基于改进YOLOv5的工程场景行人检测方法

Pedestrian Detection Method in Engineering Scene Basedon Improved YOLOv5

针对复杂工程场景常用的行人检测方法(尤其在小目标检测方面)精度低、复杂度高的问题,提出一种基于YOLOv5网络的改进识别方法。在骨干网络与颈部网络引入ECA注意力机制,提升模型对通道特征的关注度以抑制背景噪声;使用加权双向特征金字塔结构BIFPN对颈部网络进行修改,加强模型对不同尺度特征融合;使用Ghost模块替换骨干网络与颈部网络的部分卷积,减少模型参数、缩小体积。结果表明:提出的改进模型检测精度达到了88.4%,同时,模型的复杂度(参数量与模型大小)仅为13.5×10^(6)与6.67 MB;与目前主流的深度学习方法相比,该算法在检测精度与复杂度上具有更好的性能,在复杂的场景下具有较好的识别效果。

Aiming at the problems of low accuracy and high complexity of pedestrian detection methods commonly uses in complex engineering scenes(especially in small target detection),an improved recognition method based on YOLOv5 network is proposed.Firstly,in order to enhance the model's attention to channel features and suppress background noise,ECA attention mechanism is introduced in the backbone network and neck network;secondly,in order to strengthen the fusion of different scale features,the weighted bidirectional feature pyramid BIFPN is used to modify the bottleneck layer;finally,in order to reduce model parameters and the volume,Ghost module is used to replace part of the convolution of backbone network and bottleneck layer.The results show that the detection accuracy of the improved model proposes in this paper reaches to 88.4%,and the parameter numbers and model size is only 13.5*10^(6) and 6.67 MB,respectively.Compared with the mainstream deep learning methods,the modified algorithm has better performance in the detection accuracy and complexity,and has better recognition effect in complex scenes.