【作者】 程卓；

【导师】 翟象平；

【作者基本信息】 南京航空航天大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 佩戴安全帽是配电站房安全生产运维的必要条件,目前主要依靠人工对视频中的人员进行安全帽佩戴实时监控。人工监控的方式不仅效率低下,而且长时间关注视频也极大损害了监控人员的身体健康。虽然基于传统图像处理和机器学习的方法可以在一定程度上进行自动化的识别检测,但是该类方法往往对环境要求较为严格,模型开发周期较长、泛化能力较弱、检测精度和实时性较差,难以进行实际的应用。随着时代的发展,物联网、云计算、大数据、人工智能等新兴技术正在蓬勃发展,尤其以卷积神经网络为代表的深度学习算法可以通过训练自主提取图像特征,利用迁移学习快速适应新任务,极大推动了视频智能检测的发展。因此,本文面向配电站房应用场景研究了一种基于卷积神经网络的安全帽佩戴检测算法,主要内容如下:(1)提出一种双模自适应的特征融合方式。传统的特征金字塔网络存在两个问题:最高层金字塔特征信息损失,不同特征层之间具有语义鸿沟。为了解决这两个问题,本文首先利用镜像特征层合并的操作直接将不同层级的金字塔特征进行融合,从而保证特征在传输过程中的不会损失;然后通过对每个尺度特征图的融合空间权重进行自适应学习,解决不同特征层之间的语义差异。(2)设计了一种基于位置-通道协调的注意力模块。针对传统的通道注意力机制只考虑编码通道间信息而忽略了位置信息的问题,本文利用两个一维的全局池操作分别将垂直方向和水平方向上的输入特征聚合成两个独立的方向感知特征映射,然后将携带方向信息的特征映射编码成两个独立的注意映射,分别对输入特征映射在各自空间方向上的长期依赖关系进行捕获,更加有利于CNN在空间范围中精确对待检测的目标区域进行定位。(3)面对数据集采集和制作繁琐的问题,本文根据实际应用场景研究了多种样本增广的方法,并制作了一组配电站房安全帽检测数据集。首先,对图像添加高斯噪声和椒盐噪声模拟摄像头工作及传输数据过程的模糊;其次,针对人员的位置、角度、环境光照等情况依次做了旋转、拉伸、灰度变化等操作;最后针对人员遮挡等情况采用了掩码遮挡和图像覆盖等方法。基于自制的配电站房安全帽检测数据集,经系列实验证明,本文提出的方法检测效果优于主流的目标检测算法,检测的平均精度达到90.78%,速度达到28.93 FPS,在满足实时性需求的同时达到高精度的检测效果,对配电站房安全生产监督具有重要的意义。更多还原

【Abstract】 It is a necessary condition for safe production for workers in substations to wear safety helmets.At present,it mainly relies on manual monitoring of the wearing of safety helmets by the personnel in the video in real time.The manual monitoring method is inefficient,and it will cause harm to the health of the monitoring personnel when focusing on the video for a long time.The method based on traditional image processing can automatically detect targets to a certain extent,but it often has stricter environmental requirements,long model development cycle,weak generalization ability,poor detection accuracy,and it is always difficult to put into practical application.With the development of emerging technologies such as the Internet of Things,cloud computing,big data,and artificial intelligence,deep learning algorithms represented by convolutional neural networks can autonomously extract image features through training,and use transfer learning to quickly adapt to new tasks,it greatly promotes the development of intelligent detection technology.Therefore,this paper studies a safety helmet wearing detection algorithm based on convolutional neural network for the application scenario of power distribution station.The main contents are as follows:(1)This paper proposes a bimodal adaptive feature fusion method.There are two problems in traditional feature pyramid networks: the loss of feature information in the highest-level pyramid,and the semantic gap between different feature layers.In order to solve these two problems,this paper firstly uses the mirror feature layer merging operation to directly fuse the pyramid features of different levels to ensure that the features will not be lost during the transmission process;then adaptively learn the fusion space weights of the feature maps of different scales to resolve the semantic differences between different feature layers.(2)This paper designs an attention module based on position-channel coordination.Aiming at the problem that the traditional channel attention mechanism only considers the information between encoding channels and ignores the position information,this paper uses two one-dimensional global pooling operations to aggregate the input features in the vertical and horizontal directions into two independent directional feature maps.Then encoding the feature maps that carrying direction information into two independent attention maps,which capture the long-term dependencies of the input feature maps in their respective spatial directions.It is beneficial for CNN to accurately locate the target area in the spatial range.(3)Faced with the tedious problem of data set collection and production,this paper studies a variety of sample augmentation methods according to practical application scenarios,and produces a set of safety helmet detection data sets for distribution substations.First,we add Gaussian noise and salt and pepper noise to the images to simulate the impact of the work of camera and data transmission process.secondly,by rotating,stretching,and grayscale changes on the images to simulate changes in personnel positions,angles,and ambient lighting.Finally,we use Gridmask and Cutmix to simulate scenes where people are occluded.Based on the self-made distribution station safety helmet data set,a series of experiments show that the detection effect of the method proposed in this paper is better than the mainstream target detection algorithm,the average detection accuracy reaches 90.78%,and the speed reaches 28.93 FPS.It is of great significance for the safety production supervision of the distribution station to achieve high-precision detection results while meeting the real-time requirements.更多还原