基于平面颜色分布的增强现实自然特征注册算法

Augmented reality registration from nature features based on planar color distribution

针对增强现实中自然特征注册算法其自然特征的复杂和无规律而难以兼顾算法速度和精度的问题,提出一种基于平面颜色分布的自然特征注册算法.算法提取彩色连通域作为局部不变特征,计算简单的色调、几何信息作为特征描述,结合颜色分布在视角变换下的几何限制进行全局的匹配优化.该算法无须追踪连续帧间的运动特性,可以在独立帧上完成.算法采用Mikolajczyk标准库验证其注册效果,在800×600的图像尺寸下实现15帧/s的实时注册.与加速鲁棒特征(SURF)算法的对比表明:本算法能满足更苛刻的注册条件,并能保持较好的注册精度.动态视频的注册结果也表明算法面对动态模糊也十分鲁棒.

Aiming at the problem that augmented reality nature feature（or markerless）registration is difficult to guarantee both high accuracy and real-time speed because of the complexity and irregular of nature features,an algorithm based on planar color distribution was proposed.The algorithm takes colorful connected areas as invariant features,calculates their descriptors simply by hue and geometry information,and matches them by global optimization based on geometric constraints on color distribution under undefined view transformations.The algorithm does not rely on recursive tracking between frames.It achieves registration in every frame independently.Mikolajczyk＇s dataset was applied to test the algorithm,and the real-time registration for 800×600image sets over 15frames per second was achieved.The comparison against the SURF shows that the algorithm can keep good registration precision even facing much harder conditions.And the live video registration results demonstrate that the algorithm is robust to motion blur.