Research on Deep Learning Based Optimal Combination of Multidimensional Features in Large-Scene Laser Point Clouds Classification

As the self-occlusion or occluded 3D point clouds objects in complex scenes, which could affect the accuracy of objects classification, we propose Optimal Combination of Multidimensional Features based on deep learning for large-scene laser point clouds in classification. We construct the optimal combination matrix of multidimensional features by extracting the three-dimensional features of the three-dimensional point cloud and the two-dimensional features in multiple directions. The multidimensional optimal combination features are introduced into the convolutional network. The experimental results show that effectiveness of classification for large-scale point clouds, the effectiveness of 3D feature of point cloud is higher than that of 2D feature. The classification accuracy of our method can reach 98.8% on the Large-Scene Point Cloud Oakland data set, which obtains the better classification accuracy than other classification algorithms the paper mentioned.