基于空地交叉视图匹配的无人系统视觉定位

Visual Localization of Unmanned Systems Based on Ground-to-Aerial Image Matching

空地交叉视图定位是一项极具挑战性的任务,因为2个视图在视觉外观和空间布局上有巨大差异。目前大部分方法将该问题视作图像检索任务,并设计了各种双分支卷积神经网络来学习这2个视图的全局特征嵌入。现有方法未充分利用部分级特征,增加了对不可见数据表示学习的风险。为了应对上述问题,提出了一种新的深度学习算法,首次将部分级特征这一概念引入交叉视图定位领域。该算法通过对全局特征进行切分直接生成部分级特征,并引入部分级监督信息来引导模型学习更具区分度的特征。试验结果表明,该算法在CVUSA数据集上r@1召回率可达93.22%,在不增加参数和计算复杂度的前提下有效提高了匹配准确度。

Ground-to-aerial cross-view geolocalization is a very challenging task, due to the huge perspective differences on visual appearance and spatial arrangement between two views. The problem is regarded as a standard image retrieval task in most of current approaches, and various dual branch convolution neural networks are designed to learn the global feature embedding for the views. However, part-level feature isn′t utilized in current methods, thus the risk on learning invisible data representations is rising. For the above problems, a new novel deep learning method is proposed, and the concept on part-level feature is firstly introduced into cross-view geolocalization domain. Part-level feature representations are automatically generated by segmenting the global feature, and part-level surveillance information is introduced to lead the model learning more distinguished features. Experimental result shows that the method can achieve rank-1 recall rate of 93.22% on CVUSA dataset, and it can efficiently improve the matching accuracy without increasing model parameters and calculating com-plexity.