基于DeepLabv3+的鸣笛监测系统定位误差检测  

Positioning Error Detection of Vehicle Honking Systems Based on DeepLabv3+

作　　者：周云 郑慧峰 姚润广 朱勤丰 ZHOU Yun;ZHENG Huifeng;YAO Runguang;ZHU Qinfeng(College of Metrology&Measurement Engineering,China Jiliang University,Hangzhou Zhejiang 310018,China)

机构地区：[1]中国计量大学计量测试工程学院,浙江杭州310018

出　　处：《传感技术学报》2022年第8期1086-1092,共7页Chinese Journal of Sensors and Actuators

基　　金：浙江省自然科学基金(LQ20A040007、LQ19A040004);国家重点研发计划资助(2017YFF0205004);国家自然科学基金(11904346);浙江省质量技术监督系统科研计划项目(20180103);浙江省‘仪器科学与技术’重中之重学科人才培育项目(JL150506);浙江省教育厅科研资助项目(Y201942513);浙江省大学生科研创新活动计划资助项目(2021R409052)。

摘　　要：用传统方法检测鸣笛监测系统的定位误差容易受声相云图中复杂背景干扰且效率低下,提出了一种基于DeepLabv3+的定位误差检测方法。首先定义了横向定位误差和纵向定位误差,然后采用轻量级的主干网络并增加通道注意力机制,对DeepLabv3+网络结构进行改进,预测出鸣笛声源和伪彩图的中心像素坐标;其次根据现场安装信息建立坐标系转换模型,结合几何推导法获得对应的世界坐标,从而得到定位误差;最后开展了鸣笛监测系统的动静态定位误差室外实验。结果表明:改进DeepLabv3+网络的平均交并比为87.82%,较原始算法提升了3.66%,对鸣笛声源和伪彩图的分割效果更好,检测精度满足鸣笛监测系统的现场定位误差检测要求。Using traditional method to detect the positioning error of the monitoring vehicle honking monitoring system is inefficient and easily interfered by the complex background in the acoustic phase image.A positioning error detection method is proposed based on DeepLabv3+.The horizontal and vertical positioning errors are first defined,then the DeepLabv3+structure is improved by adopting a light weight network and increasing a channel attention mechanism,and the center pixel coordinates of the sound source and the pseudo-color image are predicted.Subsequently,the coordinate system conversion model is established according to the on-site installation information,and the corresponding world coordinates are obtained in combination with the geometric deduction method,therefore the positioning error is obtained.Finally,the dynamic and static positioning errors of the vehicle honking monitoring system are launched.The results show that the improved DeepLabv3+network has an average intersection over union(mIOU)ratio of 87.82%,which is 3.66%higher than the original algorithm,and a better segmentation effect on the sound source and the pseudo-color image,and the detection accuracy satisfies the on-site positioning error detection requirements.

关 键 词：误差检测 声源定位 鸣笛监测系统 深度学习 DeepLabv3+ 

分 类 号：TB95[一般工业技术—计量学]