Application of nonlinear optimization method in stereo visual odometry

Nonlinear optimization approaches are becoming more prevalent in the field of navigation with visual sensors. Therefore, this paper presented a refinement motion estimation approach which showed the advantages compared to the typical method. The proposed framework included established observation model and the mathematical derivation of nonlinear optimization algorithm. Speaking of model part, the relationship between frame pixels and the coarse pose estimated by RANSAC was quantized by a formula which became a prerequisite for optimization. While in the nonlinear optimization part, one strategy was taken into consideration to reduce the cumulative error. That is to say, the current pose was optimized each time a new frame comes. Finally, tests on the Malaga urban datasets verified that this approach achieved higher accuracy than the RANSAC estimation. As a result, benefits of using nonlinear optimization in motion estimation refinement were verified, which could also provide important priori information for simultaneous localization and mapping (SLAM).