适用于单目视频的无标记三维人体运动跟踪

Markerless 3D Human Motion Tracking for Monocular Video Sequences

在无标记人体运动跟踪过程中,由于被跟踪目标缺乏明显的特征以及背景复杂而使得跟踪到的人体运动姿态与真实值偏差较大,不能进行长序列视频跟踪.针对这一现象,提出一种基于形变外观模板匹配进行单目视频的三维人体运动跟踪算法,其中所用的人体外观模型由三维人体骨骼模型及二维纸板模型组成.首先根据人体骨骼比例约束采用逆运动学计算出关节旋转欧拉角;然后利用正向运动学求得纸板模型中像素在三维空间中的坐标,将这些像素根据摄像机成像模型投影到二维图像中得到形变外观模板;最后采用直方图匹配得到人体运动跟踪结果.实验结果表明,该算法对于一些复杂的长序列人体运动能够得到较为理想的跟踪结果,可应用于人机交互和动画制作等领域.

In markerless human motion tracking, the reconstructed human motion pose has great difference from the ground-truth value due to the absence of obvious markers and complex background. To overcome this problem, we present an approach to track 3D human motion from uncalibrated monocular video sequences based on deformable appearance template matching where the human appearance model adopted in this research contains 3D human skeleton model and 2D cardboard model. Firstly, the Euler angles of joints are estimated by inverse kinematics based on human skeleton constraint secondly, the coordinates of pixels in the cardboard model in the scene are determined by forward kinematics, and the region of morphing appearance template in the image is obtained by projecting these pixels in the scene onto the image plane under perspective projection finally, the human motion can be tracked by histogram matching. The experimental results show that favorable tracking results on a number of long complex human motion sequences can be generated by the method. This approach can be applied to several areas such as human-computer interaction and human animation.