Robust methods for structure from motion using video sequences.

摘要

This dissertation presents robust methods for solving the structure from motion problem using video sequences. Two methods are proposed to handle cases where existing structure from motion methods have difficulties in practical applications. The first method fuses inertial information obtained from a camera-mounted inertial sensor into the structure from motion algorithm using an extended Kalman filter. The second one uses a sequential Monte Carlo method such as sequential importance sampling to compute an approximation to the posterior distribution of the camera motion and scene structure. Furthermore, approaches to camera self-calibration and simultaneous estimation of motion and structure parameters of multiple independently moving objects are proposed using sequential importance sampling. Both theoretical analysis and experimental results show that the challenging issues in solving the structure from motion problem such as errors in feature tracking, feature occlusion, motion/structure ambiguity, mixed-domain sequences, mismatched features/independently moving objects, and the need for good initialization and critical motion sequences in camera self-calibration can be well modeled and effectively solved by the proposed methods.