基于自适应交互式多模型粒子滤波的分布式说话人跟踪算法

Distributed speaker tracking algorithm based on adaptive interacting multiple model and particle filtering

针对分布式麦克风网络中的说话人跟踪问题,提出一种自适应交互式多模型粒子滤波算法,以实现复杂环境下对说话人的分布式跟踪.首先,对分布式麦克风网络中的说话人跟踪问题建立状态空间模型,并利用贝叶斯滤波理论求解该问题.然后,将交互式多模型与粒子滤波相结合,提出一种双粒子滤波方法对运动模型的转换概率进行自适应估计,以更好地对多种运动模式的说话人进行跟踪.最后,应用一致性算法对分布式麦克风网络中各节点说话人位置矢量的后验分布进行最优融合,从而可能得到全局的最优估计结果.该算法不要求状态空间模型中运动模型转换概率已知,相比传统IMMPF算法对声源复杂运动具有更好的鲁棒性.仿真实验结果验证了该算法的有效性.

To achieve distributed speaker tracking in complex environment,an adaptive interacting multiple model and particle filter algorithm is proposed for speaker tracking in distributed microphone networks.Firstly,a state space model is established for speaker tracking problem in distributed microphone networks,and Bayesian filtering theory is used to solve the problem.Then,the interacting multiple model is combined with the particle filtering,and the dual-particle filter method is presented to adaptively estimate the transition probability of the motion model in order to better track the speakers with multiple motion modes.Last,the consensus algorithm is used to realize the approximate optimal fusion of the posterior distribution estimates of the speaker position vector calculated by each node in the distributed microphone network,so as to nearly achieve the global optimal estimation results.The proposed algorithm does not require that the transition probability of the motion model in the state space model is known,and improves the robustness of the traditional IMMPF algorithm to the complex trajectory of the sound source.Simulation test results show the effectiveness of the proposed algorithm.