Abstract:

In view of the requirements of long-term tracking registration in Augmented Reality（AR） system and the problems of registration failure easily caused by complex environment such as scale change, illumination change and internal and external rotation of registration area in 3D tracking registration, an AR long-term 3D registration method is proposed to improve Tracking-Learning-Detection（TLD） tracking and Oriented FAST and Rotated BRIEF（ORB） detection. The TLD long-term robust tracking algorithm is used to track the registration area, and the adaptive dynamic detection strategy is used to optimize the real-time performance and precision of the tracking learning detection structure mechanism of the TLD algorithm. Then, the ORB algorithm is used to detect the registered location and calculate the 3D registration matrix. Finally, the virtual information is superimposed on the registration location. The experimental results show that compared with TLD algorithm, the improved TLD long-term tracking algorithm can improve the accuracy of One-Pass Evaluation（OPE） by about 5% and the tracking rate of the location to be registered is increased by about 2 times, which makes the AR system have better stability and robustness in the complex environment.

摘要：

针对增强现实系统长时跟踪注册需求以及三维跟踪注册时待注册区域尺度变化、光照变化、内外旋转等复杂环境易导致注册失败问题，提出改进TLD跟踪及ORB特征检测的增强现实长时三维注册方法。通过采用TLD长时鲁棒跟踪算法对待注册区域跟踪，并利用待注册位置自适应动态检测策略对TLD算法跟踪-学习-检测的结构机制的实时性和精确度进行优化；利用ORB算法对待注册位置进行特征检测及三维注册矩阵计算；将虚拟信息叠加至待注册位置。选取长时视频序列进行跟踪、检测及注册仿真实验。实验结果表明，改进TLD长时跟踪算法相比TLD算法在OPE精确度方面提升约5%，对待注册位置的跟踪速率提升约2倍，使得AR系统在复杂环境下的长时跟踪注册具有更好的稳定性和鲁棒性。

关键词: TLD算法, ORB算法, 长时跟踪, 注册区域检测, 长时注册