A Novel 3D Human Action Recognition Method Based on Part Affinity Fields

Abstract

The analysis of human actions based on 3D skeleton data becomes popular recently due to its succinctness, robustness, and view-invariant representation. Recent attempts on this problem suggested to use human body affinity fields to efficiently detect the 2D pose of multiple people in an image. In this paper, we extend this idea to 3D domains and develop a 3D human action recognition system with ability to understand the cooperative action of several people. To achieve this, we firstly extract the human body affinity fields to robustly represent associate 2D human skeleton with individuals in the image. Inspired by the triangulation techniques in stereo vision analysis, 3D human skeleton data can be obtained. To handle the noise in constructed 3D human skeleton data, we introduce an enhanced light-weight matching algorithm based on Dynamic Time Warping (DTW) to compute the matching cost. The real-life experiments demonstrate the efficiency and applicability of our approach.