基于DCNN的井下行人检测系统的研究与设计

发布时间：2018-03-22 19:43

本文选题：井下行人检测　切入点：卷积神经网络　出处：《西安科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：煤炭在我国能源利用中占据着举足轻重的地位,煤矿安全尤其是井下生产环境的安全则一直是煤矿行业的重中之重。目前煤矿企业对于井下工作人员的检测主要依托于已装备的井下人员定位系统等,这些技术的应用可以有效地进行人员的定位和识别,但是在使用过程当中也出现替下、捎卡等情况,其精准度不高,智能化水平较低,特别是当监控人员疏忽时,存在很大的安全隐患。基于这样的背景,本文结合DCNN(深度卷积神经网络)在视频图像识别领域中的应用和井下装备的工业视频监控系统,提出了一种基于DCNN的矿井井下行人检测技术。为提高检测速度,采用了 YOLO目标检测系统,并针对井下特殊环境的特点对其进行了改进,最终利用Java Web技术对基于改进YOLO的井下行人检测系统进行了简单实现。本文以神经网络为基础,首先对卷积神经网络、深度学习网络等的理论做了介绍与分析,在深度卷积神经网络的基础上对YOLO目标检测系统的网络结构以及检测过程等原理进行了详细的剖析,分析了 YOLO系统精确度不高的缺陷,针对矿井下的视频质量差、背景单调、检测目标单一等特点对原有的YOLO系统在数据集和网络结构上进行了改进。利用煤矿井下的监控视频重新制作了训练集,网络结构上利用浅层的表征信息与深层的语义信息相结合的思想将网络中第八层的特征提取出来与最后层的输出相加作为整个网络最后的输出,在提取第八层提取特征的基础上提.提出了三种方案,分别为先卷积后采样、先采样后卷积、最后层输出利用反卷积扩大特征图再与第八层相加。通过在Caffe框架上进行实验并分析结果,综合考虑后选择了第二种方案为最终改进方案,证明.了改进后的YOLO系统在井下特殊环境的行人检测性能得到了提升。最后,利用Java EE技术构建了关于Java Web的井下行人检测系统,该系统包含系统管理、权限管理、检测管理、考勤信息、设备管理五个模块,对DCNN的井下行人检测系统进行了测试分析及功能性验证,说明了所设计系统的可行性。通过本文的实验可以看出,改进后的YOLO系统对井下特殊环境的检测有比较好的检测效果。
[Abstract]:Coal occupies a pivotal position in the utilization of energy in China. Coal mine safety, especially the safety of the underground production environment, has always been the top priority of the coal mining industry. At present, the inspection of underground workers by coal mining enterprises mainly depends on the positioning system of the underground personnel that has been equipped. The application of these technologies can effectively locate and identify the personnel, but in the process of use, there are replacement, cards, etc., their accuracy is not high, and the level of intelligence is low, especially when the monitoring personnel are negligent. Based on this background, this paper combines the application of DCNN (depth convolution neural network) in the field of video image recognition and the industrial video surveillance system of underground equipment. This paper presents a kind of underground pedestrian detection technology based on DCNN. In order to improve the detection speed, the YOLO target detection system is adopted, and it is improved according to the characteristics of the special underground environment. Finally, using Java Web technology, a simple realization of underground pedestrian detection system based on improved YOLO is carried out. Firstly, the theory of convolution neural network and depth learning network is introduced and analyzed based on neural network. On the basis of deep convolution neural network, the network structure and detection process of YOLO target detection system are analyzed in detail, and the defects of low accuracy of YOLO system are analyzed. The video quality under mine is poor and the background is monotonous. The original YOLO system has been improved in data set and network structure with the characteristics of single detection target, and the training set has been remade by using the monitoring video of underground coal mine. In the network structure, the feature of the eighth layer in the network is extracted and the output of the last layer is added as the final output of the whole network by the idea of combining the shallow representation information with the deep semantic information. On the basis of extracting features from the eighth layer, three schemes are proposed, which are first convolution and then sampling, first sampling and then convolution. The final layer output uses deconvolution expanded feature map to add to the eighth layer. Through the experiment on the Caffe framework and the analysis of the results, the second scheme is selected as the final improvement scheme. It is proved that the improved YOLO system has improved the performance of pedestrian detection in the special underground environment. Finally, the underground pedestrian detection system about Java Web is constructed by using Java EE technology. The system includes system management, authority management, detection management, etc. Five modules of attendance information and equipment management are used to test and analyze the underground pedestrian detection system of DCNN and verify the function of the system. The feasibility of the designed system is demonstrated by the experiment in this paper. The improved YOLO system has a good effect on the detection of underground special environment.
【学位授予单位】：西安科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TD76;TP391.41

【参考文献】