Cross-View Geolocalization Based On Deep Learning

Gross-view geolocalization refers to localizing a real-time query image by matching it to an accurate geo-tagged reference database stored in memory,which is a key technology to real-ize autonomous positioning of unmanned vehicles under navigation satellite system unavailable conditions.This project focuses on a common cross-view positioning task where the query im-age is a panoramic street-view image and the reference navigation map are aerial images.This task is very challenging due to the huge perspective differences in visual appearances and ge-ometric configurations between two views.Current works mainly focus on designing Siamese two-branch convolutional neural networks and related loss functions to learn cross-view feature embeddings.However,such representations easily get overfitted on discriminative scenes(i.e.,roads,grass),thus increasing the risk of representation learning for unseen data.In response to the above challenges,this paper conducts an in-depth study on the cross-view geolocalization task,and proposes a novel and efficient cross-view geolocalization structure with the help of deep learning technology.Main contributions of this article are as follows:1.Propose a novel deep learning framework exploiting part-level feature representations and apply it to cross-view geolocalization task.The framework is based on the Siamese network,firstly introduces the concept of partial-level feature representation to cross-view geolocaliza-tion,then combines part-level losses with global retrieval loss to enforces the deep network to learn representations for different parts and gain the discriminative power on unseen scene.Ex-periments on benchmark CVUSA dataset demonstrate that our proposed network outperforms the state-of-the art approaches without increasing model complexity,achieving rank-1 recall rate of 93.22% for aligned two view modalities.2.Research a novel cross-view geolocalization problem with unaligned orientations,and propose two solutions: first one is unaligned cross-view geolocalization based on global fea-tures,this method is improved on the basis of cross-view geolocalization model exploiting part-level feature representations,it combines the global feature representation which is generated by several part-level features and then output through the fully connected layer,with another global feature representation based on the attention mechanism to supervise and train the model.Second one is unaligned cross-view geolocalization based on correlation layer,this method es-timates the correlation coefficient of two views by correlation layer and regression network,then designs a triple loss function based on the correlation coefficients to optimize the network.In addition,an unaligned cross-view dataset is generated for this task by data simulation.Ex-periments show that the two solutions proposed in this paper significantly outperform current state-of-the-art methods for unaligned two view modalities.3.Considering the limitation of hardware resources on the mobile platform,a cross-view geolocalization model based on lightweight neural network is designed.Taking the model of cross-view geolocalization exploiting part-level feature representation as an example,this lightweight network simplies the backbone network by deep separable convolution and the fea-ture encoding module from the definition of the attention mechanism,reducing the amount of parameters and calculations.In the case of limited hardware resources,this method can save storage space and improve matching efficiency.