In this paper, we propose a traditional semi-dense direct visual odometry (VO) based on our preliminary study using low-order Gaussian derivative functions for solving a VO problem with pure frame-by-frame point tracking. With the off-line fitting analysis of residual sets that we firstly performed to determine the coarse-to-fine framework, this study employs a simple local interpolation to enrich the searching space of the subsample of the original image. Without any processing for dealing with implementation acceleration, tracking lost and divergence problems, the proposed approach achieves relatively acceptable performance compared with baseline algorithms of both the direct approach and the matching-based data association algorithm. An experimental study is conducted using a group of TUM datasets and the reference VO algorithms.
In this paper, we propose a traditional semi-dense direct visual odometry (VO) based on our preliminary study using low-order Gaussian derivative functions for solving a VO problem with pure frame-by-frame point tracking. With the off-line fitting analysis of residual sets that we firstly performed to determine the coarse-to-fine framework, this study employs a simple local interpolation to enrich the searching space of the subsample of the original image. Without any processing for dealing with implementation acceleration, tracking lost and divergence problems, the proposed approach achieves relatively acceptable performance compared with baseline algorithms of both the direct approach and the matching-based data association algorithm. An experimental study is conducted using a group of TUM datasets and the reference VO algorithms.