LiDAR 3D Object Detection Based on Improved PointRCNN

Fig. 5. Visualization of point cloud before and after preprocessing.(a) Point cloud image before preprocessing;(b) point cloud image after preprocessing

Download full size

Fig. 6. Visual comparison of detection effect between proposed algorithm and PointRCNN algorithm

Download full size

Table 1. Comparison of detection accuracy results of different algorithms on Car category

View table

Table 1. Comparison of detection accuracy results of different algorithms on Car category

Model	3D AP			3D mAP
Model	Easy	Moderate	Hard	3D mAP
VoxelNet	81.25	66.73	61.52	69.83
SECOND	87.23	77.85	76.15	80.41
F-PointNet	84.22	71.65	62.25	72.71
PointPillars	86.75	76.74	73.77	79.09
PointGNN	87.76	77.26	74.19	79.74
PointRCNN	88.43	79.55	77.24	81.74
3DSSD	87.39	80.73	73.26	80.46
Part-A²	88.41	78.53	77.08	81.34
Proposed	88.77	81.55	79.34	83.22

Table 2. Comparison of detection accuracy results of different algorithms on Pedestrian category

View table

Table 2. Comparison of detection accuracy results of different algorithms on Pedestrian category

Model	3D AP			3D mAP
Model	Easy	Moderate	Hard	3D mAP
VoxelNet	57.16	53.83	47.93	52.97
SECOND	51.27	45.22	40.39	45.63
F-PointNet	58.09	50.22	47.20	51.84
PointPillars	58.33	50.13	44.41	50.96
PointGNN	52.04	42.55	41.17	45.25
PointRCNN	54.18	47.55	41.74	47.82
3DSSD	54.75	43.23	40.31	46.10
Part-A²	55.12	44.50	42.83	47.48
Proposed	59.41	52.72	46.95	53.03

Table 3. Comparison of detection accuracy results of different algorithms on Cyclist category

View table

Table 3. Comparison of detection accuracy results of different algorithms on Cyclist category

Model	3D AP			3D mAP
Model	Easy	Moderate	Hard	3D mAP
VoxelNet	67.73	46.74	44.82	53.10
SECOND	82.77	64.23	57.01	68.00
F-PointNet	75.21	62.05	54.63	63.96
PointPillars	77.13	58.76	52.03	62.64
PointGNN	78.54	62.50	57.83	66.29
PointRCNN	85.06	64.59	60.51	70.05
3DSSD	82.36	64.11	56.87	67.78
Part-A²	82.03	68.22	62.03	70.76
Proposed	86.04	66.69	62.65	71.79

Table 4. mAP evaluation results calculated using 40 recall locations in all categories

View table

Table 4. mAP evaluation results calculated using 40 recall locations in all categories

Method	Spatial auto correlation	Manifold self-attention	Group-wise self-attention	Car 3D	Pedestrian 3D	Cyclist 3D	3D mAP
Baseline				81.74	47.82	70.05	66.54
Experiment 1	√			81.91	48.56	70.60	67.02
Experiment 2		√		82.13	50.37	71.26	67.92
Experiment 3		√	√	82.64	51.92	71.15	68.57
Proposed	√	√	√	83.22	53.03	71.79	69.35

Table 5. Comparison of detection accuracy results of different self-attention mechanisms
View table
Table 5. Comparison of detection accuracy results of different self-attention mechanisms
Method Car 3D Pedestrian 3D Cyclist 3D 3D mAP
Self-attention 81.66 49.95 71.10 67.57
Manifold self-attention 82.13 50.37 71.26 67.92

Table 6. Comparison of the influence of different modules on the performance of network models
View table
Table 6. Comparison of the influence of different modules on the performance of network models
Method Parameter count /10⁶ Time /ms
Baseline 3.0 1146
Baseline+MSA 10.8 1213
Baseline+MSA+GSA 8.7 1149

Table 7. Comparison of detection accuracy results of network model with different point cloud preprocessings
View table
Table 7. Comparison of detection accuracy results of network model with different point cloud preprocessings
Method 3D mAP /%
Random sampling 60.73
Uniform sampling 61.55
Farthest point sampling 66.26
Proposed 67.02

Table 8. Comparison of detection accuracy results of algorithm with different group numbers
View table
Table 8. Comparison of detection accuracy results of algorithm with different group numbers
Method Time /ms 3D mAP /%
［1，1，1，1，1］ 1238 67.70
［8，8，8，8，8］ 1195 68.11
［1，c/1，c/2，c/4，c/8］ 1149 69.35

Tools

Get Citation

Copy Citation Text

Han Gao, Ying Chen, Lizheng Ni, Xiuhan Deng, Kai Zhong, Chengzhi Yan. LiDAR 3D Object Detection Based on Improved PointRCNN[J]. Laser & Optoelectronics Progress, 2024, 61(22): 2215004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Dec. 14, 2023

Accepted: Mar. 25, 2024

Published Online: Nov. 20, 2024

The Author Email: Ying Chen (cheny8262@163.com)

DOI:10.3788/LOP232672

CSTR:32186.14.LOP232672

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Comparison of detection accuracy results of different algorithms on Car category

Table 1. Comparison of detection accuracy results of different algorithms on Car category

Table 2. Comparison of detection accuracy results of different algorithms on Pedestrian category

Table 2. Comparison of detection accuracy results of different algorithms on Pedestrian category

Table 3. Comparison of detection accuracy results of different algorithms on Cyclist category

Table 3. Comparison of detection accuracy results of different algorithms on Cyclist category

Table 4. mAP evaluation results calculated using 40 recall locations in all categories

Table 4. mAP evaluation results calculated using 40 recall locations in all categories

Table 5. Comparison of detection accuracy results of different self-attention mechanisms

Table 5. Comparison of detection accuracy results of different self-attention mechanisms

Table 6. Comparison of the influence of different modules on the performance of network models

Table 6. Comparison of the influence of different modules on the performance of network models

Table 7. Comparison of detection accuracy results of network model with different point cloud preprocessings

Table 7. Comparison of detection accuracy results of network model with different point cloud preprocessings

Table 8. Comparison of detection accuracy results of algorithm with different group numbers

Table 8. Comparison of detection accuracy results of algorithm with different group numbers