基于局部和全局特征视觉单词的人物行为识别

基于视觉单词的人物行为识别由于在特征中加入了中层语义信息,因此提高了识别的准确性。然而,视觉单词提取时由于前景和背景存在相互干扰,使得视觉单词的表达能力受到影响。提出一种结合局部和全局特征的视觉单词生成方法。该方法首先用显著图检测出前景人物区域,采用提出的动态阈值矩阵对人物区域用不同的阈值来分别检测时空兴趣点,并计算周围的3D-SIFT特征来描述局部信息。在此基础上,采用光流直方图特征描述行为的全局运动信息。通过谱聚类将局部和全局特征融合成视觉单词。实验证明,相对于流行的局部特征视觉单词生成方法,所提出的方法在简单背景的KTH数据集上的识别率比平均识别率提高了6.4%,在复杂背景的UCF数据集上的识别率比平均识别率提高了6.5%。

Human Action Recognition by Visual Word Based on Local and Global Features

Different from the method based on low-level features,the human action recognition based on visual word adds mid-level semantic information to features and then improves the accuracy of recognition.For complex background or dynamic scenes,the efficiency of visual words might deteriorate.We proposed a new method which is a combination of local and global feature to generate visual words.Firstly,our approach uses saliency map to detect the rectangles around human.And then inside these rectangles,3D-SIFT will be calculated around interest points detected from dynamic threshold matrix to describe local features.We also added HOOF to describe the global motion information.These visual words provide the important semantic information in the video such as brightness contrast,motion information,etc.The performance of this method in action recognition can be improved 6.4% on KTH dataset and 6.5% on UCF dataset compared with state-of-the-art methods.The experiment results also indicate that our visual dictionary has more advantages in both simple background and dynamic scene than others.