This study describes a posture classification method for a marker-free depth camera. The method consists of an object identification procedure, feature extraction, and a naıve Bayesian classification approach with a supervised training. Point clouds obtained from the depth camera are split into objects. For each object a set of features is extracted. A method of feature pre-processing is proposed and compared against a statistical orthogonalisation method. Using a manually labelled training data set, the probability distributions for the Bayesian classification are obtained. As a result of the classification, the most likely gesture is assigned to each object in real time. Classification performance was tested on a separate data set and reached about 80%.

Three different applications are described: Automatic estimation of user postures to estimate the influence of hearing devices on user behaviour in communication situations, the control of an interactive audio-visual art installation, and interactive light control on a dance-floor setup with multiple dancers. Classification performance in these applications was measured and discussed.