ZATLAB: recognizing gestures for artistic performance interaction

Most artistic performances rely on human gestures, ultimately resulting in an elaborate interaction between the performer and the audience. Humans, even without any kind of formal analysis background in music, dance or gesture are typically able to extract, almost unconsciously, a great amount of relevant information from a gesture. In fact, a gesture contains so much information, why not use it to further enhance a performance? Gestures and expressive communication are intrinsically connected, and being intimately attached to our own daily existence, both have a central position in our (nowadays) technological society. However, the use of technology to understand gestures is still somehow vaguely explored, it has moved beyond its first steps but the way towards systems fully capable of analyzing gestures is still long and difficult (Volpe, 2005). Probably because, if on one hand, the recognition of gestures is somehow a trivial task for humans, on the other hand, the endeavor of translating gestures to the virtual world, with a digital encoding is a difficult and illdefined task. It is necessary to somehow bridge this gap, stimulating a constructive interaction between gestures and technology, culture and science, performance and communication. Opening thus, new and unexplored frontiers in the design of a novel generation of multimodal interactive systems. This work proposes an interactive, real time, gesture recognition framework called the Zatlab System (ZtS). This framework is flexible and extensible. Thus, it is in permanent evolution, keeping up with the different technologies and algorithms that emerge at a fast pace nowadays. The basis of the proposed approach is to partition a temporal stream of captured movement into perceptually motivated descriptive features and transmit them for further processing in Machine Learning algorithms. The framework described will take the view that perception primarily depends on the previous knowledge or learning. Just like humans do, the framework will have to learn gestures and their main features so that later it can identify them. It is however planned to be flexible enough to allow learning gestures on the fly. This dissertation also presents a qualitative and quantitative experimental validation of the framework. The qualitative analysis provides the results concerning the users acceptability of the framework. The quantitative validation provides the results about the gesture recognizing algorithms. The use of Machine Learning algorithms in these tasks allows the achievement of final results that compare or outperform typical and state-of-the-art systems. In addition, there are also presented two artistic implementations of the framework, thus assessing its usability amongst the artistic performance domain. Although a specific implementation of the proposed framework is presented in this dissertation and made available as open source software, the proposed approach is flexible enough to be used in other case scenarios, paving the way to applications that can benefit not only the performative arts domain, but also, probably in the near future, helping other types of communication, such as the gestural sign language for the hearing impaired.