【作者】 于虹；

【导师】 李金；

【作者基本信息】 哈尔滨工程大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 人体运动分析是计算机视觉领域一个非常有前途的研究方向，对它的研究涉及计算机视觉、人工智能、模式识别和图像处理等学科领域，是一个跨学科的挑战性研究课题。人体运动分析的主要任务是对视频序列中的人体运动进行跟踪。由于人体运动的实质是骨骼围绕关节点的运动，因而人体关节运动跟踪技术的研究在人体运动跟踪中最具有代表性，它的准确跟踪使其在运动员动作分析、辅助临床诊断、计算机动画等方面有着广阔应用前景的研究课题。本文针对人体关节目标跟踪中存在的难点问题，进行了人体关节运动跟踪技术的研究。本文首先研究了核密度估计及无参密度估计均值偏移(Mean Shift)理论，其次，研究了应用于目标跟踪中的Mean Shift算法。针对目标在运动过程中，由于其自身等条件发生变化时，需要对模板进行更新的问题，进行了模板更新条件及模板更新方法的研究，本文提出了采用人体关节目标之间的Bhattacharyya距离作为模板更新的条件和新旧模板加权的更新方法，使传统Mean Shift算法对人体关节运动目标的跟踪精度得到提高。同时针对目标受光照变化等环境条件影响时，在传统Mean Shift算法中，若仅利用目标单一颜色特征对人体关节运动目标进行跟踪的不可靠性，研究了多种矩不变量方法，提出了基于Mean Shift的人体关节运动跟踪算法。该算法利用人体关节目标的速度、小波矩不变量和颜色分布特征进行目标跟踪，大大减轻了传统Mean Shift对运动目标受光照变化而造成的跟踪不准确性，且对运动目标的遮挡具有一定的容忍能力。进行了采用卡尔曼滤波缩小(Kalman Filter)对人体关节目标搜索范围的研究，提出了基于Kalman Filter和Mean Shift的人体关节运动跟踪算法。该跟踪算法充分利用卡尔曼滤波对人体关节目标在当前帧可能位置的预测，并发挥了运动目标的小波矩不变量和颜色分布特征在目标跟踪中的优势，为减少Mean Shift的迭代运算和提高Mean Shift的搜索效率提供了一种较为有效的途径。研究了对于遮挡等造成的多峰值、非高斯分布和非线性问题的目标跟踪算法，并针对利用目标单一特征信息往往很难实现对运动目标鲁棒跟踪的情况，提出了一种将目标特征信息的观测模型结合到无迹粒子滤波(UPF)中的人体关节运动跟踪算法。该方法充分利用无迹粒子滤波算法具有的多重假设以及最新的观测信息。实验结果表明，本论文提出的基于目标特征信息和UPF的人体关节运动跟踪算法很好地解决了目标被频繁遮挡的跟踪问题，并具有良好的鲁棒性。对人体关节目标在运动过程中出现较长时间遮挡导致无迹粒子滤波存在样本贫化现象，需要增加样本集的多样性，进行了智能优化算法的研究，提出了基于智能优化和无迹粒子滤波的人体关节运动跟踪算法，使无迹粒子滤波的样本贫化现象得到了改善。本论文在均值偏移、粒子滤波、无迹粒子滤波及智能优化算法等方面进行了较为深入的研究，取得了一些有益的成果。这些成果对人体关节运动目标的准确、鲁棒、可靠的跟踪起到了重要的作用。更多还原

【Abstract】 Video human motion analysis is a very promising research area which combines subjects of computer vision, artificial intelligence, pattern recognition, image processing, and so on. Thus this work is an interdisciplinary and challenging research topic. Tracking human motion from image sequences is main task for human motion analysis. Due to the essential of human motion is the bone of movement around human joints, thus the tracking techniques of human joints are most representative in the human motion tracking. The tracking technique of the veracity human joints is a research topic which has a wide range of applications in athlete action analysis, computer-aided clinical diagnosis, computer animation, and so on.For the difficult problems of human joints tracking, the tracking techniques of human joints movement are researched in the dissertation.The kernel density estimation and the Mean Shift theory of nonparametric density estimation are studied in this dissertation. Then the application of Mean Shift algorithm (MSA) in the targets tracking is researched. Aiming at the motion process of the targets, the template update is needed when its own conditions changes. The condition and method of template update are studied. The Bhattacharyya distance between the human joints as the condition of template update, at the same time, the weighted new and old template of update method is put forward for the Mean Shift in this dissertation. The traditional Mean Shift is improved the tracking precision of the moving human joints.For the unreliability of single color character in the traditional MSA when targets are influenced by the illumination and other environmental conditions, diversified moment invariants are researched and a tracking method of human joints movement based on Mean Shift is proposed. The velocity, wavelet moment invariants and color distribution characteristic are utilized by the proposed algorithm, in which the tracking inaccuracy by the illumination effect is greatly reduced. In the mean time, this new algorithm has the tolerant ability for the occlusion of the moving targets.The research based on Kalman Filter (KF) is adopted to reduce the search range of human joints, and the tracking algorithm based on Kalman Filter and Mean Shift is brought forward. In current frame, Kalman Filter is adequately used to forecast the possible positions of human joints, and the advantages of wavelet moment invariants and color distribution in the targets tracking are played by the proposed algorithm. The method can decreases the iterative operation, but offers an effective approach to improve the searching efficiency.The tracking algorithm of the targets is researched for the multi-mode, non-Gaussian distribution and nonlinear problem by the occlusion. And aiming at the single feature information of the target, the robust tracking can not be completed. Thus a tracking algorithm of human joints motion is proposed, in which the observation model based on the character information is included in the unscented particle filter (UPF). The multi-hypothesis and the latest observation information of UPF are made the best of the proposed method. Experiment results demonstrate that the human joints tracking algorithm based on feature information and UPF is proposed in the dissertation, which can settles well the tracking problem for the targets of the frequent occlusions, and have good robust.Aiming at a longer time occlusion in the moving process of human joints, the sample impoverishment phenomenon of the unscented particle filter is produced. And the diversity of sample collection needs be added. So the intelligence optimization algorithm is researched and a tracking algorithm of moving human joints based on intelligence optimization and UPF is put forward. The sample impoverishment phenomenon is improved by the proposed tracking algorithm in this dissertation.This dissertation makes deeply research on the Mean Shift、Particle Filter (PF)、Unscented Particle Filter and intelligence optimization algorithm. And some beneficial results are obtained. These results give the important effect for the veracity, robust and reliable tracking of the moving human joints.更多还原