基于多特征融合的人体行为识别

HUMAN ACTION RECOGNITION BASED ON MULTI-FEATURES FUSION

为了提高视频序列中人体行为的识别率和增强在复杂环境下的适用性,通过选取人体行为区分度较高的运动方向特征、形状特征和光流变化特征进行行为描述,提出一种基于运动方向直方图(MOH)特征、2D-SIFT特征和光流方向直方图(HOOF)特征相结合的人体行为识别方法。改进运动方向直方图特征,使其在有符号梯度空间下对人体全局运动方向具有更为鲁棒的表示。使用视觉词袋模型既解决了不同动作提取的兴趣点点数不同的问题,又实现了局部特征的有效融合。实验在Weizmann数据库和KTH数据库上识别率分别高达97.83%和91.38%,并具有较好的鲁棒性。

In order to improve the human actions＇ recognition rate in video sequence and enhance the applicability in complex environment, by selecting the features of higher differentiation in regard to human actions such as motion orientation, shape and optical flow change for representing the actions, we proposed a new human actions recognition algorithm which is based on the combination of motion orientation histograms （MOH） feature, 2D-Sift feature and histograms of oriented optical flow （HOOF） feature. The MOH feature was refined and was made to have more robust representation on body＇ s global motion direction in symbol gradient space. We used visual bag-of-word model to have solved the problem of various numbers of interest points extracted from different actions while achieved effective fusion of local features. Experiments performed on Weizmann and KTH databases showed that this algorithm had high recognition rate up to 97.83% and 91.38% and had better robustness as well.