Genetic algorithm-based stereo vision with no block-partitioning of input images

Stereo correspondence could be formulated as an optimization problem. Most of the existing solutions, however, adopt the gradient-based approaches, requiring an initialization close to the correct solution. This paper presents an alternative approach, which is genetic algorithm based, that has larger tolerance toward the quality of the initialization. Each candidate for the three-dimensional description of the imaged scene is encoded as an individual that embraces thousands or even millions of chromosomes, and a population of such individuals are allowed to evolve to reach a globally optimal or near-optimal solution. Our solution framework also includes a coarse-to-fine matching strategy to reduce the matching ambiguity and the computations needed. Experimental results on synthetic and real images are shown to illustrate the performance of the approach.