Acta Photonica Sinica, Volume. 52, Issue 9, 0912002(2023)
3D Object Detection with Fusion Point Attention Mechanism in LiDAR Point Cloud
Figures & Tables(13)
Fig. 2. The overall framework of the improved PointPillars algorithm
Fig. 6. Comparison of the visualization results of PointPillars and the algorithm in this paper
Table 1. CSPNet network structure of this paper
View tableView in ArticleTable 1. CSPNet network structure of this paper
Layer Repeat Kernel size Stride Output channels Conv2d 1 3×3 2 64 Stage1 Conv1 3 1×1 1 32 Conv2 1×1 1 32 BottleNeck 1×1
3×3
32
32
Conv3 1×1 1 64 Conv2d 1 3×3 2 128 Stage2 Conv1 5 1×1 1 64 Conv2 1×1 1 64 BottleNeck 1×1
3×3
64
64
Conv3 1×1 1 128 Conv2d 1 3×3 2 256 Stage3 Conv1 5 1×1 1 128 Conv2 1×1 1 128 BottleNeck 1×1
3×3
128
128
Conv3 1×1 1 256
Table 2. Data division in three scenarios
View tableView in ArticleTable 2. Data division in three scenarios
Easy Moderate Hard Min height of bounding box 40 pixels 25 pixels 25 pixels Max blocking level Fully visible Partially obscured Hard to see Maximum cut-off 15% 30% 50%
Table 3. Experimental environment configuration
View tableView in ArticleTable 3. Experimental environment configuration
Experimental environment Configuration Operating system Ubantu 16.04 Processor Intel Xeon Silver 411 Memory 64 GB Video card NVIDIA TITAN V Deep learning framework Pytorch 1.5 Development language Python 3.7
Table 4. Comparison of AP for different methods（%）
View tableView in ArticleTable 4. Comparison of AP for different methods（%）
Method/ Car-3D（IoU=0.7） Car-BEV（IoU=0.7） Easy Moderate Hard Easy Moderate Hard F-PointNets［ 25 ］82.19 69.79 60..59 91.17 84.67 74.77 VoxelNet［ 10 ］87.93 75.37 73.21 89.35 79.26 77.39 SECOND［ 12 ］83.34 72.55 65.82 89.39 83.77 78.59 PointPillars［ 13 ］86.29 76.77 73.92 91.89 88.07 87.02 TANet［ 26 ］84.39 75.94 68.82 75.70 59.44 52.53 SegVoxelNet［ 27 ］86.04 76.12 70.76 91.62 86.37 83.04 PointRCNN［ 8 ］86.96 75.64 70.70 92.13 87.39 82.72 Part- ［ 28 ］87.81 78.49 73.51 91.70 87.79 84.61 Ours 88.52 79.02 76.22 92.63 88.53 87.16
Table 5. Inference speed comparison among different methods
View tableView in ArticleTable 5. Inference speed comparison among different methods
Method Reasoning speed/（frame· ） F-PointNets［ 25 ］0.169 VoxelNet［ 10 ］0.033 SECOND［ 12 ］0.380 3DSSD［ 9 ］0.04 TANet［ 26 ］0.035 SegVoxelNet［ 27 ］0.04 PointRCNN［ 8 ］0.067 Part- ［ 28 ］0.08 SA-SSD［ 30 ］0.04 Ours 0.037 2
Table 6. Average precision of 3D detection for ablation experiments in the KITTI test set（%）
View tableView in ArticleTable 6. Average precision of 3D detection for ablation experiments in the KITTI test set（%）
Method Car-3D（IoU=0.7） Easy Moderate Hard PointPillars 86.29 76.77 73.92 PPPA 87.72 78.23 75.13 PPCSP 87.83 78.30 75.60 PPCSP+PPPA 88.52 79.02 76.22
Table 7. Average precision of detection in the BEV scenario of the KITTI test focused ablation experiment（%）
View tableView in ArticleTable 7. Average precision of detection in the BEV scenario of the KITTI test focused ablation experiment（%）
Method Car-BEV（IoU=0.7） Easy Moderate Hard PointPillars 91.89 88.07 87.02 PPPA 92.56 88.60 87.24 PPCSP 92.13 88.02 86.68 PPCSP+PPPA 92.63 88.53 87.16
Tools
Get Citation
Copy Citation Text
Weili LIU, Deli ZHU, Huahao LUO, Yi LI. 3D Object Detection with Fusion Point Attention Mechanism in LiDAR Point Cloud[J]. Acta Photonica Sinica, 2023, 52(9): 0912002
Paper Information
Category: Instrumentation, Measurement and Metrology
Received: Apr. 4, 2023
Accepted: May. 17, 2023
Published Online: Oct. 24, 2023
The Author Email: ZHU Deli (463453339@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20