"In the computer vision field, methods used for 3D reconstruction can be applied also for image understanding, pose estimation, visual tracking, robot navigation, camera calibration, visual measurements, among others. Due to the importance of its applications, this thesis covers three different problems that are closely related with the sparse 3D reconstruction pipeline. Thus, the work presented in this dissertation covers the following problems: i) Estimating geometric relations between two different views of the same scene; ii) Detecting image vanishing points; and iii) Extracting circular markers from digital images. Since these problems can be visualized as modeling estimations commonly formalized as optimization problems, traditional optimization techniques are generally used. These are based on the gradient or the Hessian of the cost function such as Gauss-Newton, Levenberg-Marquardt or Barzilai-Borwein methods. However, when a considerable number of unwanted abnormal data is present, these methods might fail. Other solution methods relies on accumulator space techniques like the Hough Transform (HT), while others employ a heuristic approach such as the Random Sample Consensus algorithm (RANSAC). However, HT-like solutions are slow, whereas RANSAC-like methods are not optimal. To propose a different solution technique, in this work we explore the utilization of metaheuristics, such as evolutionary and swarm-based algorithms. Therefore, the solutions presented in this dissertation require less computational cost in comparison with HT methods and perform better than RANSAC-based solutions. Under the proposed mecha\-nism, new candidate solutions are iteratively built by considering the quality of models that have been generated by previous candidate solutions, rather than relying over a pure random selection as it is the case with classic RANSAC. Further, our solutions explore the search space optimally requiring less computational cost than HT methods, and at the same time having the capability of escape local optima differently from traditional optimization methods. As a result, our metaheuristc-based algorithms present a nice balance between accuracy and computational time. To validate the efficacy of the proposed approaches, several tests and a comparison with other techniques were carried out."