提出一种基于自然特征的增强现实并行注册方法。基于双核CPU双通道的注册策略，一个线程利用KLT算法进行跟踪注册，另一个线程使用Surf特征匹配算法进行宽基线注册校正。延迟单应阵概念的提出，解决了特征跟踪与匹配的线程同步问题。同时，提出特征关键帧的思想，提高了并行注册的准确性与稳定性。实验结果表明，该方法在实时的情况下，有效地解决了基于特征跟踪方法的误差累积问题，是一种准确、稳定、有效的自然特征注册方法。

A robust natural feature based parallel registration method is proposed. The method uses two threads for registration on a dual-core computer, one thread using KLT for tracking, while the other using Surf to reset the accumulated-errors. Delayed-homography is proposed to solve the problem of thread Synchronization. At the same time, the idea of key-frame, which is defined in this paper, increased the accuracy and stability of the registration. The experiment result proved that the multi-thread method solved the problem of accumulated-errors, and it is a robust feature based real-time registration method.