Based On Feature Points Extracted Monocular Visual Odometer

The realization of a monocular vision odometry is proposed. Deviations brought in the traditional odometry when the surrounding is slippery can be well choked back, and the course information can also be got. All of these benefit the intelligence vehicle finishing the navigating or other vision tasks.The proposed monocular vision odometry needs only one camera. It facilitates the modified SIFT(Scale Invariant Feature Transform) to extract keypoints in the captured image frames. Then Tsai's calibration method is used to mapping the image coordinates and the world coordinates. For the scale invariant and the rotation invariant property of extracted keypoints, accurately matching the keypoints between the series image frames is successfully done and the transformation between two vehicle coordinates has been achieved. This transformation is just the estimations of the distance and rotation angles of the vehicle.The strong points of this monocular vision odometry are that it can be widely adopted in some situations, and comparing the traditional odometry better location result comes out especially on roads with less friction. Not reducing the accuracy, the equipment and realization in this method are much simpler. Of course the cost is less than the binocular vision odometry proposed by others recently. When one camera in binocular vision odometry has troubles, this method can also be a good remedy. The significance of our way for practical realization is the image process basing on relative image frames, not the video, which greatly reduces the time and memory demands.For future design of the monocular vision system, some modifications should be well considered. The main problems of our method are the real time method for exacting keypoints and the robustness of the system.In this paper, we focus on the techniques of monocular vision odometry, and make a detail research on the extracting and matching keypoints in images (Harris corner, SUSAN corner, SIFT keypoints and so on ), camera calibration (Tsai's calibration method, Zhang Zhengyou's calibration method ,self calibration and so on), coordinates transformation(quaternions and so on). We realize these methods by analyzing and programming. At last, the best way for our monocular vision odometry is selected, and the mathematical model is given. Some result details and experiments are carried out in the end for our design test.