Abstract
Knowledge of the camera trajectory, which may be determined from the motions between consecutive frames of a video clip, can be used to register images for constructing image mosaics. We discuss a mosaic-based positioning framework for building photo-mosaics and concurrently utilizing them for improved positioning. In this approach, the mosaic is directly exploited in bounding the accumulation of position errors as we integrate the incremental motions of the camera. It is also shown that two earlier closed-form solutions for the estimation of motion directly from spatio-temporal image gradients, as for most gradient-based techniques based on the application of linear(ized) image motion constraint equations, are corrupted with systematic biases. These can be reduced significantly by incorporating the higher-order terms. We propose recursive methods to solve the new nonlinear constraint equations, and investigate the performance of the new solutions in a number of experiments with synthetic and real data.
