3 Dimensional Shape Recovery from Image Sequences by Non-Linear Optimization

研究了一种利用非线性优化方法从图像序列中恢复三维形状的方法。(1)提高特征点的检测和跟踪精度我们开发了一种在图像序列的仿射和深度几何约束下检测和跟踪特征点的方法。首先对包含运动视差的一对图像进行仿射变换，然后在仿射几何约束下检测图像之间的对应点。在恢复初始三维形状后，应用深度几何约束去除离群点并检测新的对应点。经过多次迭代，仿真研究得到25帧图像中含有11277个点的三维模型，真实图像中含有15751个点的三维模型。实验结果表明，该方法可以有效地从图像序列中恢复出密集、精确的三维模型。(2)提高计算非线性优化的稳定性将二维图像序列中物体形状和相机运动的三维恢复表述为非线性优化问题，其初始值对于避免搜索空间中称为局部最小值的错误解非常重要。我们开发了一种确定恢复物体初始值的方法，该物体的图像是由围绕物体进行圆周运动的相机拍摄的。提出了一种从假解中搜索最优解的方法。合成图像和真实图像的实验结果表明，该方法在给定的初始值下可以稳定地进行三维恢复。

We have investigated a method for recovering 3 dimensional shapes from image sequence by using non-linear optimization.(1) Improvement of precision for detecting and tracking feature pointsWe have developed a method for detecting and tracking feature points under the affine and depth geometric constraints from the image sequences. First, the affine transform is applied to a pair of images including motion disparity, and then the corresponding points between the images are detected under the affine geometric constraint. After recovering the initial 3D shape, the depth geometric constraint is applied to remove the outlier points and to detect new corresponding points. After several iterations of this procedure, 3D model with 11,277 points was recovered from images of 25 frames in a simulation study, and that with 15,751 points was from 25 frames using real images. These experimental results have shown the effectiveness of our method to recover the 3D dense and precise models from image sequences.(2) Improvement of stability for computing non-linear optimization3D recovery of object shape and camera motion from 2D image sequence is formulated as a non-linear optimization problem, and the initial values are very important to avoid the false solutions called local minima in the search space. We have developed a method to determine the initial values for recovering object whose images are taken by a camera that moves on circular path around the object. We also proposed a procedure to search the optimal solution from the false one. Experimental results with synthetic and real images showed that 3D recovery was stably performed with the initial values given by our method.