基于流形假设的骨架序列动作识别算法

骨架数据是通过对动作的空间几何位置进行编码获取,可以避免冗余背景信息的干扰, 是动作识别领域常用的数据类型之一.现有骨架数据的动作识别主要分为经典的骨架数据表征和基于深度学习的骨架动作识别应用.相较于传统欧氏度量下的识别方法,流形为更好地研究非线性结构提供了重要数学工具. 然而,目前仍缺乏利用流形假设对骨架数据进行动作识别的相关总结. 因此,从骨架表示、轨迹时间对齐、动作序列表征以及动作分类 4 个关键步骤出发,系统地总结了基于流形假设的动作识别工作,对比了各项工作在基准数据集上的表现. 最后,根据当前动作识别工作的发展趋势,对流形假设在动作识别方向上的进一步改进进行了展望.

Skeleton-based action recognition by manifold assumption

Skeletal data are obtained by encoding the spatial geometricposition of the action, which can prevent the interference ofredundant background information. It is one of the commonly useddata types in the field of action recognition. The existing reviewof action recognition related to skeletal data is mainly dividedinto the classical skeletal data representation and the applicationof skeletal action recognition based on deep learning. Compared withthe action recognition methods based on the traditional Euclideanmetric, manifolds provide an important mathematical tool for abetter study of nonlinear structures. However, there is still a lackof summaries about action recognition from skeletal data using themanifold assumption. Therefore, starting from the four steps ofskeleton representation -- trajectory temporal alignment, actionsequence characterization, and action classification -- thisarticle systematically summarizes the action recognition work basedon the manifold assumption, and compares the performance of eachwork on the benchmark datasets. Finally, according to the currentdevelopment trend of action recognition, further improvement of themanifold assumption in thedirection of action recognition is prospected.