Abstract: In order to improve the performance of dynamic hand gesture recognition in complex real scenes for human-computer interaction, a novel method named Temporal Locality Sensitive Histograms of Oriented Gradients（TLSHOG） is proposed to represent sequential variations and changeable spatial postures of human hands. And it certainly achieves effective dynamic hand gesture recognition in real time. Firstly, a training set for machine learning is established by recording several videos of human hands using normal PC cameras. After that, a single-frame feature is proposed to describe the spatial feature of hands, and a Temporal Pyramid（TP） method is further applied to describe the trajectories of dynamic hand gestures. At last, a multidimensional Support Vector Machine（SVM） classifier is build, and all dynamic hand gestures in testing set are then correctly recognized. The experimental results indicate that this new method is highly discriminative of dynamic hand gestures and robust to effects of complex backgrounds and illuminations.

Key words: dynamic hand gesture, locality sensitive, histogram of oriented gradients, spatial-temporal features

摘要： 为了提高实际复杂场景的人机交互中动态手势识别的准确性和实时性，提出了一种时序局部敏感直方图（Temporal Locality Sensitive Histograms of Oriented Gradients，TLSHOG）特征新方法，用于描述手势运动的时序变化和空间姿态，实现了快速而精确的动态手势识别。采用普通网络摄像头获取手部的二维图像序列作为训练样本，然后构造单帧图像特征描述手部的空间姿态，并结合时间金字塔（Temporal Pyramid，TP）来描述手势运动轨迹的时空特征，运用多维支持向量机（Support Vector Machine，SVM）算法进行模型训练，对测试样本中的多种手势进行精确的分类。实验结果表明，该方法准确度高，实时性好，对于复杂背景干扰、光照强度变化有较强的鲁棒性。

关键词: 动态手势, 局部敏感, 梯度方向直方图, 时空特征