针对动态场景的快速视差图像预测方法  

Rapid Disparity Prediction for Dynamic Scenes

作　　者：陈宝文[1] 姜军[2] 谭旭[1] 

机构地区：[1]深圳信息职业技术学院软件学院,深圳518172 [2]中科院深圳先进技术研究院,深圳518055

出　　处：《科学技术与工程》2014年第18期94-98,共5页Science Technology and Engineering

基　　金：国家自然科学基金(71101096);深圳市科技计划(JC201005270364A)资助

摘　　要：要在机器人导航、视频监控和人机交互等领域,从立体图像序列中计算动态场景的三维结构时,大多数空-时立体方法和场景流方法,没有充分利用前一帧视差图像的信息计算下一帧视差图像,导致在计算每一帧视差图像时,对应点的搜索空间降低了视差计算的速度。针对这一问题,提出了一种有效的视差图像预测方法,利用稳定的基于角不变量特征的三维配准算法,估计立体装置在相邻两帧之间的自主运动参数;并建立了相邻两帧视差图像之间的变换关系。所提法能够快速计算具有时间一致性的视差图像序列,仿真和真实实验证明了其有效性。When computing 3D structures of dynamic scenes from stereo sequences, most spatiotemporal stereo and scene flow methods do not utilize the previous disparity map sufficiently to compute the next disparity map in robotic navigation, video surveillance and human-computer interaction, etc. This leads to the low speed of disparity computation because of the large searching space of correspondences. To solve this problem, an effective scheme was proposed to predict disparity maps from stereo sequences. In particular, a robust 3D registration algorithm was applied based on the angular-invariant feature to estimate the ego-motion of the stereo rig between consecutive fi-ames, and the transformation between consecutive disparity maps was presented. The scheme can produce a sequence of temporally coherent disparity maps rapidly. Synthetic and real experiments indicate the effectiveness of the new scheme.

关 键 词：深度预测 三维配准 三维重建 动态场景 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]