融合多特征和压缩感知的手势识别

Hand Posture Recognition Based on Multi-feature and Compressive Sensing

基于压缩感知理论,提出了一种新的手势识别方法,考虑到单个特征的局限性,结合Zernike矩和HOG描述符从全局和局部角度描述手势外观和形状.训练阶段提取手势训练图像的Zernike矩和HOG特征构建字典,识别阶段提取待测样本特征,将其表示成相应训练字典的稀疏线性组合,采用求解l1范数的最优化问题实现分类.实验结果证明,和目前应用较广的手势识别方法相比,该方法具有较强的竞争性,而且通过融合两种形状特征,对光照、尺度、旋转等变化更具鲁棒性.

A method was introduced for hand posture recognition based on compressive sensing. Con- sidering the limitations of a single feature, Zernike moment and HOG descriptors were fused to improve the robustness. Firstly, we constructed training dictionaries according to the characteristics, then the can- didate target was expressed as a sparse combination of the corresponding training dictionary, and classifica- tion results were done through solving a l1-norm based optimization problem. The proposed method can take full advantage of each feature, which is robust to rotation, noise and varying illumination. Experi- ment results show that the algorithm is competitive to the state-of-the-art hand posture recognition meth- ods, and is suitable for real-time application.