摘要
为减少背景特征对行为识别的影响,提出一种基于前景置信的人体行为识别方法。该方法在基于稠密时空兴趣点的行为识别基础上,结合像素前景置信估计对特征描述器进行加权分类,再利用词袋模型判别行为。融合运动、外观及视觉显著性的像素前景置信的引入,提高了算法处理复杂背景视频的能力。该方法在UCF50和HMDB51视频库中进行训练和测试,平均识别率为66.4%。
In order to reduce the effect of background features on action recognition,this paper proposes a foreground confidence-based human body action recognition method. On the basis of dense spatiotemporal interest points-based action recognition,the method combines the pixels estimation with foreground confidence to carry out weighted classification on feature descriptors. Then it uses the bag-of-words model to discriminate actions. The introduction of foreground confidence of pixels fusing the motion,appearance and visual saliency improves the ability of algorithm in dealing with complex background video. To be trained and tested on UCF50 and HMDB51 video datasets,the method obtains the average recognition rate of 66. 4%.
出处
《计算机应用与软件》
CSCD
2016年第10期191-193,197,共4页
Computer Applications and Software
关键词
行为识别
前景置信
加权分类
词袋模型
复杂背景
Action recognition
Foreground confidence
Weighted classification
Bag-of-words model
Complex background
