Laser Radar 3D Target Detection Based on Improved PointPillars

Feng Tian; Chao Liu; Fang Liu; Wenwen Jiang; Xin Xu; Ling Zhao

doi:10.3788/LOP231493

Journals >Laser & Optoelectronics Progress >Volume 61 >Issue 8 >Page 0812007 > Article

Laser & Optoelectronics Progress
Vol. 61, Issue 8, 0812007 (2024)

Laser Radar 3D Target Detection Based on Improved PointPillars

Feng Tian, Chao Liu, Fang Liu^*, Wenwen Jiang, Xin Xu, and Ling Zhao

Author Affiliations

School of Computer and Information Technology, Northeast Petroleum University, Daqing 163318, Heilongjiang , China

show less

DOI: 10.3788/LOP231493 Cite this Article Set citation alerts

Feng Tian, Chao Liu, Fang Liu, Wenwen Jiang, Xin Xu, Ling Zhao. Laser Radar 3D Target Detection Based on Improved PointPillars[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0812007 Copy Citation Text

show less

Fig. 1. Network structure

Download full size

Fig. 2. Structure of the pillar feature network of the original PointPillars model

Download full size

Fig. 3. Structure of improved pillar encoding network

Download full size

Fig. 4. Structure of ConvNeXt module

Download full size

Fig. 5. Structure of backbone network based on ConvNeXt module

Download full size

Fig. 6. 3D object detection renderings and 2D images of the proposed algorithm on different scenes. (a) Scene one; (b) scene two; (c) scene three; (d) scene four

Download full size

Fig. 7. Comparison of detection performance between proposed algorithm and PointPillars algorithm: complex scenes

Download full size

Fig. 8. Comparison of detection performance between proposed algorithm and PointPillars algorithm: long-distance scenes

Download full size

Model	AP /%			mAP /%
Model	Easy	Moderate	Hard	mAP /%
VoxelNet	87.93	75.37	73.21	78.84
SECOND	88.61	78.62	77.22	81.48
PointPillars	87.50	77.01	74.77	79.76
3D-GIoU	87.83	77.91	78.84	82.60
TANet	88.17	77.75	75.31	80.41
PointRCNN	89.01	78.77	78.10	81.96
Point-GNN	89.33	79.47	78.29	82.36
Part-A²	89.56	79.41	78.84	82.60
Ours	89.28	79.56	79.62	82.82

Table 1. Comparison of mAP for different algorithms under car category

Model	AP /%			mAP /%
Model	Easy	Moderate	Hard	mAP /%
VoxelNet	67.81	63.52	58.87	63.40
SECOND	56.00	50.02	43.64	49.89
PointPillars	66.73	61.06	56.50	61.43
3D-GIoU	67.23	59.58	52.69	59.83
TANet	70.80	63.45	58.22	64.16
PointRCNN	62.69	55.36	51.60	56.55
Point-GNN	61.92	53.77	50.14	55.28
Part-A²	65.69	60.05	55.45	60.40
Ours	71.33	63.75	58.63	64.57

Table 2. Comparison of mAP for different algorithms under the pedistrian category

Model	AP /%			mAP /%
Model	Easy	Moderate	Hard	mAP /%
VoxelNet	77.69	58.72	51.63	62.68
SECOND	80.97	63.43	56.67	67.02
PointPillars	83.65	63.40	59.71	68.92
3D-GIoU	83.32	64.69	63.51	70.51
TANet	85.21	65.29	61.57	70.69
PointRCNN	84.48	65.37	59.83	69.89
Point-GNN	86.60	67.48	62.58	72.22
Part-A²	85.50	68.90	64.53	72.98
Ours	87.88	68.75	64.26	73.63

Table 3. Comparison of mAP for different algorithms under the cyclist category

Method	Average pooling	Attention pooling	ConvNeXt	AP /%			mAP /%	FPS /（frame/s）
Method	Average pooling	Attention pooling	ConvNeXt	Car	Pedestrian	Cyclist	mAP /%	FPS /（frame/s）
Baseline				79.76	61.43	68.92	70.04	42.2
Experiment 1	√			80.31	62.03	70.90	71.08	38.3
Experiment 2		√		80.54	62.43	71.47	71.48	36.4
Experiment 3	√	√		81.26	62.78	72.39	72.14	33.1
Ours	√	√	√	82.82	64.57	73.63	73.67	26.1

Table 4. Results of ablation experiment

Feng Tian, Chao Liu, Fang Liu, Wenwen Jiang, Xin Xu, Ling Zhao. Laser Radar 3D Target Detection Based on Improved PointPillars[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0812007

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information