复杂场景下基于自适应分块的多目标跟踪方法研究

发布时间：2018-04-21 11:10

本文选题：多目标跟踪 + 自适应分块　；参考：《山东大学》2017年硕士论文

【摘要】：随着计算机硬件和多媒体技术的发展,以及各国政府和民众对安防的高度重视,智能视频监控的应用变得越来越广泛,而多目标跟踪技术作为智能视频监控领域最基本的核心技术,具有重要的研究意义和广阔的应用前景,受到来自世界各地的学术界和工业界科研人员的普遍关注和研究。目前,多目标跟踪技术的研究取得了长足进步,但仍存在许多难题需要解决,如复杂的跟踪场景、非刚体目标的姿态变化、目标遮挡以及跟踪的实时性等。本文针对复杂场景下存在目标遮挡、表观变化以及相似目标的问题,对多目标跟踪进行了研究,主要研究内容及成果为:(1)介绍了多目标跟踪的基本理论。对贝叶斯理论框架下的卡尔曼滤波和粒子滤波的基本原理做了简单介绍,并分析了算法的优缺点。介绍了均值漂移算法和模糊C均值算法的基本原理,并研究了算法的基本步骤。(2)在对多个目标进行跟踪过程中经常存在遮挡、相似目标的情况,为此研究了一种基于自适应分块的粒子滤波多目标跟踪方法。该方法根据目标的灰度分布进行自适应分块,提高遮挡情况下准确跟踪多目标的能力;在粒子滤波跟踪时,利用均值漂移和模糊C均值聚类获取每个目标对应的粒子群,得到目标最优状态估计;引入加权Bhattacharyya距离计算子块的匹配度,考虑了子块可靠性对粒子权重的影响。(3)为了解决多目标跟踪过程中还经常存在的相似目标相互遮挡以及目标表观变化问题,提出了一种基于自适应分块的多特征融合多目标跟踪方法。该方法在上一方法的基础上加入了多特征融合策略,融合颜色直方图和HOG特征对目标进行描述;在粒子滤波跟踪时,依据子块可靠性以及粒子的空间分布及时调整目标模型中子块的权重;并且为减少过程中目标变化对跟踪结果的影响,采取权重更新方法动态更新目标特征模型。实验结果表明,该方法在多目标跟踪过程中存在表观变化、目标相似以及目标遮挡或者相似目标相互遮挡的复杂情况下,均能准确鲁棒地跟踪多个目标。
[Abstract]:With the development of computer hardware and multimedia technology, as well as the governments and people of various countries attach great importance to security, the application of intelligent video surveillance has become more and more widespread. As the most basic core technology in the field of intelligent video surveillance, multi-target tracking technology has important research significance and broad application prospects, and has been widely concerned and studied by researchers from academia and industry all over the world. At present, the research of multi-target tracking technology has made great progress, but there are still many problems to be solved, such as complex tracking scene, attitude change of non-rigid object, target occlusion and real-time tracking. Aiming at the problems of object occlusion, apparent variation and similar targets in complex scenes, this paper studies multi-target tracking. The main research content and result is: (1) the basic theory of multi-target tracking is introduced. The basic principles of Kalman filter and particle filter based on Bayesian theory are introduced, and the advantages and disadvantages of the algorithm are analyzed. This paper introduces the basic principle of mean shift algorithm and fuzzy C-means algorithm, and studies the basic steps of the algorithm. In this paper, a particle filter multi-target tracking method based on adaptive blocking is studied. According to the gray level distribution of the target, the method adaptively divides blocks to improve the ability of accurately tracking multiple targets under occlusion, and obtains the corresponding particle swarm of each target by means of mean shift and fuzzy C-means clustering in particle filter tracking. The optimal state estimation of the target is obtained, and the weighted Bhattacharyya distance is introduced to calculate the matching degree of the subblock. The influence of sub-block reliability on particle weight is considered. A multi-feature fusion multi-target tracking method based on adaptive block is proposed. Based on the previous method, a multi-feature fusion strategy is added to describe the target with color histogram and HOG feature. According to the reliability of the subblock and the spatial distribution of the particle, the weight of the neutron block of the target model is adjusted in time, and in order to reduce the influence of the target change on the tracking result, the target feature model is dynamically updated by the weight updating method. The experimental results show that the proposed method can track multiple targets accurately and robustly in the case of the apparent changes in the process of multi-target tracking, the similarity of targets, and the complexity of object occlusion or mutual occlusion of similar targets.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【相似文献】