Research On The Surface Classification Method Of Unmanned Vehicles Based On Hyperspectral Images

The basic premise that an unmanned vehicle can complete autonomous navigation and various unmanned operation tasks is to correctly sense the surrounding complex environment.At present,the research and test environment of intelligent unmanned vehicles is mainly based on highways and urbanized roads,and rarely considers the distribution of vehicle driving force by unknown surface types,which causes the problem of speed matching and overturning of the vehicle body.However,there are certain differences between urbanized roads and outdoor roads.The ground structure in the two scenarios is different,and the underlying control of the vehicle should be different,which is more important for military ground unmanned vehicles.Based on this,this article focuses on the surface classification of unmanned vehicles on structured and unstructured roads.Considering the limitations of traditional RGB images in surface classification and the exploratory nature of using hyperspectral surface classification,this paper mainly conducts research on surface classification based on hyperspectral images.The research content mainly includes the following points:(1)Set up a vehicle-based hyperspectral data acquisition system,design a surface data acquisition scheme,and perform data acquisition in a real outdoor environment.The main types of surface include asphalt road,dirt road,cement road,and vegetation.The image is marked at the pixel level to establish a surface classification data set.(2)Research and analyze some key technologies in hyperspectral image classification,including hyperspectral image dimensionality reduction method,CNN-based hyperspectral image classification method,classification evaluation index,etc.The differences between band selection and feature extraction in the method of dimensionality reduction of hyperspectral images are analyzed in detail,and the rationality and necessity of band dimensionality reduction using the band selection method in the future are clarified.This paper focuses on the process of 1D-CNN,2D-CNN,and 3D-CNN for hyperspectral image classification,and clarifies that the spatial-spectrum joint feature extraction is usually more effective than the classification method just based on spectral feature extraction.It establishes a general direction for the main research work of this article.(3)An SNMF band selection method combining class-specific separability factors is proposed and implemented,which is used to reduce the dimensionality of the surface hyperspectral data.By specifying the range of the number of band clusters,and then repeatedly testing and evaluating with a certain step size,24 bands were finally selected for the research of surface classification.(4)A 2D-CNN framework that considers both pixel spatial information and spectral information is proposed and implemented for surface classification,and compared with2D-CNN based on spectral information and 2D-CNN based on RGB three-channel image.On the one hand,it is verified that the 2D-CNN combined with spatial spectrum information is better than the 2D-CNN classification using only spectral information.On the other hand,it also proves the advantages of hyperspectral images over traditional RGB images.(5)An end-to-end spatial spectrum residual 3D-CNN network framework is proposed and implemented for surface classification.This method can directly train the original three-dimensional hyperspectral data without performing band dimensionality reduction.The design idea of spectral residuals can solve the problem of network degradation caused by the increasing number of network layers and greatly improve the classification performance.The experimental results are compared and analyzed with the SPC,SPA,and 2D-CNN surface classification effects.The experimental results show that the spatial spectral residual 3D-CNN can achieve better classification performance.