Fig. 1. Flowchart of indoor mobile robot localization based on deep learning doorplate detection

Fig. 2. MobileNet-SSD model for doorplate area detection

Fig. 3. Doorplate pre-processing procedure. (a) Detection of doorplate area; (b) initial cutting of the doorplate; (c) threshold segmentation; (d) contour detection; (e) background filling

Fig. 4. Improved rotation and projection methods.(a) Doorplate rotation diagram; (b) doorplate traversal diagram

Fig. 5. Doorplate area after tilt correction

Fig. 6. Flowchart of the improved rotation projection method

Fig. 7. Doorplate template and SURF feature point matching graph. (a) Detected front view of doorplate template in doorplate area; (b) SURF feature point matching graph

Fig. 8. Relationship between the camera and world frame

Fig. 9. Mobile robot platform and experimental scene

Fig. 10. Doorplate area detection result based on MobileNet-SSD

Fig. 11. Time cost of the improved and traditional rotation projection methods

Fig. 12. Comparison of localization accuracy. (a) Position error; (b) orientation error

Fig. 13. Experiment trajectory of the robot. (a) Planned the route; (b) travel trajectory based on the AMCL method; (c) travel trajectory after adding doorplate positioning