Multi-spectral Pedestrian Detection Based on Deformable Convolution and Multi-Scale Residual Attention

Fig. 9. Visualization comparison of detection results by different algorithms. (a) Complex background scene; (b) heavily occlusion pedestrian target scene; (c) blur background small target scene; (d) weak feature pedestrian target scene

Download full size

Table 1. Experimental environment configuration
View table
Table 1. Experimental environment configuration
Parameter Experimental environment
Operating system Windows 11
CPU Intel Core i5-12500H @ 2.40 GHz
GPU GeForce RTX 3060
Memory 32 Gb
Python 3.8
Deep learning framework PyTorch 1.7.0、CUDA 11.1

Table 2. Ablation experimental results

View table

Table 2. Ablation experimental results

Group	DCN	MRA	Small target	mAP@0.5 /%	mAP@0.5∶0.95 /%	FPS /（frame/s）
1				49.7	26.3	63.8
2	√			51.1	26.9	60.7
3		√		50.6	26.5	61.4
4			√	51.7	27.0	60.3
5	√	√		51.9	27.3	58.7
6	√		√	53.5	27.7	56.4
7		√	√	52.6	27.5	56.8
8	√	√	√	54.8	28.2	54.4

Table 3. Comparative experimental results of fusion methods
View table
Table 3. Comparative experimental results of fusion methods
Method Platform FPS /（frame/s）
ACF MATLAB 0.37
Halfway Fusion TITAN X 2.36
IAF-RCNN TITAN X 4.76
CIAN 1080Ti 14.28
DSMN 3090Ti 1.32
MBNet 1080Ti 14.29
Ours 3060Ti 54.40

Table 4. Comparative experimental results of different methods
View table
Table 4. Comparative experimental results of different methods
Method mAP@0.5 /% FPS /（frame/s）
Faster R-CNN 46.7 14.8
SSD 38.6 42.9
YOLOv4-tiny 42.5 53.6
YOLOv5s 49.7 63.8
YOLOv7 51.6 39.3
Ours 54.8 54.4

Tools

Get Citation

Copy Citation Text

Guoli Zhang, Shuai Chang, Yansong Song, Tianci Liu. Multi-spectral Pedestrian Detection Based on Deformable Convolution and Multi-Scale Residual Attention[J]. Laser & Optoelectronics Progress, 2024, 61(10): 1037004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites