目录
摘 要 Ⅰ
ABSTRACT Ⅱ
第一章 文献综述 7
1研究背景及意义 7
2 国内外研究现状 7
3 研究内容与创新点 8
4 论文内容安排 9
第二章 卷积神经网络的目标检测算法研究 10
1 图像处理技术 10
2卷积神经网络 11
2.1 卷积神经网络结构 11
2.2卷积神经网络特性 12
2.3 常用卷积神经网络结构 13
3 目标检测方法分析 14
3.1 基于传统特征目标检测算法 14
3.2 基于卷积神经网络的目标检测算法 15
3.3不同目标检测方法对比 16
第三章 基于YOLOV5安全帽佩戴检测算法 17
1 目标检测模型 YOLO 算法 17
1.1 Yolo基本原理 17
1.2 非极大值抑制 18
1.3 损失函数 19
2 YOLOV5网络结构 20
3 基于密集强化的网络结构 22
第四章 实验结果和分析 24
1数据集介绍 24
2 实验环境 25
3训练过程与网络参数设置 26
4 实验评价标准 27
5实验结果和分析 28
第五章 结论与展望 29
参考文献 31
附 录 33
致 谢 41
基于YOLO V5电动车安全帽佩戴检测研究
摘 要
随着城市道路交通堵塞状况及汽车尾气污染日益增多，电动车以然变成人们短途出门的最佳选择，可某些驾驶人员缺乏相应的安全意识，导致了交通事故频频发生，因此佩戴安全帽成为当今电动车安全驾驶的重要保障。一般情况，电瓶车都处于高速移动的情况下，因此对于视频中目标的检测往往存在一定的难度，当前主流安全帽检测算法在检测速度和密集人群检测上都存在一定缺陷。当在密集的人群环境下，高度重叠的区域会出现漏检等情况的发生。本文采用一种在YOLOV5目标检测算法的结构基础上添加一 *51今日免费论文网|www.51jrft.com +Q: ^351916072^ 
个预测分支方式，实现在密集场景中安全帽的快速检测。YOLOV5是一种基于端到端的网络结构，可以实现视频中待检目标的快速检测，在重叠的场景采用EMD Loss 损失函数和Set NMS 后处理方法，实现一个候选框只检测一个待测目标，而不是预测所有的待测行人，从而提升漏检率。通过对最终实验训练模型测试识别效率达到90.49%，充分证实了本文算法的有效性和可行性。
关键字：深度学习；安全帽检测；YOLOV5；密集环境
RESEARCH ON WEARING DETECTION OF ELECTRIC VEHICLE SAFETY HELMET BASED ON YOLOV5
ABSTRACT
In recent years, electric vehicles have gradually become a means of short distance
travel. However, due to the weak traffic awareness of some drivers, electric vehicle traffic accidents occur frequently. Therefore, wearing safety helmet has become an important guarantee for the safe driving of electric vehicles. Generally, the battery car is moving at a high speed, so it is difficult to detect the target in the video. The current mainstream helmet detection algorithm has some defects in the detection speed and dense crowd detection. When in a dense crowd environment, high overlapping areas will be missed. In this paper, a prediction branch method is added to the structure of yoov5 target detection algorithm to realize the rapid detection of safety helmet in dense scenes. Yoov5 is an endtoend network structure, which canrealize the fast detection of the target in the video. In the overlapping scene, EMD loss function and set NMS postprocessing method are used to realize a candidate box to detect only one target to be tested, instead of predicting all the people to be tested, so as to improve the miss detection rate. Through the final experimental training model test, the recognition efficiency reaches 90.49%, which fully proves the effectiveness and feasibility of the algorithm.

原文链接：http://www.jxszl.com/jtgc/jjtys/610118.html