Aiming at the problem of path estimation of Kinect camera in unknown scene

a visual odometry based on graph optimization is proposed. Key frame is screened out by optical flow matching of depth images

and the initial pose estimation of the key frame is obtained. Initial pose estimation of the key frame is taken as the vertex and the transformation between the poses as side to form a connected graph model. And the loopback is added to the graph by loopback detection. Non-linear least squares is used to optimize the initial pose in the graph model

and the visual odometer is obtained. The experimental results show that the method can reduce the error effectively on the basis of meeting real-time performance

which plays an important role in the application based on visual odometer.