P-SANet:一种高精度实时点云语义分割框架

Perception in autonomous system is an important task to guide decision execu-tion. Lidar point cloud is a type of dataset to complete perception task, it is rich in original information, easy to collect, and convenient to store. Compared to camera image, point cloud contains precise spatial information and adapts to various en-vironments, nevertheless, more information means more computing power con-sumption. The processing speed and accuracy are two key metrics of neural net-work framework. The traditional methods have to pay the price of reducing accu-racy for increasing processing speed. Though some frameworks preprocess point cloud into projected image, the 2D image tensor also contains a large number of redundant channel features in the traditional 2Dconvolution operation. In this pa-per, we propose a point clouds semantic segmentation framework, we replace the standard convolutional layer with a new sub-module, and it greatly reduces the amount of computation, besides, we introduce a sub-module to fuse the coordi-nate values and middle tensors. The framework in this paper is divided into three parts: spherical projection preprocessing module, En-Decoder module and data post-processing module. We use the SemanticKITTI dataset to conduct experi-ments, and the results show that our framework outperforms other frameworks both in prediction accuracy and prediction speed. We also use sparse point cloud dataset to test the generalization of our framework, and the experiments show that it performs better than other frameworks. Code is available at: https://github.com/windtries/P-SANet