摘要
针对当前行人检测方法计算量大、检测精度低的问题,基于YOLOv4-tiny提出一种改进的行人检测算法。引入通道注意力和空间注意力模块(CBAM)至CSPDarknet53-tiny网络中,通过学习图像的位置信息和通道信息得到更加丰富的特征;在骨干网络CSPDarknet53-tiny之后引入空间金字塔池化模块,能够极大地增加感受野,分离出最显著的上下文特征;使用CIoU损失函数对算法的多任务联合损失进行了优化。实验分别使用INRIA和WiderPerson中的训练集作为训练模型,INRIA和WiderPerson中的测试集分别作为测试集来验证模型。实验表明,对比YOLOv4-tiny目标检测模型,改进后的网络在INRIA测试集中精确度、召回率和平均精度值分别提升了6.23%、3.15%和6.12%;改进后的网络在WiderPerson测试集中精确度、召回率和平均精度值分别提升了3.65%、3.28%和4.41%。改进后的模型在几乎不影响检测实时性的前提下,更易于行人特征提取,提高了检测精度。
As the current pedestrian detection method has the problems of large computation and low detection accuracy,an improved pedestrian detection method based on YOLOv4-Tiny was proposed.This method introduces Convolutional Block Attention Module into CSPDarknet53-tiny network to get richer features by learning the position information and channel information of the image,adds the spatial pyramid pooling module following CSPDarknet53-tiny,which can greatly increase the receptive field and isolate the most significant context features,and uses CIoU loss function to optimize the combined loss of multiple tasks.In the experiment,the training set in INRIA and WiderPerson are used as the training model,and the test set in INRIA and WiderPerson are used as the test set to verify the model.Compared with YOLOv4-Tiny,the precision,recall and average accuracy of the improved YOLOv4-Tiny network in INRIA test set are increased by 6.23%,3.15% and 6.12%,respectively,and the improved network increased the precision,recall,and average accuracy in the WiderPerson test set by 3.65%,3.28%,and 4.41%,respectively.It is found that this improved model can extract pedestrian features more easily and improve the detection accuracy on the premise that the real-time detection is hardly affected.
作者
常青
韩文
王清华
李振华
CHANG Qing;HAN Wen;WANG Qinghua;LI Zhenhua(School of Science,Nanjing University of Science and Technology,Nanjing 210094,China;School of Information Science and Technology,Nanjing Forestry University,Nanjing 210037,China)
出处
《光学技术》
CAS
CSCD
北大核心
2022年第1期80-85,共6页
Optical Technique