This paper proposes an online competitive learning paradigm, Self-Branching Competitive Learning(SBCL), which uses K-Nearest Neighborhood(KNN) and iterative variance estimation for clustering analysis. SBCL adopts the incremental learning strategy, starts clustering data from one initial prototype and then branches if the bias between vectors is larger than the pre-specified scale. SBCL is unrelated to initial cluster number or data distribution, avoids the dead node problem and suits to analyze the online input data. We apply SBCL to two classical problems: clustering data with mixed Gaussian distributions and segmenting MRI images. The experimental results shew that SBCL has good performance in these problems.