宽基线主动视觉中感兴趣目标的对应技术

Correspondence of Object-of-Interest in Wide Baseline Active Vision

在主动视觉系统中,通常需要多个代理对同一场景中的感兴趣目标进行协同处理,以提高系统智能分析感兴趣目标的能力。其中,基于多视几何关系解决感兴趣目标的对应问题是协同处理的基础。一方面,主动视觉系统一般工作在宽基线条件下,这增加了对应问题描述的复杂性;另一方面,主动视觉系统以最佳视角观察目标,因此摄像头需做实时的姿态调整,由此导致的视间几何关系变化进一步加深了对应问题的解决难度。本文基于仿射不变的几何特征,建立宽基线条件下的多视几何关系,并针对频繁使用几何特征不能满足主动视觉系统实时要求的问题,提出一种快速更新多视几何关系的方法,并在多视几何约束下实现对应感兴趣目标的鲁棒标识。实验结果表明,该方法能解决宽基线主动视觉系统中感兴趣目标的复杂对应问题,并能达到实时要求。

In active vision system, intelligent analysis of objects-of-interest（OOI） in 3D scene generally needs multi-agent cooperation, which depends heavily on the accuracy of multi-view geometry computation and the robust correspondence of OOI. In most cases, active vision system captures multiple views under wide-baseline stereo model, which results in distinct affine distortions between views and complex description of correspondence problem. Furthermore, the difficulty of depicting correspondence problem in active vision is aggravated when the cameras perform real-time pose adjustment for active OOI tracking with the best viewpoint. For abovementioned issues, this paper proposes a method to achieve real-time multi-view geometry updating on the basis of constructing geometric feature invariants. Then a robust OOI correspondence algorithm is advanced with the acquired multi-view geometry constraint. The experimental results demonstrate that our method can obtain robust OOI correspondence under wide baseline stereo model and be suitable for real-time application in active vision system.