Reconstructing 3D Motion And Structure With Straight-line Optical Flow Based On Sequences Of Monocular Images

The purpose of this paper investigates a theory and method for three-dimensional (3D) recovery of structure and motion from straight-line optical flow in a sequence of monocular images containing moving rigid objects.This paper proposes a new representation available for central projection model which describes a 3D line with four parameters in clustering mode. Based on the representation, differential coefficients of two parameters of a beeline are defined as straight-line optical flow. Under the central projection, deduce the relationship between the rotational components of rigid motion in three dimensions and optical flow of 2D lines in projection plane,and express the relationship with straight-line optical flow equations set. Using the two group of optical flow of two lines in successive frames of a monocular image sequence, this paper can solve the equations set to reconstruct the rotational components, translational components and 3D structure of a rigid. The features of this model involve fewer lines, more algorithms available and easy to carry out.Based on the relationship has been founded between straight-line optical flow and rigid motion, establish the model tracking lines based on Extended Kalman Filter, which can track the lines in the image sequence at real time stably. The study also establishes an Adaptive Linear Neural Network (ALNN) algorithm for solving beeline optical flow equations. The algorithm transforms straight-line optical flow into input and output of the ALNN, weighted value and bias of network are considered as part rotational parameters of the rigid approximately after trains the network for some time. And then, solves translational parameters and coordinates of 3D lines, thereby 3D reconstruction would come true.This paper also presents a genetic algorithm (GA) based on straight-line optical flow to recovery motion and structure of rigid. The algorithm regards motion parameters as its individuals and the equations as fitness function to evolve, then chooses excellent individual as motion parameters value searching for. In both ALNN and GA, if only can capture and track at least two lines in images, the rotational components, translational components and relative depth of rigid can be reconstructed successively. Results are tested for validity of the algorithm in real and synthetic image sequences, and validate robustness and precision of it.