Batch implementations of standard one class support vector machine (SVM) are inefficient on an on-line setting because they must be retrained from scratch every time when the training set is incremental modified. To solve this problem, a detailed incremental one class SVM algorithm is given, and the feasibility and the finite convergence of the algorithm are proven through theoretical analysis. It is ensured that each adjustment step in the C&P algorithm is reliable, and the algorithm will converge to the optimal solution within finite steps. The experimental results on benchmark datasets verify the correctness of theoretical analysis.