Reducing accumulated errors in ego-motion estimation using local bundle adjustment

Incremental motion estimation methods involve a problem that estimation accuracy gradually becomes worse as the motion trajectory becomes longer and longer. This is due to accumulation of estimation errors incurred in each estimation step. To keep estimation accuracy stable even for a long trajectory, we propose to locally apply the bundle adjustment to each estimated motion so that the modified estimation becomes geometrically consistent with time-series frames acquired so far. To demonstrate the effectiveness of this approach, we employ an ego-motion estimation method using the binocular fixation control, and show that (i) our modification of estimation is statistically significant; (ii) in order to reduce estimation errors most effectively, three frames are optimal for applying the bundle adjustment; (iii) the proposed method is effective in the real situation, demonstrating drastic improvement of accuracy in estimation for a long motion trajectory.