作者：李福进, 李军, 宫海洋

作者单位：华北理工大学电气工程学院, 河北 唐山 063210

关键词：稀疏子空间;卷积神经网络;粒子滤波;目标跟踪;相异系数矩阵

摘要：

针对粒子滤波目标跟踪过程中初始化和权值退化的数据处理情况，在粒子滤波框架下提出一种基于稀疏子空间的卷积神经网络目标跟踪算法。以仿生学为基础，在目标跟踪过程中引入稀疏子空间和卷积神经网络。首先，利用稀疏子空间模型筛选出与目标状态相似度较高的候选区域进行后续跟踪处理，减少冗余计算并降低跟踪的复杂性；然后，将稀疏子空间输出用作卷积神经网络的输入，并利用卷积神经网络模型对图像数据处理的优点进行目标跟踪的数据处理；最后，通过对目标数据的不断更新来减少目标表观变化的影响。实验表明，该算法能够更好地处理目标跟踪中的目标遮挡、运动模糊、光流与尺度变化，提高算法的准确性和数据处理能力。

Convolutional neural network target tracking based on sparse subspace
LI Fujin, LI Jun, GONG Haiyang
College of Electrical Engineering, North China University of Science and Technology, Tangshan 063210, China
Abstract: Aiming at the data processing situation of initialization and weight degradation in particle filter target tracking process, a particle tracking algorithm based on sparse subspace was proposed in the particle filter framework. Based on bionics, sparse subspace and convolution neural network were introduced in the target tracking process. Firstly, the candidate regions with high similarity with the target state was selected by the sparse subspace model, which can reduce the redundancy calculation and reduce the complexity of the tracking. Then, the output of the sparse subspace is used as the input of the convolution neural network, and the convolutional neural network model was used to perform the target tracking data processing on the advantages of image data processing. And finally, through the continuous updating of the tracking data to reduce the impact of the apparent changes in the target. Compared with the current mainstream tracking method, the experimental results show that the algorithm can deal with the problem of target occlusion, motion blur, optical flow and scale change in target tracking, and improve the accuracy and data processing ability of the algorithm.
Keywords: sparse subspace;convolution neural network;particle filter;target tracking;coefficient of dissimilarity matrix
2019, 45(7):122-127  收稿日期: 2018-05-16;收到修改稿日期: 2018-07-08
基金项目: 国家自然科学基金（61203343）；河北省自然科学基金（E2014209106）；河北省高等学校科学技术研究青年基金项目（QN2016102，QN2016105）
作者简介: 李福进(1957-),男,河北唐山市人,教授,硕士生导师,博士,主要研究方向是智能控制与智能仪表
参考文献
[1] WANG H, YUAN C, LUO G, et al. Action recognition using linear dynamic systems[J]. Pattern Recognition, 2013, 46(6):1710-1718
[2] YANG S, YUAN C, WANG H, et al. Combining sparse appearance features and dense motion features via random f orest for action detection[C]//IEEE International Conference on Acoustics:IEEE, 2013.
[3] 张才千, 葛磊, 韩东. 基于目标跟踪的粒子群粒子滤波算法研究[J]. 计算机仿真, 2014, 31(8):392-396
[4] 许伦辉, 丛晓野. 改进粒子滤波对人物跟踪的应用[J]. 计算机仿真, 2014, 31(1):344-347
[5] 许婉君, 侯志强, 余旺盛, 等. 基于颜色和空间信息的多特征融合目标跟踪算法[J]. 应用光学, 2015, 36(5):755-761
[6] 陈芸, 吴飞, 荆晓远, 等. 鲁棒低秩稀疏表示的在线目标跟踪[J]. 计算机工程与设计, 2016, 37(4):1062-1066
[7] 赵二群. 视觉神经系统仿生模型及其应用研究[D]. 长沙:湖南大学, 2014.
[8] HINAULT T, DUFAU S, LEMAIRE P. Strategy combination in human cognition:a behavioral and ERP study in arithmetic[J]. Psychonomic Bulletin & Review, 2015, 22(1):190-199
[9] OLSHAUSEN B A, FIELD D J. Natural image statistics and efficient coding[J]. Network, 2009, 7(2):333
[10] 王保宪, 赵保军, 唐林波, 等. 基于双向稀疏表示的鲁棒目标跟踪算法[J]. 物理学报, 2014(23):174-184
[11] YIN H. Sparse representation with learned multi-scale dictionary for image fusion[J]. Neurocomputing, 2015, 148(148):600-610
[12] DONG C, CHEN C L, HE K, et al. Image super-resolutionusing deep convolutional networks[J]. IEEE Transactions onPattern Analysis and Machine Intelligence, 2016, 38(2):295-307
[13] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Imagend classification with deep convolutional neural[C]//Advances in Neural Information Processing Systems, 2012.