• Acta Optica Sinica
  • Vol. 41, Issue 22, 2215002 (2021)
Hao Chen1, Kailun Yang2, Weijian Hu1, Jian Bai1, and Kaiwei Wang1,*
Author Affiliations
  • 1National Engineering Research Center of Optical Instrumentation, Zhejiang University, Hangzhou, Zhejiang 310058, China
  • 2Institute for Anthropomatics and Robotics, Karlsruhe Institute of Technology, Karlsruhe 76131, Germany
  • show less
    DOI: 10.3788/AOS202141.2215002 Cite this Article Set citation alerts
    Hao Chen, Kailun Yang, Weijian Hu, Jian Bai, Kaiwei Wang. Semantic Visual Odometry Based on Panoramic Annular Imaging[J]. Acta Optica Sinica, 2021, 41(22): 2215002 Copy Citation Text show less
    Panoramic annular imaging
    Fig. 1. Panoramic annular imaging
    Camera model. (a) PAL camera model; (b) image plane; (c) sensor plane
    Fig. 2. Camera model. (a) PAL camera model; (b) image plane; (c) sensor plane
    Flowchart of PASVO algorithm
    Fig. 3. Flowchart of PASVO algorithm
    Panoramic annular semantic segmentation (PASS)
    Fig. 4. Panoramic annular semantic segmentation (PASS)
    Correspondence search along epipolar curve under semantic guidance. (a) One of the keypoints; (b) searching matching points along polar curve under semantic guidance
    Fig. 5. Correspondence search along epipolar curve under semantic guidance. (a) One of the keypoints; (b) searching matching points along polar curve under semantic guidance
    Keypoints extraction under semantic guidance. (a) Semantic segmentation result; (b) confidence distribution; (c) original keypoints extraction; (d) keypoints extraction under semantic guidance
    Fig. 6. Keypoints extraction under semantic guidance. (a) Semantic segmentation result; (b) confidence distribution; (c) original keypoints extraction; (d) keypoints extraction under semantic guidance
    Experiment setup. (a) Remote control vehicle; (b) PAL image expands to a perspective image
    Fig. 7. Experiment setup. (a) Remote control vehicle; (b) PAL image expands to a perspective image
    Effectiveness test of keypoints extraction under semantic guidance. (a) Proportion of inliers with/without semantic guidance; (b) RANSAC convergence probability with/without semantic guidance
    Fig. 8. Effectiveness test of keypoints extraction under semantic guidance. (a) Proportion of inliers with/without semantic guidance; (b) RANSAC convergence probability with/without semantic guidance
    Error between trajectory estimated by PASVO and reference trajectory in the accuracy test. (a) S1; (b) S5; (c) S6
    Fig. 9. Error between trajectory estimated by PASVO and reference trajectory in the accuracy test. (a) S1; (b) S5; (c) S6
    Trajectories estimated by the different algorithms on the large-scale dataset. The black dot represents the starting point, and dots with different colors denote the end points of the trajectory estimated by different algorithms
    Fig. 10. Trajectories estimated by the different algorithms on the large-scale dataset. The black dot represents the starting point, and dots with different colors denote the end points of the trajectory estimated by different algorithms
    DatasetLength /mAbsolute translation error /m
    PASVOPALVO[35]CubemapSLAM[11]SVO[3]ORBSLAM2[4]
    S1630.66680.67850.68650.73501.4700
    S2600.40510.43730.49980.62760.4859
    S390.15970.19370.42820.1513
    S4160.21740.21680.24050.3293
    S5250.39060.39430.40380.3727
    S6430.54510.59860.62690.6840
    S7190.08800.11960.07300.11780.1233
    S8180.25240.26160.27320.43720.3445
    S9411.42941.58591.82712.04011.7519
    Table 1. Absolute translation error
    DatasetLength /mLoop closure error /%
    PASVOPALVO[35]CubemapSLAM[11]SVO[3]ORBSLAM2[4]
    Caolou1901.12602.57190.99893.67021.2254
    Fountain2001.04694.22884.30816.3477
    Library2501.80313.63104.07946.5322
    Table 2. Loop closure error in the large-scale dataset