Author(s)

Jian Liu, Ziyi Meng, and Di Bai

Corresponding Author

Jian Liu

Abstract

Semantic learning on 3D point clouds model using a deep network has received great interests to address the 3D object classification problem. The complete information of the 3D object is described by the disordered data structure. A new classification method based on deep learning is proposed, which combines 3D vision with deep network. The point clouds and 3D standard geometry are regarded as input in the process of training the network. The proposed network unifies feature extraction and classification into a stage, which removed the need of manual feature engineering for 3D point clouds in traditional method. Therefore, the 3D point clouds data of the identified object is directly input to the trained network and the classification result can be obtained. The specific classification experiments on target objects have been done. The training has been completed on the RW-4028GR-TRT2 experimental platform with six Nvidia GTX 1080Ti. The experimental results show the accuracy of the proposed classification network, which is significant to improve the efficiency of object classification. Furthermore, the network has been applied to the recognition of indoor scene in the intelligent building, and it will attribute to the wide application in many fields in the future.

Keywords

Deep learning, Point clouds, Classification network