Two-Channel SSD Pedestrian Head Detection Algorithm Based on Multi-Scale Feature fusion

Yongfu Zhou; Wenlong Li; Ranran Hu

doi:10.3788/LOP202158.2415009

Laser & Optoelectronics Progress, Volume. 58, Issue 24, 2415009(2021)

Two-Channel SSD Pedestrian Head Detection Algorithm Based on Multi-Scale Feature fusion

Yongfu Zhou¹, Wenlong Li^1,2, and Ranran Hu^2、*

Author Affiliations

¹School of Management Engineering, Jilin Communications Polytechnic, Changchun, Jilin 130012, China

²School of Electronic Information Engineering, Changchun University of Science and Technology, Changchun, Jilin 130022, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Structure of SSD300 network

Download full size

Fig. 2. Structure of two-channel network

Download full size

Fig. 3. Improved two-channel SSD network model based on feature fusion

Download full size

Fig. 4. Partial depth images. (a) Image 1; (b) image 2; (c) image 3

Download full size

Fig. 5. Color images and depth images under different operations. (a) Original images; (b) 180 ° rotation; (c) x-axis reversal; (d) y-axis reversal

Download full size

Fig. 6. Training loss and accuracy curves. (a) Training loss curve; (b) accuracy curve

Download full size

Fig. 7. Relationship between precision and recall of different networks

Download full size

Fig. 8. Detection results of SSD network and two-channel SSD network. (a) Original images; (b)SSD network; (c) two-channel SSD network

Download full size

Fig. 9. Detection results of SSD network and improved SSD network with multi-scale feature fusion. (a) Original images; (b) SSD network; (c) two-channel SSD network

Download full size

Fig. 10. Detection results of each algorithm under different illumination variation conditions. (a) SSD algorithm; (b) DSSD algorithm; (c) improved algorithm

Download full size

Fig. 11. Detection results of each algorithm under different occlusion conditions. (a) SSD algorithm; (b) DSSD algorithm; (c) improved algorithm

Download full size

Table 1. Size of default box for each layer

View table

Table 1. Size of default box for each layer

Parameter	MDConv 4_3_fusion	FC 7_fusion	Conv 8_2_fusion
Feature map size/(pixel×pixel)	38×38	19×19	10×10
Number of default boxes	4	6	3
a_r	{1,1,2,1/2}	{1,1,2,1/2,3,1/3}	{1,2,1/2}
Small side length	30	60	111
Large side length	60	111	-

Table 2. Average accuracy and speed of target detection in different prior frame networks
View table
Table 2. Average accuracy and speed of target detection in different prior frame networks
Network mAP/% FPS
Simplified SSD network 83.70 33
SSD network 84.90 29

Table 3. Average detection accuracy of two-channel SSD network unit: %
View table
Table 3. Average detection accuracy of two-channel SSD network unit: %
Fusion mode Weight ratio RGB image Depth image Two-channel SSD
Concat 83.70 82.61 89.86
Eltwise
0.7∶0.3 93.59
0.5∶0.5 91.94
0.3∶0.7 87.46

Table 4. Average detection accuracy of multi-scale feature fusion network unit: %
View table
Table 4. Average detection accuracy of multi-scale feature fusion network unit: %
Parameter SSD DSSD Multi-scale SSD
mAP 83.70 88.43 91.47

Table 5. Average detection accuracy of different models unit: %
View table
Table 5. Average detection accuracy of different models unit: %
Parameter SSD RGB-D+YOLOv2 RGB-D+Faster-RCNN Ref. [18] Ref. [19] Ref. [25] Proposed model
mAP 84.90 92.95 93.14 92.01 90.78 95.47 97.80

Tools

Get Citation

Copy Citation Text

Yongfu Zhou, Wenlong Li, Ranran Hu. Two-Channel SSD Pedestrian Head Detection Algorithm Based on Multi-Scale Feature fusion[J]. Laser & Optoelectronics Progress, 2021, 58(24): 2415009

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites