Channel Attention based Network for LiDAR Super-resolution

LiDAR is indispensable in the field of autonomous driving. LiDARs with more scan-lines captures more environ-mental information and provides denser cloud points, which is beneficial for tasks like structure from motion (SFM), SLAM and autonomous driving. However, the cost for a high-resolution LiDAR is unignorable. Applying super-resolution (SR) on point cloud obtained from low-resolution LiDAR to get denser data is a promising solution. Based on the previous use of neural network SR methods, we proposed a channel attention network for LiDAR point clouds SR. Specifically ,the proposed network can restore a higher precision Hi-res point cloud via range image SR. Additionally, for the problem of edge reconstruction after projection into a range image, we proposed to use circle padding to achieve a good result. Furthermore, we conduct the experiments using dataset collected in the simulation environment and compare out methods with the traditional approaches and network methods. Results show that our method gets the best effect on the proposed evaluation metric.