基于鲁棒高斯混合概率假设密度滤波器的多目标视频对象跟踪

With the development of the computer vision technology, the intelligent multi-target tracking system has been widely applied to intelligent surveillance, visual human-computer interactions and autonomous robot navigations et al. However, the problems include the data association between the target and the measurements, and the robust tracking of the newborn target, closed moving targets and the targets in mutual occlusion, remain an important research interest in multi-target video tracking. In this project, we develop a robust multi-target video tracking system based on the conventional Gaussian mixture probability hypothesis density filter. To eliminate the inferences of the noises as well as to reduce the false positive rate, the instability of the Shannon entropy distribution is used to automatically and correctly estimate the birth intensity of the newborn target. To reduce the mismatch rate between the target and its identity, the color-based appearance model is incorporated to penalize the weights of the closed moving targets. To reduce the miss tracking rate and the false positive rate, a game theory-based method is proposed to obtain the optimal positions of the measurements in mutual occlusion.

随着计算机视觉技术的发展，多目标智能跟踪系统已广泛地应用于智能监控、虚拟人机交互以及自主机器人导航等诸多领域。然而，如何有效解决目标与测量值间的数据关联问题和新生目标、邻近目标以及遮挡目标的鲁棒跟踪问题仍然是多目标视频跟踪亟需解决的技术重点和难点。我们提出在传统高斯混合概率假设密度滤波器的基础上，开发出一套鲁棒的多目标视频对象跟踪系统。利用香农熵分布的不稳定性，自动、准确地估计新生目标出生强度，削弱噪声的干扰，减少新生目标跟踪的错误率；融合颜色外观模型对邻近目标权值进行惩罚，减少目标与身份的不匹配率；利用博弈论的方法解决遮挡区域目标测量值的位置优化问题，准确有效地跟踪遮挡区域目标，减少系统漏跟及错跟率。

本项目针对基于高斯混合概率假设密度滤波器的多目标视频跟踪中存在的噪声干扰大、邻近目标跟踪难和遮挡问题，提出了一种基于香侬熵分布的新生目标强度估计方法，滤除跟踪过程中与目标无关的噪声分量；提出了一种融合多特征的高斯权值惩罚方法，通过改进目标外观模型，惩罚邻近目标的高斯权值，提高了邻近目标跟踪的准确率；提出了一种基于博弈论的遮挡目标跟踪方法，通过优化遮挡目标外观模型，构建零和非合作博弈模型，将遮挡目标跟踪定位问题转化为求取纳什均衡问题，有效提高了遮挡区域目标跟踪的准确度和精度。实验结果验证了本项目所提方法的有效性和优越性，相关研究成果可为智能视频监控和智能机器人领域提供理论依据和技术支撑。