【作者】 王晓林；

【导师】 常发亮；

【作者基本信息】 山东大学 ， 模式识别与智能系统， 2009， 硕士

【摘要】 在复杂的自然环境中,背景干扰因素较多,对目标的特征提取与跟踪都相对困难。同时由于摄像机的分辨率和视野范围有限,限制了单摄像机跟踪系统的应用,使得多摄像机跟踪系统的使用成为必然。因此,基于复杂大场景下多摄像机的运动目标接力跟踪成为计算机视觉研究的热点之一,其在视觉自主导航机器人、多智能体协作、可疑目标的连续跟踪、智能交通监控与违章追踪、军事领域的无人驾驶飞机和车辆等实际系统中有着广泛的应用,具有重要的理论价值和实用价值,有着广阔的发展和应用前景。本文主要对复杂大场景下的多摄像机接力目标跟踪问题进行了研究。首先对复杂大场景下多摄像机接力目标跟踪的基本环节——单摄像机在有限复杂场景下的目标跟踪算法进行了深入研究,能够实现复杂场景下的运动目标跟踪;然后提出了基于单PTZ摄像机的主动跟踪算法,使被跟踪目标时刻保持在摄像机视野的中央区域;最后研究了多摄像机下的目标接力跟踪,对相关理论和算法进行了初步研究和探讨。论文的主要工作有:运动目标跟踪方面,研究了基于均值漂移的目标跟踪算法,该算法采用三维背景加权的颜色直方图描述目标特征,对目标旋转变形、部分遮挡具有较好的适应性,并且计算简单、实时性好,但是不能自适应调节跟踪窗大小,在目标快速运动和发生机动的情况下效果不理想。针对均值漂移算法存在的问题,提出了结合均值漂移和卡尔曼滤波的目标跟踪算法。该算法对跟踪过程中的干扰进行检测,根据目标所受干扰的强弱,采用不同的比例因子将卡尔曼滤波预测结果与均值漂移算法得到的结果进行线性加权。同时对被跟踪目标的特征模型进行实时在线更新,提高了跟踪的可靠性。实验结果表明,改进后的算法具有实时性好,鲁棒性高的特点。单摄像机主动跟踪方面,分别对系统的整体结构和各个组成模块进行了介绍,着重介绍了云台转动模块和目标重定位模块。在云台转动模块中,根据PTZ(pan-tilt-zoom)摄像机的运动控制参数及目标在摄像机中的相对位置,建立摄像机运动控制模型,采用云台模糊控制策略使被跟踪目标始终保持在视野的中央区域;为了实现对同一目标的连续跟踪,在云台转动后采用基于去均值归一化模板匹配算法实现目标的重定位。当定位错误时,采用基于运动历史图像的运动目标检测算法在视野的中央区域重新找到相应的运动目标。实验结果表明,该主动跟踪算法无论对于刚性目标还是非刚性目标都能很好地实现主动跟踪。多摄像机接力跟踪方面,针对多摄像机接力跟踪中的一些关键问题与技术进行分析与研究,着重介绍了基于视野分界线的目标交接算法。采用基于同步视频的摄像机视野分界线生成算法,该算法不需要摄像机标定以及环境等信息。当目标可见性判别函数值由负变为正时,表明目标即将要出现在相邻的另一个摄像机的视野范围之内,然后在该视野相关区域内检测出所有运动物体,距离视野分界线距离最近的目标即为被跟踪目标。实验结果表明,基于视野分界线的目标交接算法具有计算简单、准确率高、实时性好的特点。更多还原

【Abstract】 In the natural complex environment, it is difficult to extract the target features and track the target because of the interference factors of the background. Meanwhile, because the resolution and ken of the camera are limited, the applications of single camera tracking system are restricted, and the applications of multi-camera tracking system become inevitable. So the moving target relay tracking based on the multi-camera in large complex scene is a focal point of research in computer vision. The research results can be widely applied in vision autonomous aviation robots, cooperation of multi-agent, suspicious targets successively tracking, intelligence traffic surveillance and vehicles peccancy tracking, and pilotless aircraft and vehicles systems in military fields. Therefore the research results possess significant theoretical value and practical value, and have wide development and application foreground.This thesis mainly aims to study the target relay tracking of multi-camera in large complex scene. Firstly, the target tracking algorithm of single camera in the limited complex scene is further studied. And the research results can realize the moving target tracking in the complex scene. Then, the active tracking algorithm based on the single PTZ (pan-tilt-zoom) camera is proposed which can keep the target in the center area of the camera scene. Finally, the preliminary study about the target relay tracking theories and algorithms by multi-camera in the wide scene is performed. The major contributions of this thesis are summarized as follows:In the moving target tracking part, the tracking algorithm based on mean shift is studied, which uses three-dimensional background-weighted color histogram to characterize the target. This algorithm has good adaptability to target round, transmogrification, and part-occlusion. Furthermore, it requires simple computation and has real-time performance. However, it can not adjust the size of the tracking window adaptively, and has poor performance in tracking fast moving targets and those with maneuverable movement. The tracking algorithm based on mean shift and Kalman filter is proposed to improve the performance of the current mean shift tracker. According to the different interference circumstance, the Kalman filter prediction result and the mean shift tracking result are linear weighted by adopting different scale factors. Meanwhile, the target feature model is updated online to improve the tracking reliability. The experiments show that the improved algorithm is real-time and robust.In the active tracking of single camera part, whole structure of the system as well as every component module is separately introduced. The pan-tilt control module and the target relocation module are emphatically introduced. In the pan-tilt control module, the motion control model of the camera is established according to the camera’s motion control parameters and the relative position of the target in the camera. The pan-tilt fuzzy control strategy is applied to keep the target in the center area of the scene. In order to realize the continuous tracking of the same target, the template matching algorithm by normalized cross-correlation is applied after the movement of the pan-tilt. When the location error occurs, the moving target detection algorithm based on the motion history image is applied to find the corresponding target again in the center area of the scene. The experiments show that the algorithm can track both the rigid targets and non-rigid ones well.In the multi-camera relay tracking part, some key problems and techniques of the multi-camera relay tracking are studied. The target handoff algorithm based on the FOV lines is emphatically introduced. The camera FOV lines generation algorithm based on the synchronous video is applied and the algorithm does not need the camera calibration and environment information, etc. When the value of the target visibility discrimination function changes from negative to positive, it means the target will soon present in the scene of the adjacent camera, then detects all the moving objects in the relative area and the tracking target is the one which is nearest to the FOV lines. The experiments show that the algorithm based on the FOV lines is simple, accurate and real-time.更多还原