In this work,in order to improve spatial recognition abilities for the long-term operation tasks of the assistant robot for the elderly,a novel approach of semantic region estimation is proposed.We define a novel graphbased semantic region descriptions,which are estimated in a dynamically fashion.We propose a two-level update algorithm,namely,Symbols update level and Regions update level.The algorithm firstly adopts particle filter to update weights of the symbols,and then use the Viterbi algorithm to estimate the region the robot stays in based on those weights,optimally.Experimental results demonstrate that our proposed approach can solve problems of the long-term operation and kidnapped robot problem.