基于分布式自适应UKF的说话人跟踪方法  

Speaker tracking method with the distributed adaptive UKF

作　　者：陈阳[1] 蔡翔宇 王睿 CHEN Yang;CAI Xiangyu;WANG Rui(School of Microelectronics and Control Engineering,Changzhou University,Changzhou 213164,Jiangsu,China;School of Computer Science and Artificial Intelligence,Changzhou University,Changzhou 213164,Jiangsu,China)

机构地区：[1]常州大学微电子与控制工程学院,江苏常州213164 [2]常州大学计算机与人工智能学院,江苏常州213164

出　　处：《声学技术》2024年第5期734-742,共9页Technical Acoustics

基　　金：自然科学基金(62201527);常州市基础研究计划(应用基础研究)(CJ20220100)。

摘　　要：针布式无迹卡尔曼滤波(distributed unscented Kalman filter, DUKF)方法进行说话人跟踪时,因状态转移噪声协方差矩阵和测量噪声协方差矩阵偏离真实值而导致跟踪误差增大。文章采用塞琪-胡萨(Sage-Husa)自适应策略,在DUKF测量更新后迭代估计局部状态转移噪声协方差矩阵和测量噪声协方差矩阵,然后利用一致性滤波融合得到全局的状态转移噪声协方差矩阵,随着卡尔曼滤波器的迭代,逐渐逼近状态转移噪声协方差矩阵和测量噪声协方差矩阵的真实值,从而提高DUKF说话人跟踪精度。实验结果表明,即使在较差的噪声和混响条件下,分布式自适应无迹卡尔曼滤波方法相较于常规的DUKF方法仍具有更好的跟踪性能,在节点损坏条件下的鲁棒性更强,能够获得更准确的说话人位置信息。The speaker tracking method of the distributed unscented Kalman filter(DUKF)needs accurate values of the state transition noise covariance matrix and the measurement noise covariance matrix.Otherwise,the tracking performance will degenerate if their values deviate from the actual values.In this paper,the Sage-Husa adaptive strategy is adopted to iteratively estimate the local state transition noise covariance matrix and measurement noise covariance matrix after updating the DUKF measurement.With the iteration of DUKF measurements,the estimates of the state transition noise covariance matrix and the measurement noise covariance matrix gradually approach to their actual values,thus the speaker tracking accuracy is improved.Experimental results show that the speaker tracking method with the proposed distributed adaptive UKF(DAUKF)achieves better tracking performance than with the conventional DUKF even under poor noise and reverberation conditions,stronger robustness under node damage conditions,and can obtain more accurate speaker position information.

关 键 词：分布式无迹卡尔曼滤波(DUKF) 说话人跟踪 分布式麦克风网络 自适应策略 

分 类 号：TN912[电子电信—通信与信息系统]