Electronic digital image stabilization: design and evaluation, with applications

The analysis of long image sequences poses several challenging problems, and despite all the techniques developed to solve them, applications still remain limited to restricted domains. General use of these techniques cannot become a reality without the existence of robust, accurate, and efficient algorithms for image analysis. In this dissertation we focus on the problem of processing long monocular image sequences taken from stationary or moving platforms, for the purpose of motion compensation, or image stabilization. Electronic digital image stabilization (EDIS) is defined as the process of generating a compensated video sequence, where the unwanted components of the camera motion are smoothed or completely removed from the original sequence. The motion of the camera is described by global parametric transformations, which can also be used to align the input image frames together, creating mosaics (panoramic views of the scene). We have designed and developed several techniques for EDIS and mosaicking, based on different motion models and estimation algorithms, particularly those suited for real-time implementation. The performance of an EDIS system is directly related to the transformation model and the technique employed to estimate the motion parameters. For the purpose of evaluation, we propose several performance measures to describe the fidelity, speed, and range of displacements supported by such systems. These measures can also be used as development tools, to determine the influence of certain modules in the overall performance of a system. To demonstrate the efficiency and robustness of these techniques, EDIS and mosaicking algorithms were implemented on a real-time image processing hardware platform, and tested under severe off-road navigation conditions and also with low quality data. Global motion models were also applied in the development of a hybrid video coding scheme, which combines global and local motion compensation to improve the performance of block-based motion compensation codecs.