The aim of this Graduation Project is to develop a generic biological inspired activity recognition system for the use in intelligent spaces. Intelligent spaces form the context for this project. The goal is to develop a working prototype that can learn and recognize human activities from a limited training set in all kinds of spaces and situations. For testing purposes, the office environment is chosen as subject for the intelligent space. The purpose of the intelligent space, in this case the office, is left out of the scope of the project. The scope is limited to the perceptive system of the intelligent space. The notion is that the prototype should not be bound to a specific space, but it should be a generic perceptive system able to cope in any given space within the build environment. The fact that no space is the same, developing a prototype without any domain knowledge in which it can learn and recognize activities, is the main challenge of this project. In al layers of the prototype, the data processing is kept as abstract and low level as possible to keep it as generic as possible. This is done by using local features, scale invariant descriptors and by using hidden Markov models for pattern recognition. The novel approach of the prototype is that it combines structure as well as motion features in one system making it able to train and recognize a variety of activities in a variety of situations. From rhythmic expressive actions with a simple cyclic pattern to activities where the movement is subtle and complex like typing and reading, can all be trained and recognized. The prototype has been tested on two very different data sets. The first set in which the videos are shot in a controlled environment in which simple actions were performed. The second set in which videos are shot in a normal office where daily office activities are captured and categorized afterwards. The prototype has given some promising results proving it can cope with very different spaces, actions and activities.
|Date of Award||31 Jul 2012|
|Supervisor||Bauke de Vries (Supervisor 1), O.D. Amft (Supervisor 2) & A.J. Jessurun (Supervisor 2)|