Learning spatial event models from multiple-camera perspectives

Intelligent, interactive environments promise to drastically change our everyday lives by connecting computation to the ordinary, human-level events happening in the real world. This paper describes a new model for tracking people in a room through a multi-camera vision system that learns to combine event predictions from multiple video streams. The system is intended to locate and track people in the room, determine their postures, and obtain images of their faces and upper bodies suitable for use during teleconferencing. This paper describes the design and architecture of the vision system and its use in Hal, the most recently constructed interactive space in the authors´ Intelligent Room project