Implementation of Optical Flow Analysis System Equipped with Reliability Evaluation

1.新的光流检测方法通过引入图像噪声的精确数学模型，提出了一种计算光流并定量评价其可靠性的方案。三维重建的鲁棒计算方案如果可以通过迭代计算一个称为“流矩阵”的矩阵，则可以重建三维结构。这些迭代的收敛性对真实图像数据的收敛性比对人工噪声的模拟数据的收敛性差。提出了一种保证收敛性的新方案。事实证明，强调光流三维分析的统计理论可以广泛应用于涉及图像和传感器数据的一般几何估计问题。我们提出了“最优拟合”的一般理论，其中理论精度界是“Cramer-Rao下界”的一种特殊形式。几何信息准则扩展了统计学中使用的AIC，我们定义了“几何信息准则（AIC）”，作为预测从噪声图像数据中期望鲁棒三维重建的程度的度量。我们将其应用于立体图像分析、运动图像分析、图形界面等多种问题，证实了我们的形式主义的有效性。

1.New optical flow detection methodIntroducing a precise mathematical model of image noise, we presented a scheme for computing optical flow and evaluating its reliability in quantitative terms.2.Robust computational scheme for 3-D reconstructionThe 3-D structure is reconstructed if a matrix called the "flow matrix" can be computed by iterations. The convergence of these iterations turned out to be worse for real image data than for simulated date with artificial noise. We presented a new scheme to guarantee the convergence.3.Mathematical theory of geometric inference from noisy dataIt has turned out that the statistical theory that underlines our 3-D analysis of optical flow can be applied to a wide range of general geometric estimation problems involving image and sensor data. We presented a general theory of "optimal fitting", in which a theoretical accuracy bound was derived as a special form of the "Cramer-Rao lower bound".4.Geometric information criterionExtending the AIC used in statistics, we defined the "geometric information criterion (AIC)" as a measure for predicting to what extend we could expect robust 3-D reconstruction from noisy image data. We applied it to a various problems including stereo image analysis, motion image analysis and graphic interface, confirmed the effectiveness of our formalism.