Author Affiliations
1School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, Hubei , China2Institute of Robotics and Intelligent Systems, Wuhan University of Science and Technology, Wuhan 430081, Hubei , Chinashow less
Fig. 1. Proposed framework
Fig. 2. ASPP module in DeepLab V3+
Fig. 3. Standard convolution decomposition
Fig. 4. Schematic of corner extraction
Fig. 5. Violence matching result
Fig. 6. Feature matching results after screening
Fig. 7. Algorithm training overall loss curve
Fig. 8. Comparison of segmentation results between DeepLab V3+ and improved DeepLab V3+. (a) Input images; (b) ground truth; (c) segmentation results of DeepLab V3+; (d) segmentation results of improved DeepLab V3+
Fig. 9. 3D point cloud maps. (a) Perspective 1; (b) perspective 2
Fig. 10. 3D point cloud after segmentation. (a) Perspective 1; (b) perspective 2
Fig. 11. Semantic maps. (a) Perspective 1; (b) perspective 2
Fig. 12. Semantic segmentation results of improved DeepLab V3+. (a) Scene Ⅰ; (b) scene Ⅱ; (c) semantic segmentation result under scene I; (d) semantic segmentation result under scene Ⅱ
Fig. 13. Semantic maps from different perspectives. (a) 3D map; (b) semantic map from perspective 1; (c) semantic map from perspective 2; (d) semantic map from perspective 3
Training parameter | Value |
---|
Batch size | 4 | Learning rate | 0.0001 | Power | 0.9 | Epoch | 100 |
|
Table 1. Network parameter selection
Algorithm | mIOU /% | PA /% | Number of parameters | Time /ms |
---|
DeepLab V3+ | 78.27 | 89.43 | 9.03×107 | 318 | Improved DeepLab V3+ | 76.95 | 87.18 | 3.53×106 | 75 |
|
Table 2. Algorithm evaluation index comparison
Dataset | Number of frames | RMSE of relative trajectory |
---|
RGBD‑SLAM | ORB‑SLAM | Proposed algorithm |
---|
fr1/floor | 1242 | 0.0044 | 0.0041 | 0.0040 | fr1/xyz | 798 | 0.0058 | 0.0059 | 0.0057 | fr2/360 | 1431 | 0.035 | 0.033 | 0.0276 | fr2/desk | 2965 | 0.0037 | 0.0036 | 0.0035 |
|
Table 3. RMSE of relative trajectory
Dataset | ICP algorithm | Improved ICP algorithm |
---|
Average time /s | Average number of iterations | Average time /s | Average number of iterations |
---|
fr1/floor | 0.042 | 9 | 0.016 | 3 | fr2/desk | 0.033 | 7 | 0.012 | 2 |
|
Table 4. Speed of motion estimation and number of iterations
Method | Number of point clouds | Total map construction time /ms | Map size /Mbit |
---|
ORB +YOLOv3 | 4132273 | 508 | 145.7 | ORB+MASK-RCNN | 3740152 | 461 | 128.2 | ORB+DeepLab V3+ | 3309853 | 439 | 110.5 | Proposed method | 1176592 | 231 | 71.2 |
|
Table 5. Comparison of map construction performance of different methods
Scene | mIOU /% | Processing time /ms |
---|
Ⅰ | 76.59 | 86 | Ⅱ | 77.92 | 53 |
|
Table 6. Segmentation performance under different scenes