基于可转移字典对的跨视角动作识别

摘要

Searching discriminative appearance features of human actions is crucial for action recognition,which is usually sensitive to changes in viewpoint.This paper proposes a framework for crossview action recognition based on sparse representations using a transferable dictionary pair.The dictionary pair consists of two dictionaries belonging to the source and target views respectively.By forcing the same action videos captured from two different have the same sparse representation,the transferable dictionary bridges the gap of features between the two views.Both supervised and unsupervised algorithms for learning transferable dictionary pairs are proposed.With the sparse representation being features,a classifier trained in source view can be applied in target view directly.It evaluates the effectiveness of the approach on multi-view,multi-modality and multi-action 3M dataset.%为人体动作寻找具有分辨力的视觉特征是机器视觉研究领域的重要课题,但当视角发生改变时其效果往往不够理想.文中提出了一种视角无关的动作识别方法,利用转移字典对完成视角间信息的转移.转移字典对包含两个字典,分别对应于源视角和目标视角.字典对的学习过程是自发的,其准则是尽量使两个视角中的同一动作具有相同的稀疏表示.提出了有监督和无监督条件下的算法,用于转移字典对的学习.利用转移字典对将两个视角中的视频进行稀疏表示之后,在源视角下训练得到的分类器即可直接用于目标视角.方法的有效性在多视角、多模态的数据库3M上进行验证,取得了良好的效果.