Abstract: In the last decade, the study of human behaviour activities within the field of Ambient Assisted Living (AAL) has led to the emergence of a variety of techniques to learn, detect, and recognize human activities in monitored environments. Among them, one of the most accepted ones is Hidden Markov models (HMM). Activity learning is usually carried out offline, and the current design methodology leads to obtaining a model for each activity of interest. On the other hand, activity recognition should be performed online. Then, if the number of learnt behaviours is high, the amount of computation increases exponentially due to the increase of models to be tested in each time instant, and it might overload the system or provide activity detection speeds far from a real time execution. This problem is increased when, instead of using passive sensors, other devices such as video cameras are used, where the received amount of data per second is much higher. In this paper, it is proposed a new technique to achieve rapid learning and effective activities recognition in real time. Being compared with the baseline technique used with HMM, our proposal is able to improve both the rate of correct activity recognition and the training and detection time complexity, achieving real-time action recognition. The proposed approach is evaluated on two action recognition datasets, one created by our team and another more challenging external dataset including activities of daily life.
Keywords: Human activity recognition, Ambient Assisted Living, motion and tracking, image understanding, learning of action patterns
