Large-scale video surveillance systems typically deploy many video cameras over a wide area and display video streams on separate screens. Operators of this kind of systems have difficulties perceiving the whole pictures of intermixed views and are prone to losing track of targets. This paper presents a surveillance system which integrates videos from multiple cameras into a single comprehensive view. Operators can monitor the guarded area from any 3- D viewpoint. For integrated surveillance systems, the accuracy of image registration significantly affects the appearance of stitched views. This paper also presents a new planar image registration method in which the transformation parameters are decomposed from the homography matrix and are optimized according to a similarity measure between corresponding regions. The proposed registration method is pratical and accurate, as demonstrated in our integrated surveillance system.