Three-Dimensional Object Detection in Substation Operation Scene Based on Attention Mechanism

Wei Gao; Boyang He; Ting Zhang; Meiqing Guo; Jun Liu; Huimin Wang; Xingzhong Zhang

doi:10.3788/LOP202259.2210010

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 22 >Page 2210010 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 22, 2210010 (2022)

Three-Dimensional Object Detection in Substation Operation Scene Based on Attention Mechanism

Wei Gao¹, Boyang He¹, Ting Zhang², Meiqing Guo², Jun Liu², Huimin Wang², and Xingzhong Zhang^2、*

Author Affiliations

¹Internet Department, State Grid Shanxi Electric Power Company, Taiyuan 030021, Shanxi , China

²College of Software, Taiyuan University of Technology, Jinzhong 030600, Shanxi , China

show less

DOI: 10.3788/LOP202259.2210010 Cite this Article Set citation alerts

Wei Gao, Boyang He, Ting Zhang, Meiqing Guo, Jun Liu, Huimin Wang, Xingzhong Zhang. Three-Dimensional Object Detection in Substation Operation Scene Based on Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2210010 Copy Citation Text

show less

Fig. 1. PowerNet structure

Download full size

Fig. 2. Local area. (a) Local area input; (b) local area representation

Download full size

Fig. 3. Channel direction attention structure diagram

Download full size

Fig. 4. Point direction attention structure diagram

Download full size

Fig. 5. Serial attention structure diagram

Download full size

Fig. 6. Dataset collector and sample illustration. (a) Dataset collector; (b) sample illustration

Download full size

Fig. 7. Data annotation. (a) PCAT annotated point cloud; (b) LabelImg annotated images; (c) label format

Download full size

Fig. 8. Loss curve and performance curve

Download full size

Fig. 9. Test result. (a) RGB images; (b) point cloud diagrams

Download full size

Channel-direction attention	Point-direction attention	Parallel			Serial
		AP		mAP	$A P$		mAP
		Pedestrian	Transformer	mAP	Pedestrian	Transformer	mAP
Two-layer MLP	7×7 filter	0.550	0.776	0.663	0.572	0.797	0.685
Two-layer MLP	5×5 filter	0.559	0.781	0.670	0.576	0.800	0.688
Four-layer MLP	7×7 filter	0.572	0.794	0.683	0.591	0.849	0.720
Four-layer MLP	5×5 filter	0.579	0.802	0.691	0.602	0.867	0.735

Table 1. Comparison of effects of different combinations of attention on network performance

Channel-direction attention （four-layer MLP）	Point-direction attention （5×5 filter）	AP		$m A P$
Channel-direction attention （four-layer MLP）	Point-direction attention （5×5 filter）	Pedestrian	Transformer	$m A P$
-	-	0.545	0.775	0.660
✓	-	0.572	0.790	0.681
-	✓	0.560	0.779	0.670
✓	✓	0.602	0.867	0.735

Table 2. Choice of attention structure

Cross entropy loss	Focal loss	$A P$		$m A P$
Cross entropy loss	Focal loss	Pedestrian	Transformer	$m A P$
✓	-	0.572	0.868	0.720
-	✓	0.602	0.867	0.735

Table 3. Choice of loss function

Method	Model	$A P$		$m A P$
Method	Model	Pedestrian	Transformer	$m A P$
3D to 2D	PIXOR^［9］	0.527	0.755	0.641
3D to 2D	Complex-YOLO^［10］	0.533	0.779	0.656
Voxelization	Vote3Deep^［15］	0.537	0.733	0.635
Voxelization	VoxelNet^［13］	0.531	0.802	0.667
Original point cloud	PointNet^［17］	0.540	0.762	0.651
	PointNet++^［18］	0.545	0.775	0.660
	Proposed method	0.602	0.867	0.735

Table 4. Performance comparison results of mainstream detection models

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information