Fig. 1. Magnetic adsorption robot working on the top inner surface of the pipeline

Fig. 2. Fowchart of the proposed method

Fig. 3. Schematic before and after coordinate deflection

Fig. 4. Detection of weld contour graphics. (a) Frames with weld joints; (b) Canny edge detection; (c) set of all points which satisfies the Hough transform threshold; (d) weld contour ellipse detected by MSHT; (e) frame existing weld in the composite video; (f) weld contour ellipse detected by MSHT

Fig. 5. ORB features are involved in magazine matching from different perspectives^[14]. (a) ORB features matching from different perspectives; (b) (c) circular regios in Fig.5 (a)

Fig. 6. Images of inner surface of pipeline. (a) Frame without welds;（b）frame with circular welds

Fig. 7. Results of ORB-SLAM3 running on EuRoC MAV dataset. (a) ORB feature extraction rendering diagram running on EuRoC MAV dataset scenario; (b) field scene point cloud map and camera motion track

Fig. 8. Results of feature extraction and map construction in the pipeline by ORB-SLAM3. (a) Extraction of ORB feature points in scenes containing welds; (b) point cloud map and camera motion track constructed during running to scene Fig.8 (a)

Fig. 9. Results of LSD-SLAM running on the LSD_ROOM.bag dataset. (a) scene contained in LSD_ROOM.bag dataset; (b) point cloud map based on Fig.9 (a)

Fig. 10. Results of LSD-SLAM running in pipeline inspection environment. (a) Result of shooting which robot moves forward in the scene without welds; (b) point cloud map which the moving robot constructed in the scene Fig.10 (a); (c) result of shooting which robot moves forward in the scene with welds; (d) point cloud map which the moving robot constructed in the scene Fig.10 (c)

Table 1. Detection results of distance and swing angle when there is left and right swing