【作者】 易科；

【导师】 沙莎；

【作者基本信息】 中南大学 ， 计算机科学与技术， 2011， 硕士

【摘要】 运动目标检测和跟踪一直是计算机视觉领域内研究的热点问题,它在智能监控,武器装备,自主导航车,医学诊断等众多领域都有广泛的应用。而动态场景下运动目标检测和跟踪由于摄像机的运动而造成背景的不断变化,很多问题如目标的准确检测,目标形变跟踪及遮挡问题至今尚未解决。于是,本文将针对动态场景下中所遇到的一些问题进行研究,主要工作可以归纳如下：1针对摄像头的运动而造成的背景不断变化且对不同层次上的物体所造成的运动向量也不同,提出动态场景下基于对极几何约束判别静与动的点的基础上检测运动目标,即前一帧中静态的特征点一定存在于相继帧中对应的极线上,如果远离对应的极线,则判定为动态的点。从而在背景景物差异大情况下依然能够检测出运动目标。2针对传统的SIFT算法中特征描述子计算复杂度高的情况,提出以积分直方图快速构建特征描述子的方法,降低其计算的复杂性,使其能够满足目标检测和跟踪时的实时性。3针对动态场景下中的运动目标经常会发生尺度变化,在分析传统均值漂移算法核带宽固定的基础上,采用改进的自适应尺度变化的均值漂移算法。4动态场景下运动目标会经常遇到静态障碍物出现遮挡或者多个动态目标运动过程中粘贴存在相互遮挡等情况,于是,针对动态目标遇到静态障碍物遮挡时,采用结合卡尔曼滤波估计的均值漂移算法来估计在遇到遮挡情况下目标的位置。并针对在摄像机运动下,两个或多个目标的存在相互遮挡且这些目标形状尺度同时也在变化时,提出结合实时检测的运动目标跟踪算法,将实时的检测结果更新到跟踪算法中,并将跟踪的结果反馈检测的准确性。最后,对本文的研究工作进行了总结,并提出了后续研究工作的思路,对今后的研究具有一定的指导意义。更多还原

【Abstract】 Moving object detection and tracking remain as the hot spot issue in the field of computer vision. They have been widely used in various fields like intelligent monitoring, military hardware, autonomous navigation and medical diagnosis. However, Many questions have not been resolved yet in the dynamic scenes of moving object detection and tracking because of the flux and reflux of the moving camera. Thus the paper conducts an investigation to the encountered problem in dynamic scenes. The main tasks can be summed up as follows:1. To sovle the problem of the constantly changing background caused by the moving camera and the different motion vector caused by objects of different level, We propose geometric constraints method based on SIFT to classify a pixel as moving or static. The constraint, we use, is the multi-view epipolar constraint which requires images of static points to lie on the corresponding epipolar lines in subsequent images. If the point far away from the corresponding epipolar line, then we can set the point as a static point. Thereby, moving objects can be detected even in the case of widely difference of the background scenes.2. Aimed at the case of high computation complexity of feature descriptor in the conventional SIFT algorithm, we propose the method of rapid building feature descriptor using the integral histogram, which is to reduce the computation complexity and satisfy the real-time of object detection and tracing.3. Aimed at the scale changing of the moving objects in the dynamic scene, we propose the improved mean shift algorithm of self-adapting changing scale on the basis of analyzing the fixed kernel bandwidth of conventional mean shift algorithm.4. The moving objects frequently meet the case of static barrier shading or the several dynamic targets pasting and shading with each other in the motor process. Accordingly, aimed at the static barrier shading of dynamic objects, we propose the mean shift algorithm combined with the Kallman filtering to estimate the target location in the case of shading. And aimed at two or more dynamic objects shading with each other and their shape and scale changing in the case of camera moving, we propose the moving objects tracing algorithm combined with the real time detection and update the tracing algorithm of the real time detection result. The tracing result feeds back the detection veracity.At last, the thesis sums up the research work and raises the train of thought of the follow-up research work. It will be have directive significance to the future research.更多还原