Spatio-Temporal Feature for Human Action Recognition Using Skeleton Data

Human action recognition is a key element in many human centric applications. Development of depth imagingsystems and enhanced machine vision techniques have led to improved action recognition systems that have solved manyproblems of video-based action recognition. Conventionally, position of the human joints is extracted from the depthimage and used to extract features for human pose representation. Relative joint displacement and joint orientation arecommonly used in this regard. However, the effectiveness of these features and their combinations have not been studied.In this paper, relative joints displacement and joint orientation in spatial and temporal states and their combinations wereevaluated for action recognition. The methods were tested on 3 publicly available datasets and 3 evaluation strategieswere used for interpretation of the results.

Keywords:

Action recognition , depth image , skeleton representation , joint displacement features , joint orientation features