The　unscented　particle　filter　(UPF)　is　well　known　as　a　state　estimation　method　for　nonlinear　system.　However,　UPF　has　the　inherent　drawback　of　costly　calculation.　In　this　paper,　an　adaptive　unscented　particle　filter　by　online　change　the　number　of　particles　is　proposed　to　overcome　the　drawback　of　computational　burden　in　the　traditional　unscented　particle　filter.　Based　on　the　K-L　distance　sampling,　the　new　algorithm　calculates　the　number　of　particles　in　the　next　deviation　according　to　the　predicted　particles　in　the　state　space.　Then　the　computer　simulations　are　performed　to　compare　the　proposed　algorithm　and　other　state　prediction　and　estimation　methods,　such　as　UPF　and　particle　filter.　The　simulation　results　demonstrated　that　the　adaptive　UPF　is　very　efficient　and　smaller　time　consumption　compared　to　traditional　unscented　particle　filter.　Therefore　the　adaptive　UPF　is　more　suitable　to　the　nonlinear　statement　estimation.