摘要
为有效解决可变目标在跟踪过程中的"漂移"问题,提出一种基于自适应先验表观模型的目标跟踪方法。该方法首先在一致架构内融合HDP-EVO演化聚类模型和在线Boosting学习。以Dirichlet过程为先验分布,对总体表观示例进行聚类分析,获得随时间自适应演化的表观类先验知识,进而利用共享的表观类混合比例的权重平滑约束各时刻的表观模型。改进Gibbs抽样过程,使之能融入目标示例的分类误差,并交替迭代地从数据中自主学习聚类和表观分类器。最后,根据表观模型中各表观类的权重系数组合它们的分类评分去定位目标位置。仿真实验表明新方法学习的表观模型能较鲁棒地自适应于目标的表观变化,提高了跟踪精度。
To solve the drifting problem of a varying target object during target tracking,an adaptive prior appearance model was presented. The hierarchical Dirichlet process evolutionary clustering model and online boosting learning were combined into a coherent framework. By taking the hierarchical Dirichlet process as prior distribution,the prior appearance knowledge could adapt to change over time. Further,the appearance model was smoothly constrained by the mixture proportion of a type of appearance cluster at each moment. To balance the classification error of appearance model and the cost for splitting the clusters,the multi-modal appearance model was automatically learned by the use of Bayesian posterior inference. Finally,based on the weighting factor of appearance clusters,the target object was discriminated by combining the outputs of appearance classifiers. The simulation results showed that the learned appearance model could robustly adapt to the appearance variations,as well as the tracking results with high accuracy.
出处
《四川大学学报（工程科学版）》
EI
CAS
CSCD
北大核心
2013年第5期69-75,共7页
Journal of Sichuan University (Engineering Science Edition)
基金
中央高校基本科研业务费资助(CDJXS10180004)
重庆市自然科学基金资助项目(CSTC2008BB2191)
作者简介
作者简介I孙建中（1976-），男，博士生．研究方向：计算机视觉、视频检测与目标跟踪．E-mailtsjz3000@sina．com
