Coupling Fall Detection and Tracking in Omnidirectional Cameras

Omnidirectional cameras have many advantages for action and activity detection in indoor scenarios, but computer vision approaches that are developed for conventional cameras require extension and modification to work with such cameras. In this paper we use multiple omnidirectional cameras to observe the inhabitants of a room, and use Hierarchical Hidden Markov Models for detecting falls. To track the people in the room, we extend a generative approach that uses probabilistic occupancy maps to omnidirectional cameras. To speed up computation, we also propose a novel method to approximate the integral image over non-rectangular shapes. The resulting system is tested successfully on a database with severe noise and frame loss conditions.