弱纹理环境双目视觉稠密视差鲁棒估计方法

精确稠密视差估计是立体视觉系统恢复观测场景三维信息的关键。从立体视觉在机器人环境感知的实际应用角度出发,提出了对于弱纹理、阴影和遮挡等关键影响因素,具有良好鲁棒性、精度和处理速度的稠密视差图估计算法。针对弱纹理、阴影和深度不连续的问题,设计了基于灰度相似度概率的置信度传播算法,结合视差平滑约束,以期实现较高精度的视差初值快速估计。由视差级数定义的消息向量通过异向平行迭代进行传播,消息向量包含表征像素点灰度相似性和平滑性的能量信息,通过全局能量函数的迭代收敛,快速获得视差初始估计。根据独立连通区域通常具有相似纹理特征和视差一致性的先验知识,提出了基于Mean-Shift聚类分割算法和参数空间投票自适应视差近似面估计算法,进行稠密视差的精细优化估计。利用具有不同弱纹理特征的5组标准测试图像、4组室内环境实际图像、4组室外环境实际图像和4组月面模拟特殊光照环境的实际环境图像进行了测试实验,实验结果表明了本文算法的良好鲁棒性和有效性。

vision under textureless environment

Precise dense disparity estimation is the key for stereo visual system to recover three-dimensional information of observation scene.From practical application perspective of stereo vision in robot environment perception,a dense disparity figure estimation algorithm having good robustness, accuracy and processing speed to key influence factors (texturelessness, shadow and blocking etc.) was proposed.Aimed at texturelessness, shadow and uncontinuous, belief propagation algorithm based on gray-scale similarity probability had been designed to realize rapid and accurate estimation of initial value of disparity by combining with disparity smoothness constraint.The message vector defined by disparity class was propagated through anisotropic diffusion and parallel iteration.Message vector included energy information representing gray-scale similarity and smoothness of pixel point.Initial estimation of disparity could be gained rapidly through iteration convergence of global energy function.According to the priori knowledge that independent connected area generally had similar textural features and disparity conformance, parameter space voting self-adaption disparity approximation surface estimation algorithm on the basis of Mean-Shift clustering partitioning algorithm was proposed to perform fine optimization estimation of dense disparity.5 groups of standard test image having different textureless features, 4 groups of actual image under indoor environment, 4 groups of actual image under outdoor environment and 4 groups of actual environment image under special lighting environment through selenographic simulation were utilized to perform test experiment and experimental result shows that the proposed algorithm has good robustness and effectiveness.