Mean shift is an effective iterative algorithm widely used in clustering, tracking, segmentation, discontinuity preserving smoothing, filtering, edge detection, and information fusion etc. However, its convergence, a key property of any iterative method, has not been rigorously proved till now. In this paper, the traditional mean shift algorithm is first extended to account for both the local property at different sampling points and the anisotropic property at different directions, then a rigorous convergence proof is provided under these extended conditions. Finally, some approaches to adaptively selecting the algorithm’s parameters are outlined. The results in this paper contribute substantially to the establishment of a sound theoretical foundation for the mean shift algorithm.