【作者】 尹宏鹏；

【导师】 柴毅；

【作者基本信息】 重庆大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 随着计算机理论、技术和应用的快速发展,视频图像处理和计算能力得到了极大的提高,使得计算机视觉成为了计算机领域与人工智能领域中最热门的研究课题之一。基于计算机视觉的运动目标跟踪作为计算机视觉领域中的一个重要问题,是研究视频图像序列中运动目标的检测、提取、识别和跟踪,获得运动目标的运动参数,如位置、速度、加速度等,以及目标运动的轨迹,从而进行进一步处理与分析,实现对运动目标的行为理解,以完成更高一级的任务。作为一个有着广泛应用背景的研究领域,基于计算机视觉的目标跟踪吸引了大批研究学者参与,许多国外研究机构也将其列为重要研究方向,并已取得了很多成果。但是一般意义上的跟踪技术还远未成熟,要开发出真正鲁棒、实用的跟踪应用系统还需要解决大量的问题。论文研究了基于图像特征的目标分割、模板匹配、Mean shift算法等目标跟踪问题,在实际的序列图像场景分析中,以航天器发射中运载火箭的起飞和飞行时的轨迹和姿态获取为背景,通过高速摄像机获得的视频图像进行火箭目标的处理和跟踪,针对不同场景下的火箭目标跟踪问题,研究了基于计算机视觉的运动目标跟踪算法。论文对基于计算机视觉的目标跟踪技术的研究现状进行了探讨,讨论了当前基于计算机视觉的目标跟踪技术中目标的表示方法以及目标跟踪特征选择标准;对基于计算机视觉的运动目标跟踪算法进行了分类并指出了各种目标跟踪算法的优缺点。对计算机视觉理论框架进行了探讨,在Marr的计算理论框架下提出了本文的研究思路。重点对基于计算机视觉的火箭目标跟踪场景中的目标及背景进行了深入分析。在此基础上,提出了基于计算机视觉的火箭目标跟踪中存在的难点问题。针对火箭目标的跟踪问题,在对比不同的边缘检测算子对火箭图像序列进行边缘检测的基础上,选择使用Robert边缘检测算子对火箭目标进行边缘检测。根据目标的灰度分布性改进了最大类间差分法,提高了火箭分割的精确度与实时性。提出了一种带方向的非线性滤波方法去除背景边缘的算法,有效的解决了火箭边缘图像中存在的干扰边缘问题。仿真实验结果表明了该算法对火箭目标具有很好的分割效果。针对在火箭目标飞行过程中大小和姿态变化情况下,对火箭目标稳定跟踪的问题,提出了基于多关联模板匹配的模板匹配策略,通过仿射变换根据伸缩比和旋转角度从上帧最优模板中产生出多关联模板以自适应火箭的大小和姿态改变提高算法的匹配精度。采用了卡尔曼滤波对火箭的运动进行轨迹预测,根据预测目标位置确定图像待匹配区域,有效地减小了算法时间复杂度,提高了算法的实时性。仿真实验结果表明了该算法具有很好的匹配精度与实时性,对于目标的状态、大小变化与遮挡现象有较好的鲁棒性。在研究了Mean shift算法与火箭目标飞行特点的基础上,提出了使用Mean shift算法与帧间差分法相结合的火箭目标跟踪算法,对Mean shift算法进行了改进。采用帧间差分法提取火箭目标运动区域,然后在此基础上使用Mean shift算法实现目标的精确跟踪。仿真实验结果表明了该算法能有效地对火箭目标进行跟踪,并能很好的解决跟踪过程中的跟踪误差累积问题。针对在Mean shift目标跟踪算法框架中只使用单一固定图像特征表示火箭目标、不能自适应的根据跟踪场景选取最佳跟踪特征对火箭目标进行表示、选用的特征模板不能随跟踪环境自适应更新而经常造成模板漂移导致跟踪失败的问题,提出了一种在线自适应多特征融合算法和模板自适应更新机制。融合火箭目标的颜色、边缘、纹理特征,对火箭目标进行表示。通过构建前后相邻两帧间的相似度函数,对跟踪模板进行自适应更新。实验证明了该算法对复杂背景下的火箭目标具有较好的跟踪效果。最后,对全文的研究工作进行了总结,并指出了今后工作中进一步研究的方向。更多还原

【Abstract】 With the rapid development of computer theory, technolgy and application, the video image processing capacity and computer’s performance have been greatly improved. It made computer vision became a very popular research task in the computer science and artificial intelligence. Computer vision based moving object tracking is a very important challenge in computer vision field. Object tracking based on the computer vision is a technology of detection, extraction, recognition and tracking moving object from video image sequence and get the moving parameters such as position, velocity, acceleration and object’s moving trajectory. Based on these parameters, with deeply process and analysis to achieve behavior understand and complete advanced task. Object tracking based on computer vision as a widely used technolgoy, has attracted many researchers. Lots of institutions take it as a very important research field and have got many remarkable achievements. But computer vision based object tracking technology is still not a mature technology. There still have a great number of problems which should be solved to develop a robust and practical tracking system.Thesis focuses on feature-based object segmentation technology, template matching technologyand Mean shift object tracking technology. On real sequence image scene analysis, taking the task of acquire rocket’s trajectory and flight regime when rocket launch and fly as research background. It using high-speed video camera directly to trace and locate the rocket image. Meanwhile, it takes deeply research and discussion on rocket tracking based on different scenes.Thesis discusses the current research of computer-vision-based object tracking. The object representation method and tracking features choosing method in the computer-vision-based object tracking filed is also discussed. It classifies the algorithms of moving object tracking and points out the advantages and disadvantages of these algorithms. Thesis introduces and analyzes the theory frame of computer vision. Based on the Marr’s computer vision theory frame, our approach is presented. It focuses on the analysis of rockets and backgrounds features under computer-vision-based rocket tracking scene. Based on that, the difficulties in computer-vision-based rocket tracking are proposed.In order to solve the rocket tracking problem this thesis compares the edge detection results using the different edge detection algorithms and chooses Roberts edge detection operator to detect the rocket edge. Improved OTUS segmentation algorithm is proposed according to distribution of grey scale and that enhances the accuracy and real-time performance of rocket segmentation. To solve the edge interferences in edge images, a nonlinear filter with direction is designed to remove background edges. Simulation has proved the proposed rocket method is effective for rocket segmentation.Thesis proposes a multi-correlation-template match strategy to overcome the flaws of traditional template matching algorithms which cannot trace the targets with long time and stability due to the lack of adaptive template update for the change of rockets size and profile. Through affine transformation, the algorithm would generate multi-template to match the size and profile of the target rocket from optimal template in previous frame according to stretching and rotation. Thus the match accuracy is enhanced. To raise the real-time performance of the algorithm,Kalman filter is used to estimate the motive track of rocket, and the estimation reduces time complexity of the algorithm effectively. Simulation shows that the algorithm is robust to changes of target size and profile as well as block problems with good match accuracy and real-time performance.Thesis proposes an improved Mean shift which is acquired with combination of normal Mean shift algorithm and frame-difference methods based on the study of Mean shift algorithm and traits of rockets flying. By using Frame-difference method to exact the motive area of a rocket, we get its approximate location. Then with Mean shift algorithm, accurate tracing is achieved. Simulations prove the effectiveness of the new algorithm that it is effective to trace a target rocket and better solves the problem of error accumulation during the tracing.Thesis proposes an online multi-features fusion algorithm and adaptive template update mechanism to solve the defect that tracking template cannot be self-adaptive to various scenes which cause template drift, and that single image feature in Mean shift algorithm framework cannot be self-adaptive to trace scene to alter the best trace feature for a target rocket. This new algorithm considers a feature with greatest difference between target rocket and background as the best trace feature. Thus, rocket and the background are classified into two different classes, and a function is established to measure the differences between target rocket and background. To acquire a complete outline, the different features are combined through ratio. By establishing a similarity function of two frames before and after, tracing templates are updated adaptively. Simulations show that this new algorithm is effective in tracing target rocket with complex background.At last, the thesis summarizes all research work discussed above, and points out the further research direction in this field.更多还原