3D PostureNet: A unified framework for skeleton-based posture recognition

Abstract

Abstract Image-based posture recognition is a very challenging problem as it is difficult to acquire rich 3D information from postures in 2D images. Existing methods founded on 3D skeleton cues could alleviate this issue, but they are not particularly efficient due to the application of handcrafted features and traditional classifiers. This paper presents a novel and unified framework for skeleton-based posture recognition, applying powerful 3D Convolutional Neural Network (CNN) to this issue. Technically, bounding-box-based normalization for the raw skeleton data is proposed to eliminate the coordinate differences caused by diverse recording environments and posture displacements. Moreover, Gaussian voxelization for the skeleton is employed to expressively represent the posture configuration. Thereby, an end-to-end framework based on 3D CNN, called 3D PostureNet, is developed for robust posture recognition. To verify its effectiveness, a large-scale writing posture dataset is created and released in this work, including 113,400 samples of 30 subjects with 15 postures. Extensive experiments on the public MSRA hand gesture dataset, body pose dataset and the proposed writing posture dataset demonstrate that 3D PostureNet achieves significantly superior performance on both skeleton-based human posture and hand posture recognition tasks.