Improved PointPillar point cloud object detection based on feature fusion

Yong ZHANG; Zhiguang SHI; Qi SHEN; Yan ZHANG; Yu ZHANG

doi:10.37188/OPE.20233119.2910

Optics and Precision Engineering, Volume. 31, Issue 19, 2910(2023)

Improved PointPillar point cloud object detection based on feature fusion

Yong ZHANG, Zhiguang SHI^*, Qi SHEN, Yan ZHANG, and Yu ZHANG

National Key Laboratory of Science and Technology on Automatic Target Recognition， College of Electronic Science and Technology， National University of Defense Technology， Changsha410073， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Structure of PointPillar network

Download full size

View in Article

Fig. 2. Structure of pillar-FFNet

Download full size

View in Article

Fig. 3. Structure of CR block

Download full size

View in Article

Fig. 4. Structure of RBNet

Fig. 5. CAMA module

Fig. 6. PR curves for four comparison algorithms on KITTI validation dataset

Download full size

View in Article

Fig. 7. Visualisation of detection results

Download full size

View in Article

Table 1. Result for 3D detection on KITTI validation dateset

View table

View in Article

Table 1. Result for 3D detection on KITTI validation dateset

Algorithm	Car（IoU=0.7）			Pedestrian（IoU=0.5）			Cyclist（IoU=0.5）			mAP/% （middle）	FPS/ （frame·s^-1）
Algorithm	easy	middle	hard	easy	middle	hard	easy	middle	hard	mAP/% （middle）	FPS/ （frame·s^-1）
Second	88.25	78.56	77.14	54.72	50.65	45.93	80.72	63.92	60.89	64.37	26.25
PointPillar	87.08	77.55	75.82	54.43	49.31	45.56	81.45	63.16	59.91	63.34	36.49
PillarNet	87.25	77.72	76.49	51.43	47.53	44.42	81.95	63.20	59.12	62.81	19.96
Pillar-FFNet	87.92	78.17	76.62	56.24	51.44	46.72	85.47	65.55	61.49	65.05	10.00

Table 2. Result for 3D detection on the DAIR-V2X-I validation dateset

View table

View in Article

Table 2. Result for 3D detection on the DAIR-V2X-I validation dateset

Algorithm	Car（IoU=0.7）			Pedestrian（IoU=0.5）			Cyclist（IoU=0.5）			mAP/% （middle）	FPS/ （frame·s^-1）
Algorithm	easy	middle	hard	easy	middle	hard	easy	middle	hard	mAP/% （middle）	FPS/ （frame·s^-1）
Second	80.13	68.84	68.87	64.35	60.41	60.57	64.64	39.32	39.43	56.19	12.69
PointPillar	80.04	68.86	68.89	63.35	62.06	62.09	61.51	37.51	37.61	56.14	15.13
PillarNet	80.21	69.82	69.86	64.89	60.82	60.84	62.98	38.66	40.06	56.43	11.42
Pillar-FFNet	80.37	69.03	69.06	65.44	62.23	62.26	66.22	39.35	39.43	56.87	8.18

Table 3. Effect of different fusion methods on detection of Pillar-FFNet detection heads
View table
View in Article
Table 3. Effect of different fusion methods on detection of Pillar-FFNet detection heads
Algorithm Car Pedestrian Cyclist mAP/%
Pillar-FFNet 78.17 51.44 65.55 65.05
experiment 1 77.30 52.83 61.69 63.94
experiment 2 78.21 52.76 62.98 64.65
experiment 3 74.13 46.86 52.34 57.78
experiment 4 77.31 49.19 62.83 63.24

Table 4. Effect of different attention modules of Pillar-FFNet on detection
View table
View in Article
Table 4. Effect of different attention modules of Pillar-FFNet on detection
Algorithm Car Pedestrian Cyclist mAP/%
Pillar-FFNet 78.17 51.44 65.55 65.05
experiment 5 77.73 51.35 63.63 64.23
experiment 6 77.75 47.04 63.21 62.67
experiment 7 77.53 51.41 63.86 64.27
experiment 8 77.98 49.86 60.59 62.81

Table 5. Effect of modules designed in paper on detection
View table
View in Article
Table 5. Effect of modules designed in paper on detection
Algorithm Car Pedestrian Cyclist mAP/%
CAMA+MFHead 78.17 51.44 65.55 65.05
CAMA 78.07 48.08 61.23 62.46
MFHead 76.85 52.94 64.45 64.74

Tools

Get Citation

Copy Citation Text

Yong ZHANG, Zhiguang SHI, Qi SHEN, Yan ZHANG, Yu ZHANG. Improved PointPillar point cloud object detection based on feature fusion[J]. Optics and Precision Engineering, 2023, 31(19): 2910

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites