基于不确定分解子空间约束光流的柔性目标点跟踪

提出一种基于不确定分解子空间约束光流的柔性目标点跟踪算法,通过不确定分解理论将多帧多点光流估计矩阵变换到各向同性的具有超球状方差分布的空间中,在此变换空间中引入SVD分解得到最小均方意义下的子空间光流逼近,有效地减轻了传统L-K算法的孔径问题、深度不连续点的估计问题及长序列视频的漂移问题。标准测试序列和消费类USB摄像头采集的测试序列的实验结果都证明算法能有效地跟踪长视频序列中具有2-D和1-D甚至基本没有纹理的具有退化结构的柔性目标点。结果还可应用于柔性目标理解并可以作为半稠密的点对应来解决SFM中的对应点求解问题。

A Novel and Effective Algorithm for Robust and Precise Visual Tracking of Nonrigid Objects

Traditional methods of optical flow estimation have a number of well-known problems that make it very difficult to implement robust,precise and fast optical-flow-based tracking of nonrigid objects.We now present a novel algorithm that can effectively achieve such desirable tracking.In the full paper,we explain our novel algorithm in detail;in this abstract,we just add some pestinent remarks to listing the three topics of explanation:(1) the essentials of uncertainty factorization,(2) the essentials of subspace constraint and(3) the optical-flow-based tracking of nonrigid objects using uncertainty factorization and subspace constraint;in topic 1,we point out that,since the Hessian matrix of traditional L-K algorithm can be considered as the inverse covariance of the optical flow,we can introduce uncertainty factorization to convert the estimation problem from a hyper-ellipse space one into a hyper-sphere space one to tackle the drift problem of points with 1D or little texture;in topic 2,we point out that,since multi-frame multi-point optical flow is of low dimensions,we can use SVD to introduce subspace constraint into the warped space to get more robust estimation;in topic 3,uncertainty factorization and subspace constraint are fused together to make the proposed nonrigid tracking more robust and precise through executing an 11-step procedure proposed by us.Experimental results of both the recorded sequence of ordinary commercial camera,whose image quality is low, and the standard test sequence,shown respectively in Figs.2 and 1 in the full paper,confirm the robustness and effectiveness of our novel tracking algorithm.