应用在线随机森林投票的动作识别

提出了基于在线随机森林投票识别人物动作类别的方法。建立了在线随机森林投票模型。通过在线训练和在线检测两部分进行了算法研究,提高了检测人物动作类别的准确率。基于人物动作在时间和空间上有重要信息,该方法首先通过提取图像立体块的lab色彩空间值、一阶差分、二阶差分以及大位移光流特征值在线训练随机森林;训练结束后,形成强分类器,利用分类器对检测图像进行投票,生成动作空间图;最后,在动作空间图中寻求最大值,判断检测图像的动作类别。验证结果表明在低分辨的视频图像中,本方法能够确定人物的动作类别,对Weizmann数据库和KTH数据库的识别率分别为97.3%和89.5%,对UCF sports数据库的识别率为79.2%,动作识别准确率有所提高。该方法增加了光流能量场特征表述,将原始投票理论拓展至三维空间,并且采用向下采样的方式更新结点信息,能够判断人物动作类别,为智能视频技术提供了有效的补充信息。

Action recognition based on on-line random forest voting

An action recognition method for people is proposed based on on-line random forest voting to judge the action classification. The on-line random forest voting model is established and its algorithms are researched through the two parts consisting of on-line training and on-line detection to improve the precision of the action classfication. As people action shows important information in both space and time, the method firstly trains the random forests in line by extracting 3D image features containing a lab color space , the first order difference, the second order difference and displacement optical flow. After training, a strong classier is formed. Then, the classifier is used to vote for detection images to produce an action space map. Finally, by seeking the maximum in the map, the category of action in the detection images is complemented. Experimental results indicate that the method determines the category of people action in the low resolution video images. The accurate rates of the Weizmann data, the KTH data and the UCF sport data are 97.3%, 89.5%, and 79.2%, respectively. These results show that the accuracy of action recognition is improved. Moreover, the model proposed adds the feature representation of light flow energy field, expands the traditional forest voting theory to a 3D space, and uses to update information. It improves the stability and the reliability and will be of potential application in the intelligent video surveillances.