Abstract: Concerning the high time complexity and low recognition rate of action recognition algorithm in traditional RGB video, this paper proposes an action recognition approach based on depth image. Firstly, each depth frame in a depth video sequence is projected onto three orthogonal Cartesian planes. Secondly, the Gabor feature is extracted from the projected maps in three projective views. Lastly, the features are used to train extreme learning machine classifiers that are used to classify actions. Experiments on public dataset MSR Action 3D show the method achieves average recognition rate 97.80%, 99.10% and 88.35% respectively on different three group experiments. The elapsed time of recognizing single depth video is less than 1 second. The results show the proposed method can effectively identify the human motion in depth image sequence and satisfy the real-time requirement of the depth image sequence.

Key words: depth image, projection, Gabor feature, extreme learning machine, action recognition

摘要： 针对传统RGB视频中动作识别算法时间复杂度高而识别准确率低的问题，提出一种基于深度图像的动作识别方法。该方法首先对深度图像在三投影面系中进行投影，然后对三个投影图分别提取Gabor特征，最后使用这些特征训练极限学习机分类器，从而完成动作分类。在公开数据集MSR Action3D上进行了实验验证，该方法在三组实验上的平均准确率分别为97.80%、99.10%和88.35%，识别单个深度视频的用时小于1 s。实验结果表明，该方法能够对深度图像序列中的人体动作进行有效识别，并基本满足深度序列识别的实时性要求。

关键词: 深度图像, 投影, Gabor特征, 极限学习机, 动作识别