3D-LSR: A New Model for Skeletal Representation and Segmentation
Di Zhang, Xiaojuan Ban, Di Wu, Hao Henry Wang · 2016
Skeleton-based human action recognition with depth sensors have a renewed interest based on the development of popular depth sensors. This paper presents a new skeletal representation, Limb-based Skeleton Representation (LSR) that explicitly models 10 limbs as spherical coordinate in 3D space, and extract angular velocity feature based on the spherical coordinate representation. Using the Spherical Coordinate and Angular Velocity (SCAV) feature, human actions can be represented as spatial-temporal vector sequences. To address the classification challenge of vector sequences that can vary in duration and sometimes temporal ordering, we propose a Dynamic Segmentation Algorithm (DSA) and encode all the segments by orientations and spherical coordinates. We apply LSR and DSA on a standard UTD-MHAD human action dataset, in comparison with 3 other methods. The result shows the classification accuracy have been improved.