Fig. 1. Framework of system

Fig. 2. Results of each line extraction algorithm in scene of low-texture. (a) Hough transform algorithm; (b) EDLines algorithm; (c) LSD algorithm; (d) CannyLines algorithm

Fig. 3. Local neighborhood of line midpoint

Fig. 4. Dataset of ICL-NUIM

Fig. 5. Results of proposed algorithm on ICL-NUIM dataset. (a) Input images; (b) point and line features extracted from the images; (c) estimated camera trajectories

Fig. 6. Some challenging scenarios of visual odometry. (a) White walls; (b) ceilings; (c) corridors; (d) scenes of illumination change

Fig. 7. Estimated camera trajectories of proposed algorithm on TUM dataset. (a) fr1/desk; (b) fr2/desk; (c) fr3/long_office; (d) fr3/cabinet; (e) fr3/str_ntex_near; (f) fr3/str_ntex_far

Fig. 8. Intuitive diagram of relative pose error of proposed algorithm on TUM dataset. (a) fr1/desk; (b) fr3/long_office; (c) fr3/str_ntex_far

Fig. 9. Estimated camera trajectories of proposed algorithm on CoRBS dataset. (a) D1 sequence; (b) E1 sequence; (c) E4 sequence; (d) H1 sequence

Fig. 10. Location accuracy and convergence time of proposed algorithm using I and ξc-1,kn initial ξck0. (a) Comparison of ATE RMSE; (b) comparison of convergence time

Fig. 11. Location accuracy ATE RMSE and convergence time of our algorithm using ξc-1,kn and ξc-1,kn􀱋ξ_c-2,c-1 initial ξck0. (a) Comparison of ATE RMSE; (b) comparison of convergence time