【作者】 薛陈；

【导师】 朱明；

【作者基本信息】 中国科学院研究生院（长春光学精密机械与物理研究所） ， 机械电子工程， 2010， 博士

【摘要】 视频目标跟踪是计算机视觉领域的关键技术之一,在民用和军事的诸多领域中都具有极为广阔的应用前景,包括智能监控、基于视觉的人机交互、智能交通、机器人视觉导航、精确制导系统等。随着信息技术的飞速发展,越来越多的研究人员投身于视频目标跟踪的研究领域,提出了很多优秀的算法,这些算法在某些特定的场合下取得了良好的效果。尽管如此,研究出一套鲁棒的,能适应各种复杂环境(如复杂的背景、目标外观的变化和目标的遮挡等)的跟踪算法并且予以工程实现,仍然存在很多困难。根据工程需要,本文主要对复杂环境下的视频目标跟踪技术进行了算法的研究和工程应用的研究,主要工作如下:(1)深入研究了均值漂移算法,针对传统均值漂移算法中由于背景像素造成的目标定位偏差的问题,提出了基于最优灰度直方图特征的改进均值漂移算法。改进的算法根据初始帧目标和背景在灰度分布上的差异,建立对数似然图(log-likelihood image),筛选出与背景可区分性好的最优灰度直方图特征进行目标建模,并且以同样的方法在后续帧建立候选模型。改进后的算法能够有效减轻背景像素对目标定位的影响,提高目标的跟踪精度,同时减少算法的迭代次数,提高算法的运算速度。(2)提出了质心加权算法并予以改进。该算法通过求跟踪区域内同一灰度级所有像素的质心位置的数学期望获得目标的最终位置。除了目标的灰度统计信息外,算法还包含了灰度分布的空间信息,目标特征描述更加丰富,目标定位更加准确;算法只需一步计算,无需迭代,实时性好;较改进的均值漂移算法,改进的质心加权算法对目标遮挡有更好的鲁棒性。同时提出了目标跟踪过程中的模板更新策略,增强了算法的鲁棒性。(3)针对目标的遮挡问题进行了深入研究,提出了完整的目标跟踪算法。首先提出了基于边缘加权的Bhattacharyya系数;该系数对目标的遮挡非常敏感,能有效判断出目标的遮挡时刻。算法以卡尔曼滤波为基本的跟踪框架。遭遇遮挡时,根据遮挡程度的不同,采取不同的处理策略。对于部分遮挡,不做特殊处理,本文提出的改进的质心加权算法完全能够克服部分遮挡;对于严重遮挡,采取基于分块的目标跟踪算法;全部遮挡情况下,采用卡尔曼滤波的预测值作为目标的位置,同时停止对卡尔曼滤波的修正。算法对目标的部分遮挡、严重遮挡和完全遮挡都有很好的鲁棒性。(4)研制了以DSP+FPGA为基本架构的电视跟踪器。硬件平台中,解决了低温工作问题、FPGA配置问题和电磁兼容问题;软件系统中,完成了DSP程序的编写和优化,实现了对光照变化鲁棒的相关算法的DSP移植。该电视跟踪器已经通过了环境测试,工作可靠稳定,跟踪效果好,实时性强,满足各项指标和要求,现已应用于实际工程项目当中。更多还原

【Abstract】 As one of the crucial issues of computer vision, visual object tracking is widely used in many applications, such as visual surveillance, human-computer interaction, visual transportation, visual navigation of robots, military guidance, etc. Along with the rapid development of information techniques, more and more researchers devoted themselves to the research area of visual object tracking, and many effective algorithms have been proposed, some of which have great performance under certain environment. However, there are still many difficulties to the research and application of robust algorithm, due to the complex environment, such as complex background, change of the appearance, and occlusion, etc.This dissertation includes two parts: firstly, the research of robust tracking algorithm under complex environment, and secondly, the realization of tracking algorithm based on hardware platform. The main contributions of the dissertation are summarized as follows:(1) Tracking algorithm based on Mean-shift is deeply discussed. Due to the background pixels, the traditional Mean-shift algorithm can not locate the object exactly. Improved Mean-shift algorithm based on the most discriminative grey level features is proposed. According to the difference of grey distribution between the object and the background in the initial frame, log-likelihood image is set up to select the discriminative grey level features for object modeling. The candidate modeling is done the same way in the next frames. The improved Mean-shift algorithm may not only reduce the impact of the background pixels to object localization and increase the precision of localization, but also reduce the iteration times of the algorithm, and increase the speed of computation.(2) Centroid weighted algorithm is proposed and improved in this dissertation. The ultimate location of the object is the expectation of the centroids of the pixels of the same grey level in the tracking area. The centroid weighted algorithm has three advantages. Firstly, the algorithm includes spatial information of the color distribution besides the statistical information, which makes it more precise. Secondly, it is very simple and needs only one step computation without iterations, which makes it very suitable for real-time application. Thirdly, rather than the improved Mean-shift algorithm, the improved centroid weighted algorithm is more robust, when partial occlusion happens. On the other hand, the model updating strategy is proposed, which makes the tracking algorithm more robust.(3)Occlusion problem is deeply discussed in the dissertation and integrated algorithm of object tracking is proposed. Bhattacharyya coefficient is proposed, which is very sensitive to occlusion. Kalman filter is the main framework of tracking. According to the degree of occlusion, different strategies are used. The proposed centroid weighted algorithm is robust to partial occlusion, so no special treatment is needed to partial occlusion. Fragments based algorithm is used when serious occlusion happens. When totally occluded, the predicted location of the Kalman filter is chosen as the object location. The strategy of dealing with occlusion is robust to partial occlusion, serious occlusion and total occlusion.(4)Video tracker is developed based on the framework of“DSP+FPGA”. Three problems of the hardware are resolved, and they are object tracking in low temperature, the configuration of FPGA and EMC of the hardware. The code of DSP is written and optimized, and robust tracking under variable light condition is realized. The video tracker has now already passed the environment test, and the performance of both hardware and software all meet with the requirements, such as stability, reliability and real-time, etc.更多还原