Recognition of human activity in a non-controlled environment remained an unsolved problem. In this study, we wonder if accurate recognition of activity can be obtained using Internet of Things technology. We propose a processing methodology that can allow to an automatic system like a robot to recognize human activity. The approach consists of the classification of some activities of the subjects: walking, standing, sitting and laying. The study exploits a standard smartwatch and a smartphone carried by participants during a non-controlled experiment. We propose a pre-computation using Discrete Cosine Transform (DCT), and we identify the best window width and feature length that provides the best results. We show that Support Vector Machines (SVM) provides better results compared with Decision Tree algorithms (DT). The results also demonstrate that participants' activities were classified with an accuracy of more than 91% in a non-controlled environment, with a non-controlled position of the smartphone. We define the notion of transient that corresponds to the transition between two activities as well between two positions of the sensors. The last result shows that removing the transients provides better results for the classification, i.e. 98%. This approach shows that it is possible for a robot assistant to understand the human behavior only using standard Internet of Things technology in a non-controlled environment.