Fast Video Mosaic Construction for Observation of Large Static Scenes

To overcome the problem of registration error accumulation, this paper presents a fast video mosaic approach especially suitable for real-time observation of large static scenes. Different from the traditional video mosaic algorithm based on pair-wise local registration, the proposed new approach can directly compute the projective transform between the current input image and the panoramic image, thus effectively avoiding the accumulation and propagation of registration errors resulted from the concatenation of pair-wise transformations. Moreover, in order to tackle the inconsistent issue of global video mosaics in registering video images from a 2D scan camera, we present a novel direct registration scheme by simultaneously employing feature points from both the temporal adjacent area and the spatial adjacent area of a panoramic image, thereby successfully reducing the complexity and time cost of the general global alignment algorithm. The experimental results show that the proposed algorithm can construct seamless video mosaic quickly both for 1D scan mode and for 2D scan mode of a camera, and that the constructed video mosaic can be extended progressively with input of video frames, to meet the requirement of real-time observation of large static scenes.