Occluded Pedestrian Detection Algorithm Based on Attention Mechanism

Ziyin Zou; Shaoyan Gai; Feipeng Da; Yu Li

doi:10.3788/AOS202141.1515001

Acta Optica Sinica, Volume. 41, Issue 15, 1515001(2021)

Occluded Pedestrian Detection Algorithm Based on Attention Mechanism

Ziyin Zou^1,2, Shaoyan Gai^1,2、*, Feipeng Da^1,2,3, and Yu Li^1,2

Author Affiliations

¹School of Automation, Southeast University, Nanjing, Jiangsu 210096, China

²Key Laboratory of Measurement and Control of Complex Systems of Engineering, Ministry of Education, Southeast University, Nanjing, Jiangsu 210096, China

³Shenzhen Research Institute, Southeast University, Shenzhen, Guangdong 518063, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Overall architecture of model

Download full size

Fig. 2. Attention module. (a) SENet module; (b) our module

Download full size

Fig. 3. Diagrams of target prediction. (a) Prediction result 1; (b) prediction result 2

Download full size

Fig. 4. Comparison of detection results of pedestrians. (a)--(d) Detection results of baseline; (e)--(h) detection results of our method

Download full size

Table 1. Occlusion status partition of CityPersons dataset
View table
Table 1. Occlusion status partition of CityPersons dataset
Type Bare Partial Reasonable Heavy
Occlusion [0,0.10] (0.10,0.35) [0,0.35) [0.35,∞)

Table 2. Number of pedestrians of mutual occluding datasets
View table
Table 2. Number of pedestrians of mutual occluding datasets
Dataset CityPersons WiderPerson
#person/img 6.47 28.87
IOU is larger than 0.3 0.96 9.21
IOU is larger than 0.5 0.32 2.15
IOU is larger than 0.7 0.08 0.24

Table 3. Results for each added module $M_{MR}^{- 2}$ %
View table
Table 3. Results for each added module $M_{MR}^{- 2}$ %
Method Reasonable Heavy Partial Bare
Baseline 15.7 56.3 18.8 9.7
Baseline+AM 13.6 52.8 15.3 7.5
Baseline+AM+RRPN 13.1 52.1 14.5 7.3
Baseline+AM+RRPN+DN 12.7 51.9 13.2 6.6

Table 4. Experimental results of $M_{MR}^{- 2}$ on CityPersons%

View table

Table 4. Experimental results of $M_{MR}^{- 2}$ on CityPersons%

Method	Backbone	Reasonable	Heavy	Partial	Bare
Faster-RCNN	VGG-16	15.4	-	-	-
WiderPerson	VGG-16	11.1	-	-	-
OR-CNN	VGG-16	12.8	55.7	15.3	6.7
TLL	ResNet-50	15.5	53.6	17.2	10.0
TLL+MRF	ResNet-50	14.4	52.0	15.9	9.2
PedJointNet	ResNet-50	13.5	52.1	-	-
RepLoss	ResNet-50	13.2	56.9	16.8	7.6
Adaptive-NMS	ResNet-50	10.8	54.0	11.4	6.2
Our method	ResNet-50	12.7	51.9	13.2	6.6

Table 5. Experimental results on WiderPerson
View table
Table 5. Experimental results on WiderPerson
Method AP /% Recall /%
Baseline 84.43 89.21
Our method 85.12 89.94

Tools

Get Citation

Copy Citation Text

Ziyin Zou, Shaoyan Gai, Feipeng Da, Yu Li. Occluded Pedestrian Detection Algorithm Based on Attention Mechanism[J]. Acta Optica Sinica, 2021, 41(15): 1515001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites