人体三维运动实时跟踪与建模系统

提出了一种新的人体三维运动实时跟踪与建模系统设计方法,并基于此实现了一套鲁棒的参考应用系统.针对人机交互等对跟踪精度要求不是很高的应用场合,系统在跟踪精确性和简易性与可推广性之间做了很好的折中.系统使用多个摄像头采集图像,实时计算场景深度信息,然后结合使用深度和颜色信息进行人体跟踪.应用一个简易的人体上半身三维模型,并使用基于颜色直方图的粒子滤波算法对头部和手部进行跟踪,从而恢复出模型的各个参数.系统以人脸检测和人手肤色聚类算法为初始化方法.大量实验证明,该系统能在复杂背景下进行人体上半身的跟踪和三维模型恢复,能进行完全自动的初始化,有较强的抗干扰能力和自动错误恢复能力.系统在2.4GHz PC机上能以25帧/秒的速度运行.

A Real-time 3D Human Tracking and Modeling System

We propose a new method for real-time 3D human tracking and modeling, and implement a robust reference application system. Focusing on applications, e.g. desktop human computer interface, which require low tracking accuracy, the method makes a good compromise between the accuracy and system simplicity. It uses multiple cameras and recovers the depth maps in real-time, and then uses both color and depth information for tracking. It adopts a simple 3D human upper body model, and uses color-histogram-based particle filtering to track human head and hands, and hereby reconstructs the 3D model. By using face detection and hand color clustering algorithms, the system initialization is fully automatic. Extensive experiments demonstrate that the system can robustly track and model human motion from complex background, and can automatically recover from losing of tracking object. The system runs at 25Hz on a 2.4GHz PC.