LiDAR Point Cloud Image Interpolation via Separable Convolution

In recent years, the point cloud data generated by LiDAR has played an important role in 3D object detection, point cloud registration, 3D map splicing and other fields. However, the frame rate of a typical LiDAR sensor is limited by hardware performance so that it can't solve the problems of some scenes that require high frame rate, such as object tracking. In this paper, we propose a method of LiDAR point cloud image frame interpolation, which can be used to solve the problem of low frame rate of LiDAR. Given two consecutive point cloud images, LiDAR point cloud image interpolation aims to generate intermediate frame between them. Firstly, the point clouds are projected into 2D images, then the kernels of the Convolutional Neural Network output and the two consecutive point cloud images are used to realize the interpolation process of the intermediate frame. In order to improve the interpolation effect, we take into account the feature distribution of LiDAR point cloud images varies with space, the attention mechanism is introduced in the network model to effectively extract the features of the LiDAR point cloud images, At the same time, set the separable convolution kernels of the network output to a rectangle to meet the large aspect ratio of the LiDAR point cloud images. Both quantitative and qualitative experiments on the KITTI Dataset show that our method performs better than other mainstream methods.