Observing and Modeling Human Activity

We propose to investigate semi-supervised and unsupervised methods to discover daily routines from their long-term observation of visual behavior. Our method combines a representation of visual behavior that encodes the temporal-spatial action with usercustomization. We further propose different parameters to encode time, length of the action for their use in the activity classification model. Our method will be evaluated on a novel long-term recorded dataset that contains daily recordings of natural visual behavior in home environment. We show the ability of our method to discover activities with performance competitive with that of previously published surveillance methods. We will be able to record long-term daily routines in the Amiqual4Home smart apartment with a network of cameras and embedded sensors. This dataset will be annotated to a ground truth and used to evaluate the proposed method. So far, however, it remains unclear how much information about daily routines is contained in long-term human visual behavior, how this information can be extracted, encoded, and modeled efficiently, and how it can be used for unsupervised discovery of human activities. The goal of this work is to shed lights on these questions. We will further present an approach for unsupervised activity discovery that combines a statistical and/or structural visual behavior representation with unsupervised model.