面向户外增强现实的地理实体目标识别与检测

发布时间：2018-03-12 22:54

本文选题：户外增强现实　切入点：深度学习　出处：《武汉大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着移动互联网的快速发展,户外增强现实系统有越来越多的使用场景,如社交、购物、娱乐等。增强现实系统的核心在于跟踪注册算法的设计,主流算法为混合跟踪注册思想,在混合注册算法中基于视觉识别的跟踪注册最为关键,传统算法以提取特征点为核心,但由于计算量大且精度不高等问题影响着跟踪注册的效率。随着在计算机视觉领域中的深度学习思想的快速发展,以卷积神经网络为核心的图像识别模型也得到普遍应用,并且实验证明该算法比传统算法大大提高了图像识别精度和效率。本文研究了多种用于图像分类和目标检测的卷积神经网络模型,分析了各个模型的优缺点,并结合实际的应用在优秀的卷积神经网络SSD(单次检测模型)模型的基础上,更换了三层前置网络结构,较少了网络的深度,设计了不同长宽比的默认框,从而简化了模型用于识别户外的地理目标。参考ImageNet和PASCAL VOC 2007的训练集,采集数据并设计了训练和预测的数据集,在深度学习平台MXNET上训练简化的SSD模型,最终得到SSD模型的网络结构和权重值,该模型的mAP达到58.2%,PC端处理一张图像的时间为0.03s。本文将深度学习思想用于解决移动增强现实系统中的跟踪注册问题,采用基于传感器的非视觉感知和基于深度学习的视觉识别混合跟踪注册方式开发了移动户外增强现实原型系统,该系统是采用本地客户端的模式,利用安卓手机的传感器系统来获取手机姿态和当前位置信息,将训练好的简化SSD模型作为图像识别的引擎,使用GPU来完成计算过程,另外,针对地理目标设计了本地的二维和三维增强信息,提供了简单的交互模式,最终在实际测试中的处理图像的速度为1.5s,目标识别的置信度在90%左右。
[Abstract]:With the rapid development of mobile Internet, outdoor augmented reality systems have more and more use scenarios, such as social, shopping, entertainment, etc. The core of augmented reality system is to track the design of registration algorithm. The mainstream algorithm is hybrid tracking registration. In the hybrid registration algorithm, visual recognition based tracking registration is the key. The traditional algorithm takes extracting feature points as the core. However, the efficiency of tracking registration is affected by many problems, such as large computation and low precision. With the rapid development of depth learning in the field of computer vision, the image recognition model with convolution neural network as the core is also widely used. Experiments show that the algorithm improves the accuracy and efficiency of image recognition greatly. In this paper, several convolution neural network models for image classification and target detection are studied, and the advantages and disadvantages of each model are analyzed. Combined with practical application, based on the excellent model of convolution neural network (SSD), the three-layer network structure is replaced, the depth of the network is reduced, and the default frame with different aspect ratio is designed. Therefore, the model is used to identify outdoor geographical targets. Referring to the training set of ImageNet and PASCAL VOC 2007, the training and prediction data sets are collected, and the simplified SSD model is trained on the MXNET platform. Finally, the network structure and weight of SSD model are obtained. The mAP of the model reaches 58.2 and the time of processing an image is 0.03s. In this paper, the idea of depth learning is used to solve the problem of tracking registration in mobile augmented reality system. A mobile outdoor augmented reality prototype system is developed by using non-visual perception based on sensor and visual recognition and registration based on deep learning. The system adopts the mode of local client. Using the sensor system of Android phone to obtain the mobile phone attitude and current position information, using the trained simplified SSD model as the image recognition engine, using GPU to complete the calculation process, in addition, The local 2D and 3D enhancement information is designed for geographical targets, and a simple interactive mode is provided. Finally, the speed of image processing in the actual test is 1.5 seconds, and the confidence of target recognition is about 90%.
【学位授予单位】：武汉大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TP391.41

【相似文献】