Abstract: This paper proposes an RGB-D relocalization method based on ellipsoidal semantic objects.Firstly,in the process of camera tracking,this paper initializes the observation object in a single frame through combining the depth point cloud,and describes the observed map object by an ellipsoid.An object-level semantic map represented by an ellipsoid is constructed by associating map objects and the corresponding target detection results by using an object common view.Then,the camera is relocated based on the location of the map object.Finally,Iterative Closest Point (ICP) cloud registration algorithm is used to further optimize the camera pose.The experimental results on the OR10 data set show that the relocation method achieves a high relocation success rate in a large parallax environment where the bag-of-words (BoW) and random ferns (FERNS) methods perform poorly.In addition,the proposed algorithm has a similar running time to the above two methods.