23. Context-Based Situation Recognition in Computer Vision Systems

The availability of visual sensors and the increment of their processing capabilities have led to the development of a new generation of multi-camera systems. This increment has also conveyed new expectations and requirements that cannot be fulfilled by applying traditional fusion techniques. The ultimate objective of computer vision systems is to obtain a description of the observed scenario in terms that are both computable and human-readable, which can be seen as a specific form of situation assessment. Particularly, there is a great interest in human activity recognition in several areas such as surveillance and ambient intelligence. Simple activities can be recognized by applying pattern recognition algorithms on sensor data. However, identification of complex activities requires the development of cognitive capabilities close to human understanding. Several recent proposals combine numerical techniques and a symbolic model that represents context-dependent, background and common-sense knowledge relevant to the task. In this chapter the current challenges in the development of vision-based activity recognition systems are described, and how they can be tackled by exploiting formally represented context knowledge. Along with a review of the related literature, we describe an approach with examples in the areas of ambient intelligence and indoor security. The chapter surveys methods for context management in the literature that use symbolic knowledge models to represent and reason with context. Due to their relevance, we will pay special attention to ontology and logic-based models.