Human Action Recognition Using Depth Images

This paper presents a human action recognition algorithm using a depth image. First, 3D coordinates of the body’s joints of each frame are generated from the depth image. Then, the proposed method applies normalization and quantization processes to the body joints of all frames of the action video to obtain a 3D histogram. The histogram is projected onto xy, xz, and yz plans sequentially and combined into a one-dimensional feature vector. For dimension reduction, the principal component analysis (PCA) technique is applied to the feature vector to generate an action descriptor. To further improve the recognition performance, a decision tree method is developed to divide input actions into four main categories. The action description vectors of each category are used to design its respective support vector machine (SVM) classifier. Each SVM classifies the actions of a category into one type of actions. Experimental results verify that our approach effectively rules out the interference of background and improves the recognition rate.