基于Kinect传感器的动态手势实时识别

Real-Time Dynamic Hand Gesture Recognition Based on Kinect Sensor

本文提出了一种基于Kinect传感器的动态手势实时识别方法,在预处理阶段通过OpenNI/NITE快速获取人体骨架,并从中得到关节点数据用于建立动态手势的运动轨迹特征描述子.提出了一种在全局约束条件下的权重化多维数据的动态时间扭曲算法(WM-DTW)来对手势轨迹序列进行训练和识别.实验结果表明,本文的识别方法比LDA算法和动态时间扭曲(DTW)算法有更高的识别率.

A reliable method for dynamic gesture recognition based on Kinect sensor is proposed in this paper. Firstly, the human skeleton is extracted by using OpenNI/NITE toolbox in the pre-processing stage. Then the feature descriptor of hand area is established for dynamic gesture trajectory. Finally, an algorithm of dynamic time warping called WM-DTW is proposed to handle the weighted multi-dimensional data under the global constraint for trajectory sequence matching. The experimental results show that our WMDTW method improves the recognition accuracy compared with Latent Dirichlet Allocation （LDA） and conventional Dynamic Time Warping （DTW） methods. Key words： Kinect sensor; depth data; hand gesture recognition; dynamic time warping （DTW）