人体关节运动跟踪是非线性、非高斯系统的运动状态估计问题，粒子滤波器是实现人体运动跟踪的有效手段。粒子状态的预测与更新是影响粒子滤波器性能的关键，预测模型反映人体运动规律的程度是决定使用粒子滤波器能否进行关节运动准确跟踪的主要因素之一。本文提出一种基于人体运动模式识别的关节运动跟踪粒子滤波器架构，在将运动模式定义为运动基的基础上，利用R(2+1)D网络进行运动基类型识别。同时，根据识别所得到的运动基概率密度分布，分配每个运动基对应预测模型的粒子个数并进行关节运动状态的先验概率密度分布计算。在粒子状态更新阶段，选取颜色直方图特征计算粒子适应度，在对粒子状态进行重采样更新的基础上修正运动基的概率密度分布，从而达到了基于粒子滤波器的人体运动模式识别与状态跟踪联合实现的目的。实现结果表明，本文提出的新型粒子滤波器架构在关节跟踪准确性上远优于传统粒子滤波器方案，与基于深度学习的点对点人体关节运动跟踪方法也具有一定的可比性。

Human joint motion tracking is a non-linear, non-Gaussian system motion state estimation problem, so particle filters have become an effective means to achieve human motion tracking. The prediction and update of the particle state is the key to affecting the performance of the particle filter. The degree to which the prediction model reflects the law of human motion is one of the main factors that determines whether the particle filter can be used to accurately track the joint motion. This paper proposes a joint motion tracking particle filter architecture based on human motion pattern recognition. On the basis of defining the motion pattern as a dyneme, the R(2+1)D network is used for dyneme type recognition. At the same time, according to the probability density distribution of the dynemes obtained by the recognition, the number of particles of the prediction model corresponding to each dyneme is allocated and the prior probability density distribution calculation of the joint motion state is performed. In the particle state update stage, the color histogram feature is selected to calculate the particle fitness, and the probability density distribution of the motion basis is corrected on the basis of re-sampling and updating the particle state, so as to achieve the human body motion pattern recognition and state tracking based on the particle filter The purpose of the joint realization. The implementation results show that the new particle filter architecture proposed in this paper is far superior to the traditional particle filter scheme in joint tracking accuracy, and it is also comparable to the point-to-point human joint motion tracking method based on deep learning.