In view of the traditional ICP (Iterative Closest Points) algorithm which cannot satisfy the requirement of SLAM (Simultaneous Localization and Mapping) algorithm accuracy in dynamic environment,an improved ICP algorithm based on the fusion of feature point structure similarity judgment is proposed;By introducing triangular structure constraints in the feature points set, the dynamic matching points and false matching points in the two groups of point sets were eliminated, and the accuracy of ORB feature points matching was improved. Compared with traditional SLAM algorithm, the estimation of camera pose is more accurate. Through the simulation experiment in Linux system, the ICP algorithm combined with geometric constraints of feature points can effectively solve the impact of dynamic objects on camera pose estimation, and improve the positioning accuracy of RGB-D SLAM in dynamic scene.