基于监控视频的电梯轿厢内禁入目标检测算法研究与实现.pdfVIP

摘要 摘要 箱式电梯的使用在生活中愈加普及，但电梯轿厢空间狭小，一些乘客携带劣犬等物 体进入电梯可能会造成乘客的人身安全损失，或引起电梯运行安全问题。利用计算机视 觉技术实现基于监控视频的禁入目标检测，在电梯安防领域具有较强的学术意义和工程 意义。本文应江苏省特种设备安全检验监督研究院无锡分院要求，研究基于电梯监控视 频的禁入目标检测算法，并设计了禁入目标检测子系统，提供检测数据给电梯监控平台。 针对电梯监控视频的分析处理，先判断了电梯内是否有物体，再执行目标识别。本 文以犬类为禁入目标的典型代表，对犬类目标识别开展实验，针对电梯轿厢环境，制作 了本文所需的数据集，选用了三种不同类型的目标识别算法展开研究并改进。基于HOG 特征和 SVM 分类器的算法比较了不同人工特征的识别效果；采用 HOG 结合栈式自编 码器网络的改进算法相比原算法识别率有约 10%的提升；基于 YOLO 卷积神经网络的 算法在添加批规范化操作、改善损失函数等改进后识别效果较好，测试集下查全率达 83%。最后比较了三个算法对电梯监控视频的检测效果，改进 YOLO 算法在 GPU 支持 的条件下有较好的识别效果，并且能够满足在线检测的要求。 本文采用 B/S 架构设计实现了基于电梯监控视频的禁入目标检测子系统，完成了禁 入目标检测、目标检测数据存储以及检测结果展示网站。在实验轿厢中运行结果表明检 测系统能够识别禁入目标，系统运行稳定。 关键词：电梯轿厢，目标检测，栈式自编码器，YOLO ，卷积神经网络 I Abstract Abstract Detection of forbidden target based on surveillance video has strong academic and engineering meaning for elevator security. According to the requirement of Jiangsu Province Special Equipment Safety Supervision Inspection Institute.Branch of Wuxi, the detection algorithms of forbidden target based on elevator surveillance video are studied and the subsystem is designed. The object detection will not work if there is nothing in the elevator. Dog is selected as prohibited target to carry out experiments in this paper, and three different types of object detection algorithms are implemented. Algorithm based on HOG and SVM is applyled and the recognition effect based on different features is compared. The algorithm using HOG combined with SAE improves the recognition rate of the original algorithm by about 10%. YOLO-based algorithm is improved by adding BN operation and a better loss function. After improvement, the recall rate is up to 83%. Finally, the processing resu