Skeleton-based Human Action Recognition Using Spatio-Temporal Geometry (ICCAS 2019)

In this paper, we propose a temporal action recognition algorithm in a sequence image. Our methods consists of three main parts. First, it extracts the 3D human pose from 2D skeletal joints from a single image. Second, we investigate a spatio-temporal structure using the extrinsic and intrinsic connectivity that correspond to joints in a human action. We extract key joints defined as endpoints of the skeleton in each frame in order to reflect temporal variation of the joints. Due to these key joints, we can consider relative position variances of corresponding points in single frames. Finally, we exploit these motion vector to infer the spatio-temporal structure of human actions. We will compare the accuracy and efficiency of this approach with our method.