Research On LiDAR Data Classification Algorithms Based On DCNN

Light Detection and Ranging(LiDAR)technology is an active remote sensing measurement technology that can acquire ground object information by emitting laser light to a target.It plays an important role in the fields of terrain mapping and urban construction.In recent years,the role of traditional Convolutional Neural Network(CNN)for LiDAR data has been verified.However,with the layers of network in traditional CNN deepening,some problems appear.Such as gradient disappearance and layers redundancy.Deep Convolutional Neural Network(DCNN)can alleviate the above problems,but it also has problems with large model sizes and too many parameters.Therefore,this article will focus on the role of DCNN for LiDAR classification and improve it to obtain better classification results.The main contents of this article are as follow:First of all,the source,development,and characteristics of LiDAR data were explored and the experimental data sets were collected.At the same time,some classic LiDAR data classification algorithms were verified and reliable classification evaluation indicators were selected.Provide a theoretical basis for the design,implementation and performance analysis for the later algorithms.Additionally,a LiDAR classification algorithm STN-Dense Net combined Dense Convolutional Neural Network(Dense Net)with Spatial Transform Network(STN)is designed and implemented.STN deforms the input feature maps adaptively,according to the requirements of network by rotating,translation and scaling transformation.At the same time,the dense link of Dense Net can enhance the reuse of information on feature maps and alleviate the gradient disappearance and layer redundancy of DCNN.Compared with some classic algorithms and traditional CNN,STN-Dense Net has better classification results.Finally,LiDAR classification algorithm combined one of the lightweight networks-Squeeze Net with Octave Convolution(Oct Conv)is designed and implemented.Squeeze Net has only 1×1 and 3×3 convolution kernels in Squeeze Net which can reduce parameters of network.Oct Conv pools some input feature maps proportionally and stores them separately from the input feature maps that maintain the original size.Spatial redundancy can be reduced by sharing information between the two groups.Compared to some classic DCNN,this proposed algorithm reduces the number of network parameters and model size,thereby saving memory space and reducing training time.