Tracking human motion in indoor environments using a distributed-camera system
Qin Cai, J. K. Aggarwal · 1997
Human motion analysis is of growing interest to computer vision researchers, who are motivated by its wide spectrum of potential applications, ranging from athletics, medicine, and surveillance, to new computing technologies such as man-machine interfaces, content-based management of digital libraries, and video conferencing. Surveillance is an important issue in environments such as airports and other restricted-access areas, and the need for computer tracking and monitoring of human activities in such indoor environments is critical. In this dissertation, we address the problem of automatic tracking of human motion in indoor environments from image sequences using a distributed-camera system. The automatic human tracking system presented in this dissertation uses multiple fixed cameras to monitor an indoor area. The tracking paradigm consists of three main modules, namely, single camera tracking, transition tracking across the views of different cameras, and automatic camera switching. The tracking task starts with matching the subject image between successive frames imaged from a single fixed camera. When the current camera is not able to image the subject well, a nearby camera with a better view of the subject takes over the tracking tasks. Thus feature correspondence must be established between synchronized frames captured by different cameras. The automatic camera switching module determines when camera switching is necessary and to which camera the tracking should be switched. We apply a Bayesian classification scheme to track (spatially and temporally) a subject image of interest between consecutive frames. Multivariate normal distributions are employed to model class-conditional densities of the features for tracking, such as location, intensity, and geometric features. Tracking is based upon images of upper human bodies captured from various viewing angles. Non-human moving objects are excluded using Principle Component Analysis of general human shapes. The performance of the tracking algorithm is evaluated by testing of the prototype three-camera human tracking system in various typical indoor scenarios. Performance measures are developed to compare the efficacy of tracking using multiple vs individual tracking features. Experimental results verify the robust performance of the algorithm and the potential of the prototype system for real-time applications.