自适应融合颜色和深度信息的人体轮廓跟踪  

Adaptively Combining Color and Depth for Human Body Contour Tracking

作　　者：徐玉华[1,2] 田尊华[1,2] 张跃强[1,2] 朱宪伟[1,2] 张小虎[1,2] 

机构地区：[1]国防科学技术大学航天科学与工程学院,长沙410073 [2]图像测量与视觉导航湖南省重点实验室,长沙410073

出　　处：《自动化学报》2014年第8期1623-1634,共12页Acta Automatica Sinica

基　　金：国家重点基础研究发展计划(973计划)(2013CB733100)资助~~

摘　　要：采用活动轮廓对人体目标建模,提出一种新的水平集框架下自适应融合RGB-D图像的颜色和深度信息的人体轮廓跟踪方法.设计了一种基于超像素的局部自适应权重计算方法,自动确定深度信息在水平集演化中的重要性.基于深度信息的活动轮廓驱动外力包括由边缘生成的梯度向量流和由目标/背景深度模型生成的置信图,基于颜色信息的驱动外力由目标/背景颜色模型生成的置信图,这三种外力通过局部自适应权重融合,驱动活动轮廓向目标的边界演化.为了得到更加精确的目标轮廓和防止误差漂移,基于本文观察到的人体表面在深度图像中的两个特性,提出两个简单但有效的算法对水平集方法得到的结果进行精化调整.最后,通过实验验证了本文算法的优越性.In this paper, we present a novel human body contour tracking method, which combines color and depth cues of RGB-D images adaptively in the level set framework. We model the body object by the active contour. A superpixel-based locally adaptive weight map is designed to determine the importance of the depth cue in the evolution of the active contour. The depth-based external forces for the active contour are derived from the gradient vector flow （GVF） generated from the edges of the depth image and the confidence map generated from the depth models of the object/background, while the color-based external force is derived from the confidence map generated from the color models of the object/background. The three external forces are integrated by the adaptive weight to drive the active contour to evolve to the boundary of the object. To obtain a more accurate contour and to avoid error drifting, we propose two simple but effective algorithms based on the two properties of the human body surface in the depth image to refine the tracking result of the level set method. Experimental results demonstrate that our tracking method behaves more robustly and accurately than the latest depth-based body contour extraction method and the color-based contour tracking method on the datasets acquired in indoor environments.

关 键 词：轮廓跟踪 人体跟踪 活动轮廓 水平集 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]