Clustering　the　spatial　feature　of　the　object　region　plays　an　important　role　in　object　modeling,　detecting　and　tracking　etc.　However,　many　clustering　methods　adopt　the　pre-set　cluster　number,　which　cannot　adapt　to　full　automatic　system.　In　this　paper,　a　novel　frame　work　for　adaptively　determining　the　number　of　clusters　is　proposed　based　on　hierarchically　kernel　cutting.　We　firstly　extract　the　value　contour　of　object　region　and　rank　the　peaks　of　the　contour　in　descending　order.　And　then　we　utilize　a　group　of　gauss　kernels　located　at　peaks　to　sequentially　segment　the　contour　into　several　subintervals.　When　the　residual　area　being　not　cut　is　lower　than　a　threshold　value,　the　cutting　process　is　compulsively　terminated.　Furthermore,　we　merge　adjacent　kernels　according　to　the　intersection　area　ratio　and　take　the　retained　kernel　number　as　the　cluster　number　k.　We　finally　classify　the　object　region　with　k　and　K-means　algorithm.　Both　theoretical　reasoning　and　experimental　comparing　illustrate　the　proposed　method　is　rational,　adaptive　and　efficient.