector machine to identify and to recognize the different actions. We proved the effectiveness of our approach by following experiments carried out on 3 public databases. Thus, the performances found are comparable to those introduced by other works.