The Research Of Monocular Vision Odometry Based On Optical Flow

Real-time precise localization is significant to the autonomous navigation. The visual odometry has gradually becoming an important choice for the autonomous navigation. Because the majority of urban road surfaces satisfy the assumption of flat plane, it is reasonable and effective to utilize the monocular vision odometry. Meanwhile, monocular system has a relative lower requirement to the hardware but has a faster calculating speed, which can satisfy the demand of real-time localization. Optical flow method has achieved the requirements of accuracy and stability for the evaluation of image motion, besides due to the fast calculating speed, the visual odometry based on optical flow of feature points method can satisfy the requirement of practical application.In this paper, we research the principles of monocular visual odometry and compare the different feature point’s algorithms principles. Meanwhile, we introduce the design features of monocular visual odometry. The traditional visual odometry is designed on the base of feature points matching as well as robust to the illumination changing and high locating accuracy, whereas it has the disadvantages of low calculating speed and cannot work in real-time.To solve the problem that the processing time of traditional visual odometry is so long that it cannot work in real-time, we integrate the optical flow method into the design of monocular odometry as well as propose the applied principles of monocular odometry based on optical flow methods. The experiment results in indoor and outdoor show the accuracies and speeds of different optical flow methods. This algorithm that we propose improves the speed and accuracy of location in flat plane. However, the accuracy will decline if the illumination is changing. The optical flow of feature points tracking method has the advantage of real-time working but the accuracy is slightly worse.At last, in this paper, we design a monocular visual odometry based on the fusion of optical flow and feature points matching. This algorithm integrates the optical flow method and traditional feature points matching algorithm with kalman filter. Optical flow method can have a more locating accuracy of small motion, but the accumulated error will increase with the increase of motion distance. The feature points matching algorithm has a high locating accuracy, but the frequency is low. Therefore these two algorithms can become complement to each other. Comparing with the traditional monocular visual odometry, the algorithm based on the fusion of two methods has a low requirement of the hardware and a fast processing speed as well as great robustness. Meantime, the new algorithm can require intermediate results, which makes the final locating results smoother and integrated.