Area Based Alarm System using 3D Cameras

Abstract: Depth map cameras provide new ways of designing surveillance systems. In this thesis we evaluate three different cameras from two different depth sensor techniques, and propose a complete method for detecting thefts over a counter in a retail environment. Our algorithm covers pre-processing with noise reduction and background segmentation using the reflected signals amplitude as a confidence measurement. A plane is fitted both to the 3D points of the top of the retail counter as well as to the 3D points on the side (cashiers side) of the retail counter. The algorithm determines which foreground pixels are on the wrong side of both these planes. By running this result through a few methods to improve rigidity, we show that it is possible to detect thefts with a very high detection rate and low false positive rate. Finally we present the results from our testing of different versions on a database of activities with known ground-truth (theft/no theft).