摘要
目的提出了一个基于流形学习的动作识别框架,用来识别深度图像序列中的人体行为。方法从Kinect设备获得的深度信息中评估出人体的关节点信息,并用相对关节点位置差作为人体特征表达。在训练阶段,利用LE(Lalpacian eigenmaps)流形学习对高维空间下的训练集进行降维,得到低维隐空间下的运动模型。在识别阶段,用最近邻差值方法将测试序列映射到低维流形空间中去,然后进行匹配计算。在匹配过程中,通过使用改进的Hausdorff距离对低维空间下测试序列和训练运动集的吻合度和相似度进行度量。结果用Kinect设备捕获的数据进行了实验,取得了良好的效果;同时也在MSR Action3D数据库上进行了测试,结果表明在训练样本较多情况下,本文方法识别效果优于以往方法。结论实验结果表明本文方法适用于基于深度图像序列的人体动作识别。
Objective Human action recognition is a widely studied area in computer vision and machine learning; it has many potential applications including human computer interfaces, video surveillance, and health care. In the past decade, extensive research efforts focused on recognizing human action from monocular video sequences. Since human motion is articulated, capturing human joint characters accurately from video is a very difficult task. The recent introduction of real time depth cameras such as the Kinect sensor, give us the opportunity to use 3D depth data of a scene instead of pictures. In this paper, we present a manifold-based framework for human action recognition using depth image data captured from depth camera. Method With the recent release of Kinect sensor and the technology assessing skeleton joint position from depth image matured, recent research used 3D skeleton joint position information as human body representation and achieved good recognition performance. As we know, human action is composed of ordered posture set, and the difference between postures is only a few changes of 3D joints pairwise, most of the 3D information changes only little. In this paper, we estimated the 3D joint locations from Kinect depth images and use pairwise relative positions as the representation of human features. In the training phase, the LE （ Lalpacian eigenmaps ） is used to build action model in low dimensional space. In test phase, the nearest-neighbor interpolation technique is used to map test sequence to the manifold space, then measure the distance with the test sequence and the training data. A novel modified Hausdorff distance is used to measure similarity and fitness of the test sequence and the training data in the matching process. Result The recognition performance of the proposed method was evaluated from Kinect sensor dataset and the result confirmed the proposed method can work well in several experiments. We also tested the proposed method on the MSR Action3D dataset and achieved state of the art accuracy in our comparison with related work when the training set has many samples. Conclusion Manifold learning is an effective nonlinear dimensionality reduction method and low-dimensional motion models can be trained well when training sample size is large. We propose a novel human action recognition based on manifold learning in this paper. The experimental results show the effectiveness of the proposed method for human action recognition based on depth image sequence.
出处
《中国图象图形学报》
CSCD
北大核心
2014年第6期914-923,共10页
Journal of Image and Graphics
基金
国家自然科学基金项目(61303142
61173096
61103140)
浙江省自然科学基金项目(Y1110882
Y1110688
R1110679
LY13F020034)
教育部高等学校博士学科点专项科研基金项目(20113317110001)
浙江省教育厅一般科研项目(Y201330304)
