摘要
为了解决粒子滤波多说话人跟踪过程中粒子易发散导致多目标跟踪精度低的问题,提出了并行粒子滤波和基于GPU的K-均值聚类的多声源定位方法。该方法首先分析了粒子滤波在实现多目标跟踪时进行数据关联的过程产生较大的计算量,并且出现多个目标时,粒子会逐渐发散。针对计算量大和粒子发散的问题,提出了一种并行粒子滤波和K-均值聚类的方法。实验表明,随着粒子数和目标数的增加,计算量以指数增加,并且粒子发散严重,采用基于GPU的K-均值聚类方法的粒子滤波多说话人跟踪方法,相比传统粒子滤波跟踪方法具有更收敛的粒子集并且跟踪精度较高。
In order to solve the problem of low accuracy of multi-target tracking,particles in particle filter are easy to disperse in the process of multi-speaker tracking. This paper presented a parallel particle filter algorithm and GPU-based K-means clustering multi-source localization method. The method first analyzed the particle filter to achieve multi-target tracking,data association process had a large amount of computation,and the particles gradually diverged with the emergence of multiple targets. In order to solve the problem of large amount of computation and particle divergence,this paper proposed a method of parallel particle filter and K-means clustering. Experiments show that,as the number of particles and the number of targets increases,the amount of computation increases exponentially and the particles scatter seriously. By using the GPU-based K-means clustering method,the particle filter multi-speaker tracking method has more convergent particle sets and higher tracking accuracy than the traditional particle filter tracking method.
引文
[1]曹洁,余丽珍.基于MFCC和运动强度聚类初始化的多说话人识别[J].计算机应用研究,2012,29(9):3295-3298.(Cao Jie,YuLizhen. Multi-speaker recognition based on MFCC and motion intensi-ty clustering initialization[J]. Application Research of Compu-ters,2012,29(9):3295-3298.)
[2]屈丹,张文林.基于本征音子说话人子空间的说话人自适应算法[J].电子与信息学报,2015,37(6):1350-1356.(Qu Dan,ZhangWenlin. Speaker adaptation algorithm based on eigen-phonon speakersubspace[J]. Journal of Electronics and Information,2015,37(6):1350-1356.)
[3]张微,康宝生.相关滤波目标跟踪进展综述[J].中国图象图形学报,2017,22(8):1017-1033.(Zhang Wei,Kang Baosheng. Reviewof the progress of correlation filtering target tracking[J]. ChineseJournal of Image Graphics,2017,22(8):1017-1033.)
[4]任航.基于拟蒙特卡洛滤波的改进式粒子滤波目标跟踪算法[J].电子测量与仪器学报,2015,29(2):289-295.(Ren Hang.Improved particle filter target tracking algorithm based on quasi-MonteCarlo filter[J]. Journal of Electronic Measurement and Instru-ment,2015,29(2):289-295.)
[5]秦永元,张洪钺,汪叔华.卡尔曼滤波与组合导航原理[M]. 3版.西安:西北工业大学出版社,2015.(Qin Yongyuan,ZhangHongyue,Wang Shuhua. Kalman filter and integrated navigation prin-ciple[M]. 3rd ed. Xi’an:Northwest University of Technology Press,2015.)
[6]翟卫欣,程承旗.基于Kalman滤波的Camshift运动跟踪算法[J].北京大学学报:自然科学版,2015,51(5):799-804.(ZhaiWeixin,Cheng Chengqi. Kalman filter-based Camshift motion trackingalgorithm[J]. Journal of Peking University:Natural Science,2015,51(5):799-804.)
[7]程兰,王志远,陈杰,等.基于粒子滤波和滑动平均扩展Kalman滤波的多径估计算法[J].电子与信息学报,2017,39(3):709-716.(Cheng Lan,Wang Zhiyuan,Chen Jie,et al. Multipath estimation algo-rithm based on particle filter and moving average spread Kalman filter[J]. Journal of Electronics and Information,2017,39(3):709-716.)
[8]雷明,韩崇昭,肖梅.扩展卡尔曼粒子滤波算法的一种修正方法[J].西安交通大学学报,2005,39(8):824-827.(Lei Ming,HanChongzhao,Xiao Mei. A modified method of extended Kalman particlefilter algorithm[J]. Journal of Xi’an Jiaotong University,2005,39(8):824-827.)
[9]张应博.基于无极卡尔曼滤波算法的雅可比矩阵估计[J].计算机应用,2011,31(6):1699-1702.(Zhang Yingbo. Jacobian matrixestimation based on infinite Kalman filter[J]. Computer Applica-tion,2011,31(6):1699-1702.)
[10]李天成,范红旗,孙树栋.粒子滤波理论、方法及其在多目标跟踪中的应用[J].自动化学报,2015,41(12):1981-2002.(LiTiancheng,Fan Hongqi,Sun Shudong. Particle filter theory,methodand its application in multi-target tracking[J]. Acta Automation,2015,41(12):1981-2002.)
[11]林静,杨继臣,张雪源,等.基于稀疏表示权重张量的音频特征提取算法[J].计算机应用,2016,36(5):1426-1429,1438.(LinJing,Yang Jichen,Zhang Xueyuan,et al. Audio feature extraction al-gorithm based on sparse representation weight tensor[J]. ComputerApplication,2016,36(5):1426-1429,1438.)
[12]孙海洋,张利.无人机跟踪场景下的粒子滤波算法的改进[J].计算机仿真,2017,34(2):84-87.(Sun Haiyang,Zhang Li. Improvedparticle filter algorithm for UAV tracking scene[J]. ComputerSimulation,2017,34(2):84-87.)
[13] Goyal B,Budhraja T,Bhatnagar R,et al. Implementation of particlefilters for single target tracking using CUDA[C]//Proc of the 5th In-ternational Conference on Advances in Computing and Communica-tions. Piscataway,NJ:IEEE Press,2016:28-32.
[14]Baydoun M,Dawi M,Ghaziri H. Enhanced parallel implementation ofthe K-means clustering algorithm[C]//Proc of International Confe-rence on Advances in Computational TOOLS for Engineering Applica-tions. Piscataway,NJ:IEEE Press,2016:7-11.
[15]李晓瑜,俞丽颖,雷航,等.一种K-means改进算法的并行化实现与应用[J].电子科技大学学报,2017,46(1):61-68.(LiXiaoyu,Yu Liying,Lei Hang,et al. Parallelization and application ofan improved K-means algorithm[J]. Journal of University of Elec-tro-nic Science and Technology,2017,46(1):61-68.)
[16]陈平华,陈传瑜.基于满二叉树的二分K-means聚类并行推荐算法[J].计算机工程与科学,2015,37(8):1450-1457.(ChenPinghua,Chen Chuanyu. Binary K-means clustering parallel recom-mendation algorithm based on full binary tree[J]. Computer Engi-neering and Science,2015,37(8):1450-1457.)