【作者】 谢妤婵；

【导师】 林玉池；

【作者基本信息】 天津大学 ， 测试计量技术及仪器， 2009， 博士

【摘要】 目标识别与三维定位技术在先进制造、武器制导、航空航天等领域都具有重要意义和实用价值。针对视觉引导车,研究其目标识别与定位技术。采用从目标物提取直线特征匹配目标模型的方案来确定目标位置与姿态信息。充分利用高级视觉原理与图像处理技术,研究具有智能化程度高、系统组成简单、定位精度高的目标识别与定位方法。由于摄像机成像受光照噪声影响严重,不适合直接采用Marr的三维重建思想进行识别。构建Zernike矩特征与BP神经网络相结合的自动目标识别系统,其实现过程充分考虑系统的实时性、准确性。拍摄目标二维图像,将目标图像进行旋转、尺度变化建立不同目标模型图像库,采用BP神经网络与Zernike矩结合的目标识别方法具有较好的识别效果。直线特征在人工目标识别中较为常见,而在图像处理中由于光照、遮挡等原因常出现线段断裂、部分线段像素缺失等现象。深入研究了直线特征提取方法,提出一种应用Hough空间峰值点逆变换提取直线段特征的新方法。视觉引导车在地面做二维运动,目标图像前后幀变化不大，因此采用直线特征配准前后幀目标图像。此方法具有较强的鲁棒性、便于实现自动化配准的特点。直线提取方法对量化误差、线段断裂具有较强的鲁棒性,可以在干扰较为强烈的情况下提取线段特征,具有较高的配准精度。研究了视觉引导车地面运动的实际情况,在已知目标模型精确几何特征的基础上,提出一种结合地面约束,仅需提取目标一条直线段特征匹配目标模型即可三维定位目标物的单目视觉方法。此方法具有较高的定位精度且实现目标定位的可视化。研究了双目立体视觉测量系统,提出一种基于直线特征的双目视觉定位方法。在Hough空间确定的角度范围内搜索目标线段,根据左右像面线段中点位置关系匹配不同像面上线段信息,建立线段之间的位置对应关系。采用所提取模型线段交点而并非直接采用直线特征作为双目视觉的匹配特征,重建特征线段,快速提取目标物的三维位姿信息。更多还原

【Abstract】 Recognition of target and determination its pose in 3D space play an important role in advanced manufacturing, homing weapon and spaceflight area. For vision guide vehicle, the technology of target recognization and pose determination is studied.A aim body’s position and pose determination project is brought forward, in which straight line character is extracted and then is used to match object. The scheme takes full advantage of high level vision theory and image processing technology. The research work is focus mainly on studying object recognition and pose determination method with high intelligence, simple structure, high measuring precision.Because the camera imaging was greatly affected by light and noise, it is not suitable for recognition with Marr’s three-dimensional reconstruction theory directly. An automatic project recognition system is proposed which conbine Zernike moments and BP net work. During the course of target recognition, its real-time quality and accuracy is taken fully account. CCD sensor shot target, then build up model library with images with rotation and scale change to different target. The recognition method of combination BP neural network with Zernike moments target has better recognition results.Straight line character is commonly used in artificial object recognition, however, illumination and occlusion often lead to straight line breaking and part lossing of pixels during the course of image processing. Straight line extraction method is deeply studied in the thesis, and a new straight line segment extraction method is proposed in which peaks in Hough space is used to extract straight line segment by reverse Hough transform. As vision-guided vehicle moved on the ground plane with two-dimensional motion, object’s images change little between former and its following image, so a target registration method with one straight line is proposed with strong robust properties and it is easy to realize automate registration. The straight line extraction method is robust to discretization error, segment burst, and can extract straight line with strong disturb with higher registration precision.The real factor of vision guide vehicle moving on the ground plane was studied.On the base of accurate geometry model known before, a new pose determination scheme is presented, in which target body’s three-dimensional positioning is determined by match a target’s straight line and that of its model’s. The method possesses higher pose determination precision, achieving target position’s visualization.Binocular stereo vision measurement system is studied.A sort of stereo vision positioning method based on straight line is brought forward including the follow steps:hunting peaks in special angle range in Hough space to extract straight lines,matching straight lines between left and right image planes according to their straight line’s middle points corrdinates relations.It takes the intersections of the straight line segments but not the straight lines themselves as characters to reconstructed straight lines in 3D space by the stereo vision.This method can get project’s pose robustly and quickly.更多还原