基于关键帧的交互行为识别

Two-person interaction recognition based on key frames

微软发布的Kinect传感器相比传统相机,不仅能获取颜色图像,还能得到深度图像,这对人类行为分析提供了一种新的解决方案。因此,基于Kinect创建了八种类型双人交互行为数据集。为了减少原始序列数据并降低计算复杂度,基于DTW模型和运动能量准则提出了一种新的关键帧提取方法。实验结果表明,提出的方法能够准确、全面地反映内容梗概;同时指出选择五个关键帧,相比其他关键帧提取个数,不仅保证与完整序列的识别结果相近,也更有效地减少了原始序列数据存储量。

Compared with traditional 3D scanners,the Kinect is a new kind of range camera researched by Microsoft; it has advantages of synchronous acquisition of RGB image and depth image and provides a new way for human action recognition.Thus,this paper created 8 types of two-person interaction recognition dataset. In order to have a more compact representation of original sequence data and simplify computation complexity,this paper proposed a new key pose extraction method based on DTW model and motion energy criteria. The experimental result demonstrates that the proposed method is accurate and comprehensive enough to generalize the interaction sequence. Also,it suggests that five key poses is the best choice to represent the whole sequence,assuring the high recognition accuracy.