Improved Two-Stage 3D Object Detection Algorithm for Roadside Scenes with Enhanced PointPillars and Transformer

Liangzi Wang; Miaohua Huang; Ruoying Liu; Chengcheng Bi; Yongkang Hu

doi:10.3788/LOP240516

Laser & Optoelectronics Progress, Volume. 61, Issue 18, 1837010(2024)

Improved Two-Stage 3D Object Detection Algorithm for Roadside Scenes with Enhanced PointPillars and Transformer

Liangzi Wang^1,2,3, Miaohua Huang^1,2,3、*, Ruoying Liu^1,2,3, Chengcheng Bi^1,2,3, and Yongkang Hu^1,2,3

¹Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan 430070, Hubei, China

²Hubei Collaborative Innovation Center for Automotive Components Technology, Wuhan University of Technology, Wuhan 430070, Hubei, China

³Hubei Research Center for New Energy & Intelligent Connected Vehicle, Wuhan University of Technology, Wuhan 430070, Hubei, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Network architecture diagram

Download full size

Fig. 2. SimAM attention mechanism. (a) SimAM structure; (b) comparison diagram

Download full size

Fig. 3. Residual downsampling module. (a) Residual structure; (b) ResBlock structure; (c) BasicBlock structure; (d) BasicBlock_DS structure

Download full size

Fig. 4. Encoder implementation process

Download full size

Fig. 5. Detection results in DAIR-V2X-I roadside scenes

Download full size

Fig. 6. Detection results in KITTI vehicle scenes

Download full size

Fig. 7. Comparison of detection results for distant vehicle detection between proposed algorithm and PointPillars on the DAIR-V2X-I dataset. (a) PointPillars detection results; (b) BEV detection results of PointPillars; (c) enlarged detection results of PointPillars for distant vehicles; (d) detection results of proposed algorithm; (e) BEV detection results of proposed algorithm; (f) enlarged detection results of proposed algorithm for distant vehicles

Download full size

Fig. 8. Comparison of detection results for pedestrians on both sides of the road between proposed algorithm and PointPillars on the DAIR-V2X-I dataset. (a) PointPillars detection results; (b) enlarged BEV detection results of PointPillars for roadside pedestrians; (c) detection results of proposed algorithm in this paper; (d) enlarged BEV detection results of proposed algorithm for roadside pedestrians

Download full size

Fig. 9. Comparison of detection results for distant vehicles and pedestrians on both sides of the road between proposed algorithm and PointPillars on the KITTI dataset. (a) (d) Camera perspective; (b) (e) BEV detection results of PointPillars; (c) (f) BEV detection results of proposed algorithm

Download full size

Table 1. Comparison results between proposed algorithm and single-stage algorithms on DAIR-V2X-I validation set

View table

Table 1. Comparison results between proposed algorithm and single-stage algorithms on DAIR-V2X-I validation set

Algorithm	car /%			pedestrian /%			cyclist /%			mAP /% （moderate）	FPS / （frame/s）
Algorithm	easy	moderate	hard	easy	moderate	hard	easy	moderate	hard	mAP /% （moderate）	FPS / （frame/s）
IA-SSD^［25］	87.90	78.07	78.13	56.50	55.62	55.64	76.67	57.45	57.31	63.71	36.76
SECOND	87.74	78.62	78.66	60.87	55.72	55.89	80.55	61.83	61.85	65.39	25.77
PointPillars	87.59	78.51	78.57	59.18	57.99	58.04	80.35	60.59	60.57	65.70	30.67
Proposed algorithm	89.87	80.41	80.45	70.58	68.49	68.61	82.92	62.70	62.62	70.53	14.60

Table 2. Comparison results between proposed algorithm and a two-stage algorithm on DAIR-V2X-I validation set

View table

Table 2. Comparison results between proposed algorithm and a two-stage algorithm on DAIR-V2X-I validation set

Algorithm

car /%

pedestrian /%

cyclist /%

mAP /%

（moderate）

FPS /

（frame/s）

easy

moderate

hard

easy

moderate

hard

easy

moderate

hard

Part-A²

89.52

79.91

79.94

68.94

64.62

64.64

83.22

63.99

64.12

69.51

14.49

Proposed algorithm

89.87

80.41

80.45

70.58

68.49

68.61

82.92

62.70

62.62

70.53

14.60

Table 3. Comparison results between proposed algorithm and single-stage algorithms on KITTI validation set

View table

Table 3. Comparison results between proposed algorithm and single-stage algorithms on KITTI validation set

Algorithm	car /%			pedestrian /%			cyclist /%			mAP /% （moderate）	FPS / （frame/s）
Algorithm	easy	moderate	hard	easy	moderate	hard	easy	moderate	hard	mAP /% （moderate）	FPS / （frame/s）
IA-SSD	88.86	79.09	77.80	56.09	52.47	47.59	87.38	72.26	67.22	67.94	53.35
SECOND	88.03	77.77	75.32	53.60	48.69	44.11	81.17	68.32	63.75	64.93	42.19
PointPillars	87.33	77.01	74.61	54.35	49.55	45.07	80.65	63.76	59.95	63.44	48.08
Proposed algorithm	89.69	79.35	78.04	62.44	54.28	50.86	86.56	71.93	66.79	68.53	28.74

Table 4. Comparison results between proposed algorithm and a two-stage algorithm on KITTI validation set

View table

Table 4. Comparison results between proposed algorithm and a two-stage algorithm on KITTI validation set

Algorithm

car /%

pedestrian /%

cyclist /%

mAP /%

（moderate）

FPS /

（frame/s）

easy

moderate

hard

easy

moderate

hard

easy

moderate

hard

Part-A²

89.27

78.96

78.17

61.54

53.60

49.11

85.86

72.46

68.10

68.34

26.91

Proposed algorithm

89.69

79.35

78.04

62.44

54.28

50.86

86.56

71.93

66.79

68.53

28.74

Table 5. Impact of different modules on detection

View table

Table 5. Impact of different modules on detection

Algorithm	SimAM	ResBlock	Refinement network	moderate
Algorithm	SimAM	ResBlock	Refinement network	car /%	pedestrian /%	cyclist /%	mAP /%
Baseline				78.51	57.99	60.59	65.70
Experiment 1	√			78.76	59.01	61.98	66.58
Experiment 2		√		78.41	59.72	61.50	66.54
Experiment 3	√	√		78.69	59.46	62.12	66.76
Experiment 4			√	80.35	67.13	61.35	69.61
Experiment 5		√	√	80.37	67.92	61.88	70.06
Proposed algorithm	√	√	√	80.41	68.49	62.70	70.53

Tools

Get Citation

Copy Citation Text

Liangzi Wang, Miaohua Huang, Ruoying Liu, Chengcheng Bi, Yongkang Hu. Improved Two-Stage 3D Object Detection Algorithm for Roadside Scenes with Enhanced PointPillars and Transformer[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1837010

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Jan. 12, 2024

Accepted: Feb. 5, 2024

Published Online: Sep. 14, 2024

The Author Email: Miaohua Huang (mh_huang@163.com)

DOI:10.3788/LOP240516

CSTR:32186.14.LOP240516

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology