In conventional visual odometry (VO) systems, perspective-n-point (PnP) method and random sample consensus (RANSAC) algorithm are generally used to estimate camera poses. However, heavy computational burden is incurred, and the pose estimations are not reliable as well. Therefore, in this paper, an improved VO system is proposed, where an off-line camera calibration method is used to obtain lesser measurement errors of image features. Moreover, an improved approach of P3P algorithm is proposed to increase the efficiency of the VO system. To validate the performances of the proposed approach, several experiments are conducted based on a Kinect sensor, where accuracy of pose estimations and runtime efficiency are both improved in comparison to the conventional VO algorithms.