A study of feature combination in gesture recognition with Kinect

Abstract

Human gesture recognition is an interdisciplinary problem, with many important applications. In the structure of a gesture recognition system, feature extraction, without doubt, is one of the most important factor affecting the performance. In this paper, we desired to improve the covariance feature, which is the current state-of-the-art feature extraction method, by integrating other frame-level features extracted in the data captured by Microsoft Kinect, and experimenting the features with various classification methods such as Random Forest (RF), Multi Layer Perceptron (MLP), Support Vector Machines (SVM). The leave-person-out experiments showed that feature combination is beneficial, especially with Random Forest, to achieve the highest score in recognition, which is improved by 2%, from 90.9% to 93.0%. However, the dimensional increase sometimes exacerbated the performance, indicating the side effect of feature combination.