Unsupervised Spatio - Temporal Activity Learning and Recognition in a Stream Processing Framework

Learning to recognize and predict common activities, performed by objects and observed by sensors, is an important and challenging problem related both to artificial intelligence and robotics. In this thesis, the general problem of dynamic adaptive situation awareness is considered and we argue for the need for an online bottom-up approach. A candidate for a bottom layer is proposed, which we consider to be capable of future extensions that can bring us closer towards the goal. We present a novel approach to adaptive activity learning, where a mapping between raw data and primitive activity concepts are learned and continuously improved on-line and unsupervised. The approach takes streams of observations of objects as input and learns a probabilistic representation of both the observed spatio-temporal activities and their causal relations. The dynamics of the activities are modeled using sparse Gaussian processes and their causal relations using probabilistic graphs. The learned model supports both estimating the most likely activity and predicting the most likely future (and past) activities. Methods and ideas from a wide range of previous work are combined to provide a uniform and efficient way to handle a variety of common problems related to learning, classifying and predicting activities. The framework is evaluated both by learning activities in a simulated traffic monitoring application and by learning the flight patterns of an internally developed autonomous quadcopter system. The conclusion is that our framework is capable of learning the observed activities in real-time with good accuracy. We see this work as a step towards unsupervised learning of activities for robotic systems to adapt to new circumstances autonomously and to learn new activities on the fly that can be detected and predicted immediately.