Abstract: On the basis of fundamental matrix estimation, computer vision methods are used to reveal the three-dimensional information of an object within a series of scene images captured from different angles and distances. They are the primary solutions for cutting-edge problems such as image-based modeling and simultaneous localization and mapping. Accuracy and efficiency are two major metrics of fundamental matrix estimation methods. When the accuracy is not enough, high-cost back-end optimization is required for the correction, and its low efficiency affects the real-time performance of the system. To solve these problems, an improved quasi-affine transformation method is proposed based on the QUATRE algorithm. Firstly, a specific “gene-chromosome” mode is used for the collaboration of the particles. Besides, the way of initialization, mutation, and crossover of the original QUATRE algorithm are redefined within the discrete solution space described by the homogeneous coordinates. In addition, a confidence coefficient based iteration termination method is presented for the acceleration. Experiments show that the proposed method is useful for fundamental matrix estimation. It can effectively get rid of the disturbance of outliers resulting from the noises and mismatches, and it outperforms the methods such as the LMedS, RANSAC, and MSAC in terms of accuracy and efficiency.

Key words: epipolar geometry, fundamental matrix, quasi-affine transformation, evolutionary strategy, mutation strategy