Aero auto pilot segmentation network(AAPSN): a segmentation technology of point cloud data for UAVs autonomous driving

Abstract With the continuous development of Unmanned Aerial Vehicles(UAVs) technology, UAVs are gradually applied to the collection of remote sensing information. In the process of collection, researchers generally choose to use laser point cloud data to complete the automatic driving of UAVs. However, the accuracy of the current laser point cloud semantic segmentation technology is not ideal, which is mainly attributed to the following two aspects: (1) The point cloud of the UAVs perspective will lose a lot of spatial features during the training process; (2) The point cloud of objects with different scales under the UAVs perspective has different density and distribution. To solve this problem, this paper proposes a semantic Segmentation technology for UAVs perspective laser point cloud data, namely Aero Auto Pilot Segmentation Network (AAPSN). The technology is improved on the basis of polar coordinate segmentation network, and by adding a multi-scale feature extraction module, the point cloud data features of objects of different scales can be extracted. In addition, attention mechanisms are introduced in the pooling process to focus more on the spatial structure characteristics of cloud data, thus improving the accuracy and efficiency of segmentation. The network model proposed in this paper improves the mean Intersection over Union (mIoU) index by 3.3% and 19.2% respectively on two real urban LiDAR datasets.