A NEW TECHNIQUE FOR LINKING PERSON TRAJECTORIES IN SURVEILLANCE CAMERA NETWORK

Abstract

Indoor surveillance systems using camera network have been widely used nowadays. In order to uniquely monitor and link trajectories of moving person in camera network, stereo calibration techniques are normally deployed. However, they often result to the drawback of certain drifts in trajectory linking or computational complexity, especially when the overlapping regions of cameras are limited. In this paper, we propose a new method for calibrating multiple cameras with high accuracy. This method can be utilized when the region of overlapping camera fields of view (FOVs) is small or even in disjointed camera FOVs. We observe a ground-plane (or floor) that is shared between cameras. A 3-D coordinate system of ground-plane is defined and calculated from 2-D coordinates on the camera images. By matching the grid points on the ground-plane that are viewed simultaneously by cameras, a homography matrix is figured out in order to determine the translation and rotation between cameras. The proposed method is evaluated on the dataset of indoor-user localization service. The dataset is built with different scenarios, such as multiple moving users at the same time and with/without occlusions. The experimental results demonstrate the accuracy of the proposed method for linking trajectories of the moving person(s) in a camera network.