Learning shape and motion representations for view invariant skeleton-based action recognition
作者：
Highlights：
• Skeleton sequence space as a subset of Geometric Algebra is constructed to represent each skeleton sequence along both spatial and temporal dimensions.
• Rotor-based view transformation overcomes the view variation challenge and reserves relative motions among skeletons.
• Spatio-temporal view invariant model is constructed to model spatial configuration and temporal dynamics of skeleton joints and bones.
• Four skeleton sequence shape and motion representations are learned to comprehensively describe skeleton-based actions, which are fed to a selected multi-stream convolutional neural network for action recognition.
摘要
•Skeleton sequence space as a subset of Geometric Algebra is constructed to represent each skeleton sequence along both spatial and temporal dimensions.•Rotor-based view transformation overcomes the view variation challenge and reserves relative motions among skeletons.•Spatio-temporal view invariant model is constructed to model spatial configuration and temporal dynamics of skeleton joints and bones.•Four skeleton sequence shape and motion representations are learned to comprehensively describe skeleton-based actions, which are fed to a selected multi-stream convolutional neural network for action recognition.
论文关键词：Human action recognition,Skeleton sequence,Representation learning,View invariant,Geometric Algebra
论文评审过程：Received 20 April 2019, Revised 18 December 2019, Accepted 19 February 2020, Available online 21 February 2020, Version of Record 29 February 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107293
