张彦彬, 陈晓春. 基于特征空间切分建模的变形手势跟踪算法[J]. 机器人, 2018, 40(4): 401-412. DOI: 10.13973/j.cnki.robot.170513
引用本文: 张彦彬, 陈晓春. 基于特征空间切分建模的变形手势跟踪算法[J]. 机器人, 2018, 40(4): 401-412. DOI: 10.13973/j.cnki.robot.170513
ZHANG Yanbin, CHEN Xiaochun. The Deformed Gesture Tracking Algorithm Based on Feature SpaceSegmentation Modeling[J]. ROBOT, 2018, 40(4): 401-412. DOI: 10.13973/j.cnki.robot.170513
Citation: ZHANG Yanbin, CHEN Xiaochun. The Deformed Gesture Tracking Algorithm Based on Feature SpaceSegmentation Modeling[J]. ROBOT, 2018, 40(4): 401-412. DOI: 10.13973/j.cnki.robot.170513

基于特征空间切分建模的变形手势跟踪算法

The Deformed Gesture Tracking Algorithm Based on Feature SpaceSegmentation Modeling

  • 摘要: 为解决人机交互中手势形变和无规律运动带来的跟踪难题,提出了一种基于特征空间切分建模的非参数核密度估计算法来实现手势跟踪.首先,在检测模块中利用AdaBoost分类器检测图像中手势的存在,将检测到的手势位置信息传送给跟踪模块,该模块精确提取手势目标从而对其颜色建模.然后,利用目标的颜色模型对各帧图像进行后验概率密度估算,获取运动目标的概率密度图像,将其分解成手势运动区和同色干扰区.最后,对同色干扰区采用混合高斯建模来削弱同色目标的干扰.当目标丢失时启动再检测模块,并利用贝叶斯分类器与方差分类器实现手势目标重检.实验结果表明,该算法通过对特征空间切分建模以及不同分类器的级联解决了变形手势跟踪的同色干扰与再检测难题.该算法提高了跟踪的准确率(>81.5%),适合于非刚性物体做无规则运动的复杂场景.

     

    Abstract: In order to solve the tracking problem caused by the deformation and irregular motion of gestures in the process of human-computer interaction, a nonparametric kernel density estimation algorithm based on the feature space segmentation modeling is proposed. Firstly, the AdaBoost classifier in the detection module is used to detect the presence of gestures in the image, and the gesture position information is sent to the tracking module, which accurately extracts the gesture target for color modeling. Then, the probability density image of motion target is obtained by estimating the posterior probability density of each frame image using the color model of the target, which is decomposed into the gesture motion area and the similar color interference region. Finally, the mixed Gauss model is used to weaken the interference of close color objects in the similar color region. The redetection module is started when the target is lost, and the gesture position can be detected by adopting the Bayesian classifier and the variance classifier. The experimental results show that the proposed method resolves the similar color interference and redetection problem in deformed gesture tracking by segmenting the feature space and cascading different classifiers. The proposed algorithm improves the tracking accuracy (>81.5%) and is suitable for complex scenes involving the irregular motion of nonrigid objects, and has a high tracking accuracy.

     

/

返回文章
返回