Stereo Object Recognition Based On Multi-source Remote Sensing Imagery

With the continuous development of multi-source,multi-angle,and multi-platform remote sensing technologies,people can not only obtain traditional types of remote sensing information including optical images and spectral images by using satellites and aviation platforms,but also obtain the stereo information about remote sensing targets through stereo images captured by satellites.In this situation three-dimensional target recognition are gradually developed.In recent years,full use of the multi-source remote sensing image information acquired by multi-platform,multi-sensor for stereo target recognition has been widely used in various fields.In the civil domain,target stereo recognition can provide convenience for urban planning,autopilot,and traffic navigation;in the military field,stereoscopic recognition of targets can also provide assistance for military reconnaissance,intelligence analysis,and target strike effect evaluation.Most of the traditional target recognition methods focus on feature extraction and recognition based on the target two-dimensional information,but often ignore the feature extraction and expression of three-dimensional information from the targets.These methods can not fully describe the overall characteristics of the target,so they result in a lower accuracy in three-dimensional target recognition.Therefore,based on the acquisition of target multi-source remote sensing image information,this paper mainly focuses on the threedimensional recognition of target from remote sensing,and deeply discuss the three aspects of target detection,three-dimensional feature extraction and three-dimensional target recognition involved in the three-dimensional target recognition process.First of all,for the problem that the mean shift partitioning algorithm can't separate the regions with high jump but spectral similarity,this paper proposes a highly-assisted mean shift partitioning algorithm.The algorithm uses a weighted graph as an edge weight while applying the mean-shift segmentation algorithm.So it can limit the segmentation in free regions with a high jump,make full use of DSM data and color orthoimage data,and detect the target from the image.It laid the foundation for feature extraction and stereo target recognition.Secondly,for the issues that caused by the low accuracy of stereoscopic recognition of remote sensing targets due to the inadequate description of the overall characteristics of the target,a target feature extraction method based on multi-source information is proposed.The method uses optical images,DSM data and RGB images to extract the geometric features and texture features of the target,the target three-dimensional structure features,and the target attribute spectral features,and takes advantage of the multi-source feature fusion method to fully describe the overall characteristics of the target.By complementing the features of the multi-source information,it can provide support for further improvement of the accuracy of the target stereo recognition.Finally,on the basis of acquiring the multi-source features of the target,in order to ensure the accuracy of the three-dimensional target recognition,this paper adopts the method of support vector machine to achieve the three-dimensional recognition of remote sensing targets.Through the combination of multi-source feature fusion method and support vector machine,the overall feature description of the target is achieved,and then the stereoscopic recognition of the remote sensing target is completed.Experimental results verify the superiority and robustness of multi-source feature fusion in stereoscopic recognition for remote sensing targets.