在动态手势特征提取和识别方面,利用运动学模式解决动态手势识别问题,在光流场基础上计算出散度模式,旋度模式,对称模式,反对称模式,梯度张量第二、第三主不变模式,应变张量第二、第三主不变模式以及自旋转张量第三主不变模式;进一步提出一种基于多实例学习的方法,将每一个动态手势的所有运动主模式构成一个动态手势词袋,将未知类型动态手势的运动主模式与词袋空间中对应运动主模式进行相似度计算,利用最近邻方法对手势进行识别.实验结果表明:基于多实例运动学主模式学习的动态手势识别方法取得了较高的识别率.%Compared to static gestures, dynamic gestures had some new characteristics. The problems of dynamic gestures recognition was spewed by using kinematics mode, such as divergence modes, curl modes, symmetric and ant-symmetric modes, the second and third principal invariant modes of the gradient tensor, the second and third principal invariant modes of the strain tensor and the third principal invariant modes of spin tensor; Further, a framework based on multi-instance learning was proposed, organize all these principle modes for each gesture were organized to a dynamic gestures bag-of-words, and the similarity between the mode of unknown type dynamic gestures and the all bag-of-words were calculated. Then, the nearest neighbor method was used to recognize the dynamic gestures. The experimental results show that the dynamic gestures recognition based on multi-instance kinematics features principal mode learning methods can obtain a higher recognition rate.
展开▼