| Vision-based navigation and 3-D reconstruction, which have great value in research and application, are the key areas of the computer vision research. The traditional way to visual navigation and 3-D reconstruction is based on stereo vision, which uses parallax information to recover the depth information from the 2-D images captured in a scene by two or more cameras. The whole scene is reconstructed, which then can be used to interpret the scene or navigate through it. This dissertation mainly discusses singular camera based obstacle detection and 3-D reconstruction, which uses motion parallax to realize 3-D reconstruction or obstacle detection in visual navigation by recovering the depth information or relative depth information from the scene.To operate autonomous mobile robot safely, obstacles must be detected before any path planning and obstacle avoidance is undertaken. However, instead of reconstructing a complete 3D representation, many biological vision systems possess some "goal-oriented" mechanisms that provide fast and reliable knowledge for a specific task. The obstacle avoidance via time-to-contact is one of these mechanisms. Time-to-contact provides vital information for visual navigation with strong biological evidence. In this dissertation, two methods based on different assumptions for estimating the time-to-contact for obstacle are presented. One is based on the assumption that the robot is moving on a locally planar ground, parts of which are visible in the images captured from monocular camera. Feature based optical flow computation is applied and the homography for the planar ground is robustly estimated. Then time-to-contact for the obstacles is recovered and obstacle map for control system is constructed. This method avoids the computation of high order derivatives of the optical flow, which are very sensitive to noise. Indoor and outdoor experiments of this method show promising results; the other proposes an algorithm to detect the obstacles in outdoor unstructured environment with a single camera and linear method. It makes use of motion cues in the video streams. Firstly, optical flow is calculated at feature points. Then rotation of the camera and FOE(focal of expansion) are evaluated separately. Rotation and FOE value are refined according to an iterative linear method. Finally, inverse TTC(time to contact) with rotation and FOE is recovered and obstacles in the scene are detected. The algorithm is suitable for outdoor unstructured environment. Experiment results show that the algorithm works on different kinds of terrains.3-D reconstruction is of great value in research and application, which can be used in many areas, such as 3-D terrain reconstruction, visual navigation, 3D medical image analysis and virtual reality. This dissertation presents a high-precision 3D reconstruction method based on single camera. A planar pattern was placed in the scene to be 3D reconstructed. Two or more pictures were taken by a single camera which intrinsic parameters were calibrated beforehand. And the extrinsic parameters were calculated accurately using the planar pattern. Standard stereo steps such as stereo calibration, rectification, matching and triangular were performed. Experimental results show that the proposed method was simple, flexible and effective. |
