Key Technologies Of The Three-dimensional Visual Odometer

Recently the performance improvements in both sensors and computing hardware have made real-time vision processing possible, and as computer vision algorithms mature, we may expect to see more of visually based navigation systems. One of the most important applications in computer vision is autonomous navigation of vehicles and robots. In order to realize autonomous navigation, robot must be able to locate itself during its moving, which is the basis of visual simultaneous localization and mapping (VSLAM).This thesis aims at the study of visual odometry for navigation of vehicle.Firstly, this dissertation fully summarized the basic theory of computer vision and stereo vision .Then the state of the art of visual odometry has been discussed in detail.Secondly, Extracting and matching of image features are studied thoroughly. Harris corner detect is discussed first, then through introducing the theory of scale space, we mainly study SIFT (Scale-Invariant Feature Transform) features. SIFT features are invariant to image scaling and rotation, and partially invariant to change in illumination and 3D camera viewpoint. In addition, the features are highly distinctive, which allows a single feature to be correctly matched with high probability against a large database of features. All these features make SIFT provide good result both in matching and reconstruction. The extracting and matching of SIFT is showed as the test result, in addition to it ,the comparison between Harris and SIFT matching as the scale changes is also showed.In the third part, a 3D space based stereo visual odometry algorithm which aims at solving problems such as wheel slipping of relative odometry localization is proposed, and the detail realization is discussed. With the matching features, the algorithm uses SVD (signal value decomposition) to solve the motion parameters, and obtains better motion parameters under the model of RANSAC. In order to get more accurate parameters, non-linear least square is used.Finally, this dissertation proposed a novel method for visual odometry, which is based on disparity space. The algorithm can overcome the limitation of the anisotropic noise distribution in 3D space. Utilizing the feature that the disparity space has isotropic noise distribution to carry out motion estimation, and through iterative the energy function can achieve global minimize. Experimental results show that this method succeeds in motion estimation and has more accurate result than traditional algorithm.