The random sample consensus (RANSAC) algorithm is a classical parameter estimation method widely applied in computer vision tasks such as image matching, image stitching, and 3D reconstruction. Hypotheses are first generated by randomly sampling a minimal subset of data points, followed by model fitting and parameter evaluation based on the sampled subset. These steps are iteratively repeated until the stop criterion is satisfied, with the optimal model obtained during the process being selected as the final output. RANSAC demonstrates strong performance when handling datasets containing a single model. However, its fitting speed is influenced by the sampling strategy, and its estimation accuracy is constrained by the underlying model estimator. To advance the development of RANSAC techniques, this study provides a comprehensive review, analysis, and summary of main stream RANSAC variants. The existing methods are classified based on their respective improvement strategies. During hypothesis generation, the sampling quality is enhanced by selecting high-quality points or incorporating geometric constraints. During model refinement, parameter accuracy is improved by aggregating multiple models or integrating local optimization techniques. During hypothesis verification, pre-screening mechanisms are introduced to filter out incorrect hypotheses, thus reducing computational overhead. Numerous RANSAC variants achieve increased computational efficiency and robustness through such modifications. This study details the implementation principles of RANSAC and its variants and conducts both quantitative and qualitative performance evaluations on public datasets to assess their overall effectiveness.