一种基于YOLOv5的轻量型行人检测方法

A Lightweight Pedestrian Detection Method Based on YOLOv5

基于卷积神经网络的行人检测方法对行人检测性能有了很大的提升,但当前行人检测方法计算量大,占用了较多的计算机资源。为了解决这种问题,提出了一种改进的YOLOv5行人检测算法。该算法在保证检测精度不变、减小权重大小、提升检测速度的情况下,在原有的YOLOv5网络基础上引入轻量级卷积模块Ghost卷积,并且为了提高检测精度,在主干网络中加入轻量注意力机制ECA。为了进一步提高检测精度,将原有的特征融合网络PAN+FPN结构替换为加权双向金字塔结构BiFPN。通过实验结果表明,经过引入和替换模块后,模型网络精度保持不变,模型大小减小了约2.13倍,浮点型计算量减少了约2.51倍,检测速度(FPS)提高了约1.67倍。

Pedestrian detection method based on convolutional neural network has greatly improved the performance of pedestrian detection,but the current pedestrian detection method has a large amount of computation and occupies more computer resources.In order to solve this problem,an improved YOLOv5 pedestrian detection algorithm is proposedin this paper.In this algorithm,the lightweight convolution module Ghost convolution is introduced on the basis of the original YOLOv5 network under the condition of keeping the detection accuracy unchanged,reducing the weight and improving the detection speed.In order to improve the detection accuracy,the lightweight attention mechanism ECA is added to the backbone network.The original feature fusion network PAN+FPN is replaced by weighted bidirectional pyramid structure BiFPN.