The　Augmented　Reality　system　based　on　KAZE　algorithm　is　to　do　nonlinear　diffusion　filtering　by　the　additive　operator　splitting　algorithm.　In　this　way,　the　problem　of　blurred　boundaries　and　detail　missing　can　be　solved.　A　stable　nonlinear　scale　space　is　constructed　by　using　arbitrary　step　to　search　the　Hessian　local　maxi-mum　value　point　after　different　scales　normalizing　to　detect　feature　points.　Meanwhile,　it　combines　KLT　algorithm　when　tracking　the　target　feature　in　order　to　improve　the　efficiency.　Experiment　show　that,　since　the　KAZE　algorithm　matching　feature　points　are　better　and　the　positioning　accuracy　is　higher,　therefore　this　AR　system　has　good　effect　of　real-time　virtual　and　actual　fusion　in　special　environments　such　as　insufficient　light,　occlusion　and　jittering.