基于深度稀疏滤波的目标跟踪.pdf

第29 卷 第3 期 计算机辅助设计与图形学学报 Vol. 29 No.3 2017 年3 月 Journal of Computer-Aided Design Computer Graphics Mar. 2017 基于深度稀疏滤波的目标跟踪 邱立达, 刘天键, 傅 平 (闽江学院物理学与电子信息工程系 福州 350108) (qld.qq@163.com) : 为了在复杂环境下更好地区分被跟踪目标和背景, 设计了一种基于l 范数归一化和l 范数最小化的深度稀 2 1 疏滤波模型, 通过深度学习获取样本稀疏特征并对其进行分类, 将该模型和粒子滤波框架结合, 提出一种目标跟踪 算法. 首先使用离线训练集对深度稀疏滤波模型进行逐层无监督预训练得到权值参数的初始值, 然后在跟踪过程中 利用标签样本对模型在线更新, 通过有监督微调优化其权值参数使得模型能够更好地适应目标外观变化, 最后利用 训练好的深度稀疏滤波模型对经由粒子滤波算法给出的候选区域进行观测, 以确定跟踪目标. 在不同视频序列中的 实验表明, 文中算法在复杂条件下具有良好的跟踪精度和鲁棒性. : 目标跟踪; 深度学习; 稀疏滤波; 粒子滤波 : TP391.41 Target Tracking Based on Deep Sparse Filtering Qiu Lida, Liu Tianjian, and Fu Ping (Department of Physics and Electronic Information Engineering, Minjiang University, Fuzhou 350108) Abstract: To better distinguish tracked target from background in the complex environment, a deep sparse filtering model based on l -norm normalization and l -norm minimization was designed to learn sparse sam- 2 1 ple features and the classification by deep learning is carried out. Additionally, a tracking algorithm com- bining this model with particle filter framework was proposed. Firstly, the deep sparse filtering model is pre-trained layer by layer on the offline training samples set without supervision to get the initial value of weight parameters. Then, during the tracking process, the model is updated online with the labeled samples by supervised fine-tuning to further optimize the weight parameters for better adaptability to target appear- ance changes. Finally, the full-trained deep sparse filtering model is used to observe candidate areas calcu- lated by particle filter algorithm to determine the trac