人体运动分析的实例学习方法

Example-based approach for human motion analysis from videos

目的面向实时、准确、鲁棒的人体运动分析应用需求,从运动分析的特征提取和运动建模问题出发,本文人体运动分析的实例学习方法。方法在构建人体姿态实例库基础上,首先,采用运动检测方法得到视频每帧的人体轮廓;其次,基于形状上下文轮廓匹配方法,从实例库中检索得到每帧视频的候选姿态集;最后,通过统计建模和转移概率建模实现人体运动分析。结果对步行、跑步、跳跃等测试视频进行实验,基于轮廓的形状上下文特征表示和匹配方法具有良好的表达能力;本文方法运动分析结果,关节夹角平均误差在5°左右,与其他算法相比,有效提高了运动分析的精度。结论本文人体运动分析的实例学习方法,能有效分析单目视频中的人体运动,并克服了映射的深度歧义,对运动的视角变化鲁棒,具有良好的计算效率和精度。

Objective Human motion tracking from monocular image sequences is a challenging work in computer vision. It also has many penitential applications, such as human computer interface, computer animation, and intelligent video surveillance. Methodologies of example-based human motion tracking from monocular camera are explored in this study to meet timeliness, accuracy, and reliability requirements of human motion tracking for real applications. We focus on two main aspects： visual feature extraction and human motion modeling. Method Based on an example database that is constructed with motion capture data, an example-basod approach for human pose estimation from monocular image sequences is proposed. First, we use a motion detection method to extract human region from images. Then, an edge-tracking method is used to detect human silhouette from human region. Second, shape context is used to describe the human silhouette detected from video frames, and candidate poses are obtained from the example database by silhouette matching. Third, we build probability and statistical model of human motion and conducted pose estimation from these candidates. Result Experimental results on walking, running, and jumping videos demonstrate that shape context-based silhouette representation and matching method can effectively extract human visual feature from image. Compared with other methods, the proposed method can tackle orientation ambiguity problem effectively. Moreover, it is invariant to viewpoints. Conclusion In this paper, we proposed an example-based method for human motion analysis. Shape context is used for visual feature extraction and matching. Probability and statistical model of motion are used for pose estimation. Experimental results on different types of motion demonstrate that the proposed method can analyze 3D pose from videos effectively, thereby increasing the efficiency and accuracy of human motion analysis. Moreover, the proposed method can solve the orientation ambiguity problem, and is invariant to viewpoints.