Pantograph Safe Trigger Target Real-Time Detection and Localization Method Based on Fused Differential Convolutional

Zhanshan Yang; Ying Zhang; Hongzhi Du; Yanbiao Sun; Jigui Zhu

doi:10.3788/LOP232644

Laser & Optoelectronics Progress, Volume. 61, Issue 18, 1812001(2024)

Pantograph Safe Trigger Target Real-Time Detection and Localization Method Based on Fused Differential Convolutional

Zhanshan Yang, Ying Zhang, Hongzhi Du, Yanbiao Sun^*, and Jigui Zhu

Author Affiliations

State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Schematic diagram of detection system

Download full size

Fig. 2. Proposed lightweight detection framework

Download full size

Fig. 3. Feature map visualization. (a) Prediction results of RepGhost network; (b) prediction results of proposed model

Download full size

Fig. 4. Difference convolution. (a) Sampling and weighted aggregation; (b) difference convolution

Download full size

Fig. 5. Proposed module (RPCDC_Bolck). (a) Before re-parameterization; (b) after re-parameterization

Download full size

Fig. 6. Re-parameterization flow. (a) Initial structure; (b) fusion of dconv2 and BN2; (c) fusion of 1×1 cv1 and $δ$ ; (d) fusion of add and BN1; (e) fusion of dconv1 and BN1; (f) fusion of dconv1, dconv2, and BN3

Download full size

Fig. 7. Shared feature fusion module

Download full size

Fig. 8. Proposed Anchor Free detecting head

Download full size

Fig. 9. Pantograph trigger detection effect images of roof perspective. (a) Detection effect 1; (b) detection effect 2; (c) detection effect 3

Download full size

Fig. 10. Pantograph trigger detection effect images of oblique side viewing angle. (a) Detection effect 1; (b) detection effect 2

Download full size

Table 1. Effect of activation function on model performance
View table
Table 1. Effect of activation function on model performance
Activation function Parameter
Top1 /% Latency time /ms
ReLU 55.08 67.06
ReLU6 55.38 67.28
LeakyReLU 56.55 67.21
Hardswish 53.10 67.19
SiLU 55.09 67.14

Table 2. Multi-stage training strategy

View table

Table 2. Multi-stage training strategy

Type	Stage
Type	S₁	S₂	S₃	S₄
Pre-trained	Random	S1 Best	S2 Best	S3 Best
RandomFlip	√	√	√	√
Normalize	√	√	√	√
RandomScale	√	√	√	√
ElasticTransform	√	√	√	—
Blur	√	√	√	—
AddNoise	√	√	—	—
Mosaic	√	√	—	—
MixUp	√	—	—	—

Table 3. Distribution of large, medium, and small targets in VOC0712 dataset

View table

Table 3. Distribution of large, medium, and small targets in VOC0712 dataset

Type	Pixel range	Voc0712_training set	Voc0712_validation set	Voc07_testing set
Small	（0<a≤32）pixel×32 pixel	4389	4237	2741
Medium	32 pixel×32 pixel—（0<a≤96）pixel×96 pixel	7959	8065	5483
Large	96 pixel×（96≤a）pixel	11270	11303	6752
Total	0<a	23618	23605	14976

Table 4. Experimental results of VOC0712 dataset and PANTOGRAPH dataset

View table

Table 4. Experimental results of VOC0712 dataset and PANTOGRAPH dataset

Dataset	Method	Number of parameters /10³	FLOP/10⁶	AP50 /%	FPS /（frame/s）	mAP /%	mAPs /%	mAPm /%	mAPl /%
VOC0712	YOLOX^［20］	70.39	90.16	30.60	—	13.60	3.00	5.10	16.60
	NanoDet Plus	69.60	—	38.44	—	20.78	4.02	8.07	25.48
	FemtoDet	68.77	90.16	46.31	90	22.90	1.10	8.40	28.50
	Proposed	140.30	79.82	55.70	144	31.30	1.30	12.00	38.60
PANTOGRAPH	FemtoDet	68.77	90.16	97.40	92	75.20	—	—	75.20
PANTOGRAPH	Proposed	140.30	79.82	97.50	149	81.20	—	—	81.20

Table 5. Comparison of Neck detection performnce
View table
Table 5. Comparison of Neck detection performnce
Neck Number of parameters /10³ AP50 /% FPS /（frame/s）
FPN 176.31 45.04 81
PAN 79.81 39.91 75
FemtoDet 68.77 42.50 100
Proposed 87.65 43.61 95

Table 6. Ablation experimental results

View table

Table 6. Ablation experimental results

Method					Number of parameters /10³	FLOP /10⁶	AP50 / %	FPS / （frame/s）
Base	Backebone	Shareneck	Anchor free	Re-parameterization	Number of parameters /10³	FLOP /10⁶	AP50 / %	FPS / （frame/s）
√	×	×	×	×	68.7	90.16	46.31	90
√	√	×	×	×	124.6	85.74	52.26	124
√	√	×	×	√	124.6	85.74	52.26	135
√	√	√	×	√	130.9	83.33	54.57	139
√	√	√	√	√	140.3	79.82	55.73	144

Table 7. Multi-stage training effects
View table
Table 7. Multi-stage training effects
Task ID Stage Number of parameters /10³ FPS /（frame/s） mAP50 /%
S₁ S₂ S₃ S₄
1 √ × × × 140.3 144 45.40
2 √ √ × × 140.3 144 52.10
3 √ √ √ × 140.3 144 54.30
4 √ √ √ √ 140.3 144 55.70

Tools

Get Citation

Copy Citation Text

Zhanshan Yang, Ying Zhang, Hongzhi Du, Yanbiao Sun, Jigui Zhu. Pantograph Safe Trigger Target Real-Time Detection and Localization Method Based on Fused Differential Convolutional[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1812001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites