3D Object Recognition Of Laser Point Cloud Based On Convolutional Neural Network

With the continuous development of computer vision,machine perceives the natural world have evolved from traditional two-dimensional imaging to three-dimensional imaging.The 3D object recognition based point cloud of laser imaging technology has great application and development prospects in the fields of military investigation,unmanned driving and robotics.The 3D object recognition,tracking and detection based point cloud method has become an important research hotspot in these fields.In this paper,the laser imaging technology of 3D object recognition as the background,through the design and improvement of convolutional neural network structure,the adjustment and optimization of network structure parameters,the research on the field of 3D object recognition based point cloud is carried out.The main research work of this paper is as follows:1.First of all,the research background and significance of 3D object recognition based point cloud in laser imaging are expounded.The current research status,theories and development trend of 3D object recognition based laser point cloud and convolutional neural network are reviewed in detail.2.Three-dimensional target recognition is theoretically introduced in detail.In this paper,the traditional method of artificially designing three-dimensional features from the beginning is introduced.Due to the large increase in data,the traditional methods cannot continue to meet the needs of development,the method of deep learning becomes a mainstream solution in 3D object recognition.Finally,some pioneering works of 3D object recognition based point cloud are introduced.3.In this paper,the proposes three different improved network structure models based on the convolutional neural network model to better realize the 3D object recognition based point cloud by enriching the diversity of convolutional layers.Based on multi-resolution feature fusion convolutional neural network,by changing the distance between point clouds,four inputs are as input to the network and then these information are fused.Based on multi-level feature fusion convolutional neural networks,the network fuses information after extracting features between different levels.Based on multi-scale feature fusion convolutional neural network,the scale information of features is changed in two ways and there has four single-scale modules for information fusion.These three network models have proved their applicability and superiority through experiments.4.Subsequently,the experimental data is introduced in detail.And then,considering the influence of network parameters on network identification accuracy,the activation function,adaptive learning rate and other methods are adopted into the network.And also,Dropout algorithm is adopted to prevent over-fitting of the network.At last,the experimental are tested on the three improved network models for 3D object recognition task.5.Finally,the main work of the paper is summarized,and moreover,the last chapter analyzes some problems in the research and also points out the research direction of the next work.