Research On Hyperspectral Image Of Intelligent Vehicle Obstacles Classification Method Based On 3D-CNN

The development and application of automobile intelligent technology is of great significance to improve automobile safety.Environmental perception is the prerequisite of automobile intelligent technology,and obstacle detection is the basis of environmental perception.At present,the research status of obstacle detection technology is summarized as follows: a single sensor technology has its own advantages and disadvantages,and it is difficult to meet the requirements of detecting all types of obstacles rapidly and precisely;multiple sensor fusion technologies are still in the intail stage,and new sensor technologies and new algorithms need to be tried and explored.In recent years,hyperspectral imaging technology has become a key research direction of machine vision,and convolutional neural network has great advantages over traditional algorithms and even other deep network models.Analyzing the characteristics of hyperspectral image data,and the advantages of three-dimensional convolutional neural network on sharing weights,fewer parameters,this paper proposes a research on hyperspectral image of intelligent vehicle obstacle classifications method based on 3D-CNN,with a high hope of the application of spectral imaging technology and convolutional neural network algorithm in intelligent vehicle obstacle recognition providing new ideas.It provides a new direction for the integration of new detection methods and new algorithms with traditional sensing technologies.Firstly,in order to obtain hyperspectral images of possible intelligent vehicle obstacles,this article builds a system for acquisiting hyperspectral image of intelligent vehicle obstacles,collects hyperspectral images of six types of substance,including stone,wood,paper,metal,glass,and plastic in the wavelength range of450～950nm,selects the hyperspectral data of wood,paper,metal,and plastic to establish the original hyperspectral image database of intelligent vehicle Obstacles after spectral image analysis,extracts the region of interest(ROI,region of interest),and forms the ROI data set S.Secondly,in order to reduce the dimension of high-dimensional hyperspectral data,on the premise of not destroying the original information of each band,taking the material separability as the band principle,a method of band selection based on convolutional neural network is proposed to reduce the dimension of hyperspectral data.In this method convolutional neural network is used to locate the effective image feature position information,which is applied to the band selection of hyperspectral images.The specific method is as follows: the spectral information of each sample in the original hyperspectral image database is performed to generate a new sample set S1,with two dimensional gray reconstruction,which is input to a lightweight network squeezenet model of the convolutional neural network.and the band activation map is generated during the classification process,and the band selection is performed according to the weight ranking of the band activation map.The weights are sorted for band selection.The 10,20,30,40,and 50 bands selected by this method are combined with the ROI data set S to form a feature space data set,and input to the convolutional neural network model such as 3D-Res Net10,3D-Res Net18 and3D-Res Net34 to perform classification training.The experimental results verify the effectiveness of the method and prove that the classification accuracy of 30 bands is the best.Finally,in the design of the classification model,a 3D-BN-VGG network structure for human behavior recognition is proposed to be applied to the hyperspectral image classification task.The classification accuracy and convergence speed have been improved to a certain extent.Improved on the basis of this model,a hyperspectral image classification model based on 3D-Res-VGG is proposed,and the residual module is used to deepen the network level,so as to better achieve deep feature extraction while avoiding over-fitting.Compared with experimental verification,the classification accuracy and sensitivity are improved,but the time is slightly wasted..