Study On Object Recognition Method Based On LiDAR Point Cloud Image

Lidar can get some point cloud images under some harsh conditions such as night and fog.These data are sparse and incomplete.It is necessary to rely on very specialized knowledge and personal experience to extract features manually.Convolution neural networks can automatically extract features and classify them and they are invariant to displacement,scaling and other forms of rigid body changes.Some experts and scholars have used the convolution neural network to classify the point cloud images.Among them,the Vox Net network with the highest recognition rate loses a part of the data because of the limited number of grids.It is against this background that the method of point cloud recognition based on convolution neural network is studied in this thesis.In this thesis,we first introduce the development of convolution neural network and analyze the shortcomings of VoxNet network based on convolution neural network.After that,the three-dimensional CAD model ModeNet is used as the object model to be identified.Three-dimensional point cloud data is generated by the program simulation of the laser radar scanning process,and the data is preprocessed to obtain the data required for the experiment.Then the point cloud recognition method based on the convolution neural network is improved,and the neural network is built,and the appropriate network super parameters are obtained by the design and analysis of the architecture.Finally,the trained network is applied to point cloud classification and recognition and the experimental results are compared with the VoxNet network.The experimental results show that when using the same dataset for classification and recognition,the method of this thesis has a certain degree of improvement in recognition accuracy than the VoxNet method.