基于改进VGG-19的井下视觉指纹匹配定位方法

Underground visual fingerprint matching embedded robot positioning method based on improved VGG-19

针对传统矿井电磁波测距定位存在依赖通信链路的问题,提出了一种基于改进的卷积神经网络VGG-19的井下定位方法.根据基于视觉图像进行识别的原理,将基于深度学习的计算机视觉技术应用于井下定位.不同间隔点位置图像的特征不尽相同,所以对井下的场景进行不同间隔的划分,在所取的间隔点处采集数据集,建立图像信息指纹库;将不同间隔位置点图像进行相应的标记分类,然后用迁移学习的方法将获得的数据集用改进的VGG-19网络进行训练,获得识别分类模型;运用识别分类模型可对不同位置实时图像进行识别,获得采集图像的设备位置数据,从而实现定位.所述定位方法得到的识别分类模型可在嵌入式系统设备运行,定位过程无须通信网络支持,极大简化了系统的复杂性,完全适应矿井灾后恶劣条件,可用于救援机器人等移动装置的井下定位,具有实时性、稳定性、抗干扰性,且有很好的定位准确率.

In order to solve the problem of dependent communication link in traditional mine electromagnetic wave measurement positioning,a downhole positioning method based on improved convolutional neural network VGG-19 was proposed.According to the principle of identification based on visual images,depth learning-based computer vision technology was applied to downhole positioning.The characteristics of different interval position images are different,so different intervals were divided into different intervals.The data set was collected at the athletic point,and the image information fingerprint library was established.The corresponding marker classification of different intervals positions was then used,and then the obtained data sets will be trained using the improved VGG-19 network to obtain the identification classification model.The identification classification model was used to identify different locations real-time images.The device location data of the capture image was obtained to achieve positioning.The identification classification model obtained by the positioning method can operate in the embedded system equipment,and the positioning process does not require communication network support,which greatly simplifies the complexity of the system,and fully adapt to the harsh conditions after the mine disaster.It can be applied to the rescue robot and other mobile devices.The downhole positioning,with real-time,stability,anti-interference,has a good positioning accuracy.