摘要
提出一种基于多特征融合的动作识别方法,利用随机森林学习框架融合RGB-D图像序列中的深度特征和时空特征。从深度图像序列中获取人的关节点位置信息,利用关节点坐标提取两种新的深度特征——位移特征和部件中心特征,共同描述人体三维结构信息。从RGB图像序列中提取稠密轨迹,保留前景内的轨迹排除背景干扰,利用词袋模型构建时空特征。最后,采用鲁棒高效的随机森林学习框架融合两种互补的特征。在MSR Daily Activity3D数据集上的实验结果表明,所提出的方法和特征能够有效地识别RGB-D图像序列中人的动作。
This paper proposed a novel action recognition method based on multi-feature fusion.In this method,the spatial-temporal features and depth features were merged in a random forest framework.The human body joint coordinates obtained from depth image sequences were processed into displacement feature and part-center feature as two new depth features.We applied these two depth features to describe the threedimension structure of human.We densely sampled the trajectories from RGB image sequences,and utilized the foreground detection approach to reduce the effect of complex background.Then spatial-temporal features were constructed by the Bag-of-Words model with trajectories from the foreground.Finally,the robust random forest framework fused both the spatial-temporal features and the depth features for recognizing human actions in RGB-D image sequences.Experimental results on MSR Daily Activity 3D dataset demonstrated the effectiveness of the proposed method.
出处
《沈阳航空航天大学学报》
2017年第2期55-65,共11页
Journal of Shenyang Aerospace University
基金
国家自然科学基金(项目编号:61170185
61602320)
辽宁省博士启动基金(项目编号:201601172)
辽宁省教育庁一般项目(项目编号:L201607
L2014070)
沈阳航空航天大学校博士启动基金项目(项目编号:15YB37)
关键词
人的动作识别
特征融合
随机森林
human action recognition
multi-feature fusion
random forest
