3D Object Detection with LiDAR Based on Multi-Attention Mechanism

Jie Cao; Yiqiang Peng; Likang Fan; Lingfan Mo; Longfei Wang

doi:10.3788/LOP241407

Laser & Optoelectronics Progress, Volume. 62, Issue 4, 0412001(2025)

3D Object Detection with LiDAR Based on Multi-Attention Mechanism

Jie Cao^1、*, Yiqiang Peng^1,2,3, Likang Fan^1,2,3, Lingfan Mo⁴, and Longfei Wang¹

Author Affiliations

¹School of Automobile and Transportation, Xihua University, Chengdu 610039, Sichuan , China

²Vehicle Measurement Control and Safety Key Laboratory of Sichuan Province, Xihua University, Chengdu 610039, Sichuan , China

³Provincial Engineering Research Center for New Energy Vehicle Intelligent Control and Simulation Test Technology of Sichuan, Chengdu 610039, Sichuan , China

⁴Guangdong Xinbao Electrical Appliances Holdings Co., Ltd., Foshan 528000, Guangdong , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Overall network structure of PV-RCNN algorithm

Download full size

Fig. 2. Overall network structure of MA-RCNN algorithm

Download full size

Fig. 3. Structure of channel attention mechanism

Download full size

Fig. 4. Structure of spatial attention mechanism

Download full size

Fig. 5. Structure of point cloud self-attention encoder network

Download full size

Fig. 6. Detection results in complex scenes of the KITTI validation set by different algorithms. (a1)‒(d1) MA-RCNN; (a2)‒(d2) PV-RCNN; (a3)‒(d3) RGB images

Download full size

Fig. 7. Proposed algorithm on a real vehicle platform. (a) Hardware platform of the real vehicle; (b) autonomous driving system; (c) implementation flow of the proposed algorithm

Download full size

Fig. 8. Visualization of online detection results. (a1)‒(c1) Real images; (a2)‒(c2) detection results of point clouds

Download full size

Table 1. AP40 of different classes in the KITTI validation set detected by different algorithms

View table

Table 1. AP40 of different classes in the KITTI validation set detected by different algorithms

Algorithm	Car			Pedestrian			Cyclist
Algorithm	Easy	Mod	Hard	Easy	Mod	Hard	Easy	Mod	Hard
VoxelNet	87.93	75.37	73.21	67.81	63.52	58.87	77.69	58.72	51.63
SECOND	88.61	78.62	77.22	56.00	50.02	43.64	80.97	63.43	56.67
PointPillars	87.50	78.62	74.77	66.73	61.06	56.50	83.65	63.40	59.71
PointRCNN	89.01	77.01	74.77	62.69	55.36	51.60	84.48	65.37	59.83
Point-GNN	89.33	79.47	78.29	61.92	53.77	50.14	86.60	67.48	62.58
Part-A²	89.56	79.41	78.84	65.69	60.05	55.45	85.50	68.90	64.53
PV-RCNN	91.54	82.67	80.24	60.39	53.14	48.49	88.05	70.99	66.54
MA-RCNN	91.67	82.62	80.21	65.09	56.15	50.40	89.46	72.62	68.42

Table 2. Overall performances of various classes in KITTI validation set detected by different algorithms

View table

Table 2. Overall performances of various classes in KITTI validation set detected by different algorithms

Algorithm	mAP			Avg
Algorithm	Car	Pedestrian	Cyclist	Avg
VoxelNet	78.84	63.40	62.68	68.31
SECOND	81.48	49.89	67.02	66.13
PointPillars	80.30	61.43	68.92	70.22
PointRCNN	80.26	56.55	69.89	68.90
Point-GNN	82.36	55.28	72.22	69.95
Part-A²	82.60	60.40	72.98	71.99
PV-RCNN	84.82	54.01	75.19	71.34
MA-RCNN	84.83	57.21	76.83	72.96

Table 3. Runtime of different algorithms
View table
Table 3. Runtime of different algorithms
Algorithm Runtime /s
PV-RCNN 0.06
MA-RCNN 0.10

Table 4. Design of ablation experiments
View table
Table 4. Design of ablation experiments
Method PV-RCNN Channel attention Spatial attention Point cloud self-attention
a √
b √ √
c √ √ √
d √ √
MA-RCNN √ √ √ √

Table 5. Results of ablation experiments

View table

Table 5. Results of ablation experiments

Method	AP40						mAP
Method	Pedestrians-Easy	Pedestrians-Mod	Pedestrians-Hard	Cyclists-Easy	Cyclists-Mod	Cyclists-Hard	Pedestrians	Cyclists
a	60.39	53.14	48.49	88.05	70.99	66.54	54.01	75.19
b	60.92	53.32	48.97	88.26	71.24	66.83	54.40	75.44
c	61.55	54.02	49.24	88.67	71.61	67.12	54.94	75.80
d	63.25	54.66	49.45	88.91	72.04	67.50	55.79	76.15
MA-RCNN	65.09	56.15	50.40	89.46	72.62	68.42	57.21	76.83

Table 6. Hardware list
View table
Table 6. Hardware list
Equipment Brand Parameter
INS CHCNAV 410
IPC Advantech 610 L
GPS CHCNAV 410
LiDAR Leishen C32
MMW Continental ARS408
Camera Molose UC50
CDC Freescale XEP100

Tools

Get Citation

Copy Citation Text

Jie Cao, Yiqiang Peng, Likang Fan, Lingfan Mo, Longfei Wang. 3D Object Detection with LiDAR Based on Multi-Attention Mechanism[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0412001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites