Reconstructing 3D Motion And Structure From Straight-line Optical Flow

It is a major topic in computer vision to recover the movement and structure of a 3D moving object through the analyses of dynamic image sequences. This paper involves the research of reconstructing motion and structure of 3D rigid body with straight-line feature from straight-line optical flow.A tracking lines method based Hough Transform was proposed. Rather than tracking lines in image space directly, the method makes use of duality theory in Hough Transform in order to transform line-tracking in image space into point-tracking in Hough space. The first step designing that adopt the Kalman filter to preliminary track the motion target, then accurate characteristic point location is searched by using build-up moulding board and image matching correlation modulus.This paper first defines the derivative to the three parameters of 2D line as the straight-line optical flow, and Based on straight-line optical flow equation, a linear algorithm of reconstructing 3D motion and structure parameters from straight-line optical flows is presented which uses two couple of straight-line optical flows in three consecutive image frames of image sequence to establish a linear equations about rotation parameters and gains the 3D rotation motion parameters and the focus of camera by solving the equations. The algorithm requires little 2D straight-lines and the it is easier to be realized.Based on straight-line optical flow equation, a B-P Artificial Neural Network model of reconstructing 3D motion and structure parameters from straight-line optical flows in long sequence images of rigid body with straight-line feature is established in this paper. When given network training rate, frequency of training and training goals after error, and by using of artificial neural network to its own strong adaptability and robustness, the B-P artificial neural network uses back-propagation algorithm, and continuously adjust the output value and network threshold to achieve the target output. When network training to achieve convergence, and this time the weight and threshold is seeking the rigid rotation parameters. Some simulated experiments data shows the model is steady and robust.