Fig. 1. Flow chart of PointPillars algorithm
Fig. 2. Visual effect of original PointPillars detection
Fig. 3. Improved feature sampling module
Fig. 4. Lidar point cloud removal of ground parts
Fig. 5. Point cloud data enhancement. (a) Raw point cloud; (b) mirrored point cloud; (c) tilted point cloud; (d) zoomed point cloud
Fig. 6. Comparison of experimental results for hyperparameter selection
Fig. 7. Comparison of detection accuracy before and after algorithm improvement
Fig. 8. Comparison of detection effects before and after algorithm improvement
Model | Depth | Head |
---|
M1 | (1,3,1) | (2,4,8) | M2 | (2,6,2) | (2,4,8) | M3 | (2,2,6) | (2,4,8) | M4 | (2,6,2) | (2,4,2) | M5 | (4,8,4) | (4,8,4) |
|
Table 1. Hyperparameter configuration of Swin-T module
Model | Easy | Moderate | Hard | Average |
---|
M1 | 50.17 | 40.42 | 38.7 | 43.09 | M2 | 94.12 | 89.55 | 88.48 | 90.71 | M3 | 94.23 | 89.77 | 88.8 | 90.93 | M4 | 94.13 | 89.23 | 88.15 | 90.5 | M5 | 94.57 | 89.65 | 88.75 | 90.99 |
|
Table 2. Comparison of accuracy rates of different Swin-T hyperparameter configurations
Model | Easy | Moderate | Hard | Average |
---|
PointPillars[6] | 90.77 | 89.61 | 88.47 | 89.61 | SECOND[5] | 90.76 | 89.77 | 88.82 | 89.78 | SECOND-IoU[5] | 89.72 | 88.73 | 88.33 | 88.92 | PointRCNN[21] | 90.76 | 89.58 | 89.03 | 89.79 | PointRCNN-IoU[21] | 90.70 | 89.32 | 88.87 | 89.63 | Part-A2-Free[22] | 90.68 | 89.00 | 88.64 | 89.44 | AS-PointPillars[12] | 90.48 | 88.32 | 86.51 | 88.44 | AP-PointPillars[12] | 90.68 | 88.92 | 86.90 | 88.83 | Proposed model | 94.23 | 89.77 | 88.8 | 90.93 |
|
Table 3. Comparison of test results
Model | GPU memory /MB | Running speed /s |
---|
PointPillars | 1267 | 0.036 | Proposed model | 1359 | 0.058 |
|
Table 4. Comparison of GPU memory usage and running speed before and after algorithm improvement