Target Recognition And Location Based On Binocular Stereo Vision

Computer Vision is a simulation to biological vision using computer and related equipment. Binocular Stereo Vision can be a good simulation of human visual capabilities, and can be carried out on the three-dimensional perception of the world, as a result, it is widely used in production and life. So there is a great significance to research it. This dissertation completed research on target recognition and location based on binocular stereo vision. Using GRANDAR robot as experiment platform, appropriate algorithm was selected, and some of them had been improved according to actual situation, at the same time, new concept and idea were integrated into it. At last, a specific Binocular Stereo Vision System was designed and realized, and it was used for target recognition and location. From the experiment result, we can see this system is accurate, stable and real-time. The work of this dissertation can be summarized as below:First, in the process of camera calibration, the internal parameters are obtained using Zhang Zhengyou's method. For external parameters, a method of external parameters estimation based on head angle was studied. This method successfully resolved the re-calibration problem when the camera rotates. The result show that external parameters can be estimated accurately using this method.Second, in the part of image segmentation, threshold segment algorithm in HSV space was selected, at the same time, connected component labeling method was inducted to remove the noise. Shape Contexts algorithm was used for target recognition, and PSCE concept was proposed based on it to improve the recognition effect.Third, general three-dimensional reconstruction approach was researched based on the three-dimensional reconstruction modal. The result show that this method is also accurate.At last, actual experiment scene was designed and realized based on the GRANDAR robot platform. According to the experiment scene, a few creative ideas was proposed, for example, three-dimension distance is obtained using the method that combined average multiple measurement result and C average cluster, the robot close to the target by diatance fine tuning, in the grab section, the part of arm was carried out mathematical modeling. These methods reduced the experiment error and improved the success rate.