LiDAR Object Detection Method Based on Multi-Stage Feature Extraction

Fig. 7. Trend diagrams of training process parameters. (a) Curve of learning rate with the number of iteration steps; (b) curves of detection error with the number of iteration steps

Download full size

Fig. 8. Dataset detection effect pictures. (a) Images; (b) Velodyne

Download full size

Fig. 9. Vehicle used in the experiment

Download full size

Fig. 10. Real vehicle algorithm detection effect diagrams. (a) Original images; (b) Velodyne

Download full size

Table 1. IoU threshold and detection box settings

View table

Table 1. IoU threshold and detection box settings

Class	Different grade IoU threshold setting			Check frame size （length， width， and height） /m
Class	Easy	Medium	Hard	Check frame size （length， width， and height） /m
Car	0.7	0.7	0.7	（1.30，3.90，1.50）
Cyclist	0.7	0.7	0.7	（0.60，1.76，1.73）
Pedestrian	0.5	0.5	0.5	（0.60，0.80，1.73）

Table 2. Detection accuracy of different algorithms on dataset from BEV perspective

View table

Table 2. Detection accuracy of different algorithms on dataset from BEV perspective

Model	Car			Pedestrian			Cyclist			mAP
Model	Easy	Medium	Hard	Easy	Medium	Hard	Easy	Medium	Hard	mAP
SECOND	88.07	84.00	75.33	58.09	50.22	47.20	83.66	66.19	62.13	68.24
VoxelNet	89.35	79.26	77.39	46.13	40.74	38.11	66.7	54.76	50.55	60.33
F-PointNet	88.70	84.00	75.33	58.09	51.05	47.54	75.38	61.69	54.68	66.27
PointPillars	92.05	87.80	85.19	56.53	50.83	46.43	81.32	65.07	60.73	69.55
Proposed	91.76	87.73	85.26	61.79	54.13	49.56	84.79	68.96	64.58	72.06

Table 3. Detection accuracy of different algorithms on dataset from 3D perspective

View table

Table 3. Detection accuracy of different algorithms on dataset from 3D perspective

Model	Car			Pedestrian			Cyclist			mAP
Model	Easy	Medium	Hard	Easy	Medium	Hard	Easy	Medium	Hard	mAP
SECOND	86.44	76.97	73.39	47.47	40.47	36.26	81.28	63.49	59.29	62.78
VoxelNet	77.47	65.11	57.73	39.48	33.69	31.50	61.22	48.36	44.37	50.99
F-PointNet	81.20	70.39	62.19	51.21	44.89	40.23	71.96	56.77	50.39	58.80
PointPillars	85.03	75.76	72.74	50.08	44.18	39.53	77.13	60.94	56.91	62.48
Proposed	86.99	76.10	73.23	55.70	47.84	42.25	81.01	63.33	59.74	65.13

Table 4. Detection accuracy of different algorithms on dataset from AOS perspective

View table

Table 4. Detection accuracy of different algorithms on dataset from AOS perspective

Model	Car			Pedestrian			Cyclist			mAP
Model	Easy	Medium	Hard	Easy	Medium	Hard	Easy	Medium	Hard	mAP
SECOND	94.84	90.94	90.11	60.01	53.92	50.77	89.40	72.82	68.87	74.63
SubCNN	90.61	88.43	78.63	78.33	66.28	61.37	71.39	63.41	46.34	71.64
AVOD-FPN	89.95	87.13	79.74	53.36	44.92	43.77	67.61	57.53	54.16	64.24
PointPillars	95.02	91.24	88.46	47.33	44.40	41.31	84.75	71.35	67.24	70.12
Proposed	95.28	91.45	88.77	57.87	51.40	47.49	86.47	74.30	70.28	73.70

Table 5. Detection accuracy of the ablation experiment from BEV perspective on KITTI dataset

View table

Table 5. Detection accuracy of the ablation experiment from BEV perspective on KITTI dataset

Model	Car			Pedestrian			Cyclist			mAP
Model	Easy	Medium	Hard	Easy	Medium	Hard	Easy	Medium	Hard	mAP
PP	91.05	87.80	85.19	56.57	50.83	46.43	81.32	65.07	60.73	69.55
PP+ZC	91.66	87.65	84.94	58.06	51.58	46.98	85.96	68.29	63.82	70.99
PP+KD	91.74	87.40	84.94	57.66	51.05	46.52	85.64	66.11	61.66	70.31
PP+GA	92.03	87.99	85.21	58.27	51.61	47.43	81.88	65.00	60.73	70.02
PP+ZC+KD	91.57	87.58	85.07	57.95	51.31	47.11	84.75	66.62	62.01	70.56
PP+GA+KD	91.65	87.71	85.11	58.84	52.33	47.85	84.83	66.54	62.44	70.81
PP+ZC+KD+GA	91.76	87.73	85.26	61.79	54.13	49.56	84.79	68.96	64.58	72.06

Table 6. Detection accuracy of ablation experiment from 3D perspective on KITTI dataset

View table

Table 6. Detection accuracy of ablation experiment from 3D perspective on KITTI dataset

Model	Car			Pedestrian			Cyclist			mAP
Model	Easy	Medium	Hard	Easy	Medium	Hard	Easy	Medium	Hard	mAP
PP	85.03	75.76	72.74	50.08	44.18	39.53	77.13	60.94	56.91	62.48
PP+ZC	86.93	75.54	72.54	51.23	44.87	40.03	80.59	63.53	59.08	63.82
PP+KD	86.51	75.67	72.59	51.65	44.88	40.13	82.02	62.94	58.45	63.88
PP+GA	87.03	75.89	72.89	50.96	44.19	39.58	80.73	62.94	58.69	63.02
PP+ZC+KD	86.64	75.89	72.81	51.81	44.78	40.49	77.73	61.01	56.98	63.65
PP+GA+KD	84.94	75.69	72.66	51.93	44.93	40.34	82.32	62.31	58.24	63.71
PP+ZC+KD+GA	86.99	76.10	73.23	55.70	47.84	42.25	81.01	63.33	59.74	65.13

Table 7. Detection accuracy of ablation experiment from AOS perspective on KITTI dataset

View table

Table 7. Detection accuracy of ablation experiment from AOS perspective on KITTI dataset

Model	Car			Pedestrian			Cyclist			mAP
Model	Easy	Medium	Hard	Easy	Medium	Hard	Easy	Medium	Hard	mAP
PP	95.02	91.24	88.46	50.08	44.18	41.31	84.75	71.35	67.24	70.12
PP+ZC	95.28	91.34	88.51	51.08	47.10	44.07	85.37	67.86	63.72	70.48
PP+KD	95.26	91.14	88.36	50.49	45.72	42.63	89.05	72.81	68.01	71.50
PP+GA	95.30	91.32	88.58	49.42	45.63	42.74	88.09	71.18	66.99	71.03
PP+ZC+KD	95.40	91.57	88.60	50.92	46.71	43.59	84.65	73.11	68.90	71.50
PP+GA+KD	95.11	91.14	88.41	51.46	46.52	43.16	87.60	75.20	70.81	72.16
PP+ZC+KD+GA	95.28	91.45	88.77	57.87	51.40	47.49	86.47	74.30	70.28	73.70

Tools

Get Citation

Copy Citation Text

Zhipeng Zhai, Jinju Shao, Song Gao, Zhibing Duan, Lei Wang. LiDAR Object Detection Method Based on Multi-Stage Feature Extraction[J]. Laser & Optoelectronics Progress, 2025, 62(10): 1015004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Sep. 24, 2024

Accepted: Nov. 1, 2024

Published Online: May. 8, 2025

The Author Email: Jinju Shao (sjjgbh@163.com)

DOI:10.3788/LOP242028

CSTR:32186.14.LOP242028

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology