Aligning Overlapping Image Sequences by the Combination of Feature-Based and Direct Methods

A new method for the spatio-temporal alignment of overlapping image sequences from static cameras is presented in this paper. This method first estimates the initial spatial transformation by registering the background images. Then the temporal transformation is estimated in a feature-based manner, using the Hausdorff distance and linear regression. The parameters of the estimated spatio-temporal transformation are then refined by a direct method to give a more precise result. The combination of both feature-based estimation and direct method can cope with large displacements in the temporal domain and provide high spatial alignment accuracy as well.