复杂人机交互场景下的指势用户对象识别

Pointing User Recognition in Human-Computer Interaction with Cluttered Scene

采用指势进行人机交互,可充分发挥人类日常技能,摆脱常规输入设备束缚.实现自然的指势人机交互的关键是,如何从复杂的人机交互场景中有效提取指势用户对象,提出了基于时/空运动特征的指势用户对象识别新方法.基于多尺度小波变换在时/空域所具有的优异局部化特性,从复杂场景中提取前景运动对象,克服环境条件约束以及动态环境变化及先验假设等不足;基于多尺度小波变换的梯度积分图方法,获取稳定可靠的指势手部HOG特征,采用机器学习方法,对上述特征向量分类,并基于指势手与指势用户对象的空间关联性识别指势用户对象.通过实验对比,结果表明本文方法有效、可行.

Human being daily skill can be exerted fully and bondage can be delivered efficiently in which people use ordinary equipment as an input way if pointing gesture is used for human-computer interaction（ HCI）. One of key problems is how to reliably recognize pointing user from HCI scene with cluttered background. A novel method has been developed based on spatio-temporal motion. According to multi-scale wavelet transform（ MWT） with outstanding local characteristics both in spatial and temporal domains,it is adopted to extract foreground motion subject from cluttered scene. Some disadvantages are overcome including restrictions in environment conditions,dynamic environment variation,and a priori assumption. MWT based gradient integral graph is used to get some HOG feature vectors in pointing hand which are classified and learnt based on machine learning. Pointing user is recognized according to spatial relationship between pointing hand and its corresponding subject. Experimental results have been shown that the proposed method is efficient and viable.