基于YOLOv5s融合注意力机制的轻量化行人检测算法

Lightweight Pedestrian Detection Algorithm Based on YOLOv5s Fused Attention Mechanism

基于YOLOv5s算法的行人检测技术在自动驾驶方面具有广泛应用,对YOLOv5s算法进行轻量化可以节省计算资源、存储空间和传输带宽,该项工作具有重要的实际意义。为提高模型对关键特征的关注能力,在骨干网络中增加CBAM注意力机制以抑制无效信息;同时,为大幅降低模型的复杂度、参数量和计算量,引入GhostNet网络中的Ghost结构代替YOLOv5s原有的卷积和Neck模块的C3结构。为了验证轻量化后算法的优势,基于PASCAL VOC 2007数据集、WiderPerson数据集对原YOLOv5s以及改进算法进行测试。实验结果表明,轻量化算法能够大幅度降低参数量、计算量,同时保证了YOLOv5s算法原有的检测识别准确率。

The pedestrian detection technology based on YOLOv5s algorithm has been widely applied in autonomous driving.Lightweight improvements on YOLOv5s algorithm can be reduced computational resources,storage space,and transmission bandwidth.This work is of great practical significance.To enhance the model's attention to key features,a CBAM attention mechanism can be fused into the backbone network to suppress irrelevant information.Simultaneously,to reduce model complexity,parameter count,and computational requirements,a Ghost structure of GhostNet network is incorporated to replace the original convolutional of YOLOv5s and C3 structure of Neck module.To verify the advantages of the lightweight algorithm,based on PASCAL VOC 2007 dataset and WiderPerson dataset,the original YOLOv5s and the improved algorithm are tested.The results demonstrate that the lightweight algorithm can greatly reduce parameter count and computational requirements while maintaining the detection and recognition accuracy of the original YOLOv5s algorithm.