A Review: Recognizing and Learning Events in Cognitive Vision Systems

This paper reviews several computational frameworks for recognizing and learning events from a video stream. Approaches to event recognition are largely classified into 1) statistical pattern recognition, and 2) symbolic reasoning. In a statistical approach, the types of event patterns that can be recognized are governed by the choice of probabilistic models, and are difficult to generalize. A logic-based framework in symbolic reasoning provides a more general and systematic approach for representing and maintaining a large knowledge base of events. In symbolic reasoning, an aggregate seems to be a natural and useful construct for representing high-level concepts such as object configurations, occurrences and events. Scene interpretation can exploit the taxonomical and compositional relations between aggregate concepts while incorportating visual evidence and contextual information. Primitive occurrences (which provide visual evidence for scene interpretation) are often learned by Vector Quantization (VQ). In VQ, the continuous space of object-level features is discretized into a finite and small number of relevant prototypes. Alternatively, a hidden Markov model (HMM) may be used to detect the coherency in qualitative primitives over a time interval. A commonly observed consecutive sequence of actions can also be learned by HMMs. Other temporal patterns can be expressed by a language that places ordering constraints on either the time-points or the intervals of events. Temporal pattern mining is often realized by a generalto-specific search technique. A logic-based induction (e.g., ILP and AMA-based algorithms) seems to provide a generic solution to the learning of hierarchical event models. These techniques are differentiated by the expressiveness of the languages used for representing the events and the complexity of inductive process.