A Pedestrian Detection Network Based on EfficientDet Combined with SCConv
Shu‐Heng Chen, Manlu Liu, Li Hu, Mei Wang · 2024
Addressing the common issues of false detections and missed detections in pedestrian detection tasks, especially when dealing with multi-scale targets and occlusions, we propose a single-stage target detection network architecture based on EfficientDet, named PSC-EfficientDet-D0. This architecture aims to enhance detection speed and model efficiency while ensuring accuracy. Specifically, we first restructure the inverted residual bottleneck network to improve the backbone network EfficientNet, which boosts accuracy while maintaining relatively low parameter counts and frame rates. Additionally, we introduce NWD Loss to further enhance the model’s capability to detect targets of varying scales and improve target localization. Experimental results show that our improved method has achieved significant performance enhancements, especially in the detection of tiny targets. Rather than EfficientNet-D0, our PSC-EfficientNet-D0 has demonstrated $2.1 \%$ increase in accuracy and $A P_{50: 95}^{{small }}$ on the Pascal VOC 2007 dataset. For pedestrian detection tasks, there has been a $1.1 \%$ improvement in average precision (AP). The enhanced EfficientDet exhibits stronger robustness and generalization capabilities.