Appearance and motion information based human activity recognition

Activity recognition is an essential objective of a smart building system which responds to what is happening in a scene. In this paper, a view invariant activity recognition system is proposed to recognise human actions. Selection of applicable features is made and solutions are proposed to deal with probable challenges including differing views on actions and directionality issues. This paper explores a number of features that can be utilised in action recognition systems and chooses suitable features to mitigate the challenges properly. Motion History Image (MHI) based on historical appearance information is used in combination with local motion vectors which are computed through each iteration sequence of the MHI information using an optical flow algorithm. A multi-view dataset (MuHaVi) and a single view dataset (Weizmann) are used to demonstrate and validate the proposed method. Our method, can detect a wide range of actions in multi-view scenarios and shows competitive performance in comparison with state-of-the-art action classification techniques.