Little-YOLOv4: A Lightweight Pedestrian Detection Network Based on YOLOv4 and GhostNet

• Published in: Wireless communications and mobile computing 2022-09, Vol.2022, p.1-13
• Main Authors: Zheng, Hongtao, Liu, Hong, Qi, Wei, Xie, Hao
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Deep learning; Feature extraction; Neural networks; R&D; Research & development; Spatial data
• ISSN: 1530-8669; 1530-8677
• DOI: 10.1155/2022/5155970

Nowadays, the areas such as intelligent-assisted driving, intelligent monitoring, pedestrian analysis, and attitude detection are developing rapidly. Inseparable from these fields is pedestrian detection technology. In order to deploy pedestrian detection algorithms on devices with limited hardware resources, a recognition algorithm that takes into account both performance and speed is required. Thus, a Little-YOLOv4 network structure is proposed, where the GhostNet is used to extract image features and the PANet is ameliorated by adding BiFPN path fusion which can integrate richer semantic features and preserve spatial information. The DO-DConv and DSC take place of standard convolution and the ReLU6 replaces the Leaky ReLU, which reduce the computational cost. The squeeze-and-excitation network is added to YOLOv4 head network, which could greatly reduce the interference information. The pedestrian detection results show that the mAP is 90.11% and the FPS is 79 by using Little-YOLOv4. The Little-YOLOv4 network structure could achieve a good compromise between algorithmic accuracy and speed.