Author Affiliations
1Beijing Key Laboratory of Information Service Engineering, Beijing Union University, Beijing 100101, China2Beijing Engineering Research Center of Smart Mechanical Innovation Design Service, Beijing 100101, Chinashow less
Fig. 1. Block diagram of semantic map construction
Fig. 2. Framework of PFPN semantic segmentation network
Fig. 3. Structure of reduced PFPN
Fig. 4. Schematic diagram of coordinate system
Fig. 5. Time matching method of Lidar and camera
Fig. 6. Fusion result
Fig. 7. Construction of semantic map
Fig. 8. Experimental test route and test environment. (a) Test route; (b) test environment
Fig. 9. Map comparison. (a) SLAM map established by original point cloud; (b) semantic map after removing vehicle information in environment; (c) (d) enlarged display of rectangular area in Fig. 9 (a),(b)
Fig. 10. Mapping track
Fig. 11. Error comparison between x-axis and y-axis
Fig. 12. Comparison of overall pose error
Method | Backbone | mIoU /% | Time /s |
---|
DANet | ResNet-101 | 39.7 | | HRNet | HRNet | 42.7 | 3.89 | PSPNet | ResNet-50 | 39.9 | 0.36 | PFPN | ResNet-101-3× | 42.1 | 0.28 | PFPN | ResNet-50-3× | 40.2 | 0.21 | PFPN | ResNet-50-1× | 41.2 | 0.18 | Proposed method | ResNet-50-1× | 41.0 | 0.13 |
|
Table 1. Comparison of different methods
Method | Nnumber of point clouds |
---|
Original method | 2089 | Proposed method | 1258 | Change | 831 |
|
Table 2. Average number of point clouds per frame
Method | Time /ms |
---|
Original method | 53 | Proposed method | 31 | Change | 22 |
|
Table 3. Comparison of average single registration time
Parameter | Original method | Proposed method | Change |
---|
x /m | 0.4388 | 0.3298 | 0.1090 | y /m | 0.3508 | 0.2371 | 0.1137 | P /rad | 0.2458 | 0.1614 | 0.0844 |
|
Table 4. Average error analysis of key parameters