User Activity Recognition via Kinect in an Ambient Intelligence Scenario

The availability of an ever-increasing kind of cheap, unobtrusive, sensing devices has stressed the need for new approaches to merge raw measurements in order to realize what is happening in the monitored environment. Ambient Intelligence (AmI) techniques exploit information about the environment state to adapt the environment itself to the users' preferences. Even if traditional sensors allow a rough understanding of the users' preferences, ad-hoc sensors are required to obtain a deeper comprehension of users' habits and activities. In this paper we propose a framework to recognize users' activities via a depth and RGB camera device, namely the Microsoft Kinect. The proposed approach takes advantage of the position of relevant human body joints estimated by using Kinect depth information. In our system, significant configurations of joints positions (i.e., postures) are discovered by a clustering approach and classified by means of a multi-class Support Vector Machine. Then, each activity is modeled by Hidden Markov Models (HMMs) as a sequence of known postures. In order to maintain a high level of pervasiveness, a real prototype has been implemented by connecting the Kinect sensor to a miniature computer with limited computational resources. Experimental tests have been performed on a dataset we collected at our laboratory and results look very promising