This paper focuses on online scene learning and fast camera relocalisation which are two key problems currently limiting the performance of wide area augmented reality systems. Firstly, we propose to use adaptive random trees to deal with the online scene learning problem. The algorithm can provide more accurate recognition rates than traditional methods, especially with large scale workspaces. Secondly, we use the enhanced PROSAC algorithm to obtain a fast camera relocalisation method. Compared with traditional algorithms, our method can significantly reduce the computation complexity, which facilitates to a large degree the process of online camera relocalisation. Finally, we implement our algorithms in a multithreaded manner by using a parallel-computing scheme. Camera tracking, scene mapping, scene learning and relocalisation are separated into four threads by using multi-CPU hardware architecture. While providing real-time tracking performance, the resulting system also possesses the ability to track multiple maps simultaneously. Some experiments have been conducted to demonstrate the validity of our methods.
Keywords: adaptive random trees; augmented reality; registration; scene recognition; wide-area.