【作者】 杨敏； 梁玮； 李善青； 贾云得；

【Author】 Min Yang,Wei Liang,Shanqing Li and Yunde Jia School of Computer Science,Beijing Institute of Technology,100081,Beijing

【机构】 北京理工大学计算机学院；

【摘要】 本文提出了一种基于交互式学习的目标建模和识别方法用于穿戴计算,使穿戴视觉系统具有了和穿戴者同步的感知能力。该方法通过穿戴视觉系统的多模态标注接口,允许穿戴者使用手势和声音标注场景中的感兴趣目标,并自动地为这些目标建立模型。在目标模型的指导下,该方法能实时地检测并识别当前场景中的感兴趣目标,并将识别结果显示给穿戴者;穿戴者同时能根据识别结果选择为目标添加训练样本,进而更新目标模型,形成交互式的学习方式。实验结果表明了本文提出的方法的有效性。更多还原

【Abstract】 This paper presents an object modeling and recognition method based on interactive learning for wearable computing.Using a multimodal labeling interface,this method can create robust model for the objects of interest labeled by the wearer’s gesture and audio.The model can be updated in an interactive way.On the one side,the proposed method can recognize the objects in the current scene by matching the object model to the scene image.On the other side,the wearer can choose to add new object samples in order to update object model by judging the recognition results.This method is composed of three key steps.Firstly,a multimodal labeling interface is adopted to allow the wearer label the concerned objects in a friendly way,which can provide labeled samples for the learning phase of object model.Secondly,object model is learned by choosing several template images which can represent the object under different conditions.Finally,object recognition is achieved by robust template matching algorithm.The experimental results demonstrate the effectiveness of the proposed method.Besides,we have come up with an application service on the wearable vision system.The application illustrate that a common focus of attention is shared by a wearer and the wearable visual system is established for natural human computer interaction by using the proposed method.更多还原