Robust motion estimation for overlapping images via genetic algorithm

We propose a robust method based on genetic algorithm for the estimation of the motion between two successive overlapping images, a classic problem in computer vision. To calculate the motion parameters encoded as a chromosome, we employed roulette wheel selection and total arithmetic crossover and developed a novel adaptive mutation operator. The experimental results show that the normalized registration error of the final solution exhibits a significant improvement over those obtained by direct search approaches to such problems. Also, in contrast to other popular approaches such as the least-squares and Levenberg-Marquardt algorithm, the proposed method can escape from local extrema and can potentially produce the global optimum.