Fig. 1. Lidar point clouds in different scenes. (a) Ground point clouds; (b) calm water surface point clouds
Fig. 2. Algorithm block diagram
Fig. 3. Lidar water surface echo data. (a) (b) Calm water surface; (c) (d) wave water surface
Fig. 4. DBSCAN filtering algorithm based on water surface point clouds
Fig. 5. DBSCAN clustering. (a) Wave point clouds over water; (b) clustering result
Fig. 6. Water surface target detection algorithm
Fig. 7. Structure of feature learning network
Fig. 8. VFE layer structure
Fig. 9. Region proposal network architecture
Fig. 10. Experimental platform
Fig. 11. Measured data. (a) Spheres; (b) tri-pyramid; (c) cylindrical; (d) multi-objective
Fig. 12. DBSCAN-VoxelNet loss value in the training process
Fig. 13. Construction of surface target detection simulation environment
Fig. 14. Virtual wave fluctuation state in the first view of ship
Fig. 15. Target setting of water virtual environment
Performance | Parameter |
---|
Number of channels | TOF ranging 16 channels | Ranging | 20 cm to 150 m(target reflectivity is 20%) | Accuracy | Within ±2 cm(typical value) | Vertical view | ±15°(30° in total) | Vertical angular resolution | 2° | Horizontal perspective | 360° | Azimuth resolution | 0.09°(5 Hz)to 0.36°(20 Hz) | Rotation speed | 300/600/1200 rad·min-1(5/10/20 Hz) |
|
Table 1. RS-LiDAR-16 sensor parameters
APspheres | APtri-pyramid | APcylindrical | APmulti-objective |
---|
5 m | 10 m | 15 m | 5 m | 10 m | 15 m | 5 m | 10 m | 15 m | 5 m | 10 m | 15 m | 0.876 | 0.865 | 0.793 | 0.881 | 0.873 | 0.866 | 0.858 | 0.852 | 0.844 | 0.812 | 0.798 | 0.776 |
|
Table 2. Detection results of DBSCAN-VoxelNet on water surface target dataset
Parameter | VoxelNet | DBSCAN-VoxelNet |
---|
mAP | 0.812 | 0.824 | Average detection speed /(frame·s-1) | 0.07 | 0.08 |
|
Table 3. mAP detection results for water surface target
Environment | mAP | Average detection speed /(frame·s-1) |
---|
Without wave | 0.841 | 0.08 | With wave | 0.897 | 0.08 |
|
Table 4. mAP of DBSCAN-VoxelNet at different environments