基于YOLOv5融合注意力机制的轻量级行人检测算法研究  

Lightweight pedestrian detection algorithm based on YOLOv5 fused attention mechanism

作　　者：卢嫚[1] 刘秀平[1] 冯国栋 Lu ManLiu;Xiuping Feng;Guodong(College of Electronics and Information,Xi'an Polytechnic University,Xi'an 710048,China)

机构地区：[1]西安工程大学电子信息学院,西安710048

出　　处：《国外电子测量技术》2023年第8期96-101,共6页Foreign Electronic Measurement Technology

基　　金：国家自然科学基金(62203344);陕西省技术创新引导专项-科技成果转移与推广计划项目(2020CGXNG-009)资助。

摘　　要：为实现行人自动、快速以及准确检测,提出一种高性能轻量化网络模型。首先,利用图像增强算法对比度自适应直方图均衡化进行图像预处理;其次,针对当前行人检测算法模型参数较多,计算量和内存占用大的问题,将Stem模块和ShuffleNetV2进行融合,并在每个深度卷积中将核大小从3个增加到5个,以获得更大的感受野,同时改进YOLOv5主干网络;最后,针对环境变化导致难以被准确检测的问题,利用ECA注意力模块可以提高数据的丰富性,从而进一步提高特征表达能力和鲁棒性。实验结果表明,改进后的算法能够较好地解决行人检测时易受环境干扰影响导致检测精度下降的问题,其平均检测精度可达94.6%,检测速度为35 fps。This paper proposes a high-performance and lightweight network model to achieve automatic,fast,and accurate pedestrian detection.Firstly,the image enhancement algorithm,contrast adaptive histogram equalization,is used for image pre-processing.Secondly,to address the problem that the current pedestrian detection algorithm model has more parameters and large computation and memory consumption,the Stem module and ShuffleNetV2 are fused and the kernel size is increased from 3 to 5 in each deep convolution to obtain a larger perceptual field,and increase the kernel size from 3 to 5 in each deep convolution to obtain a larger perceptual field,while improving the original YOLOv5 backbone network.Finally,to address the problem of difficulty to be detected accurately due to environmental changes,on this basis,feature representation and robustness can be improved with the ECA attention module.Then the improved algorithm is compared with other target detection algorithms.Experiments show that the improved algorithm effectively solves the problem that pedestrian detection is vulnerable to environmental interference,thereby improving accuracy,and its average detection accuracy can reach 94.6%with a detection speed of 35 fps.

关 键 词：行人检测 嵌入式设备 注意力机制 Stem模块 轻量级模型 YOLOv5 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]