Széchenyi Plusz RRF

This paper presents a novel Lidar based, single-frame, depth data completion method for moving ego-vehicle with dynamically varying environment. The approach provides high resolution depth image from sparse point cloud input. The effectiveness is demonstrated through simulated data of three different types of low-end Lidar devices, showcasing the potential of entry-level, affordably priced instruments. A Convolutional Neural Network (CNN) is presented as the core, trained in conjunction with a discriminator model. Distance proposals are generated with pixel-wise regression at the last layer of the network, resulting in a spatially accurate complete depth image. To address this challenging objective, feature- and three-dimensional point cloud based losses are introduced to assist the training process. In experiments on our synthetic dataset, which includes simulations from three different Lidar instruments, we demonstrate that our solution outperforms two state-of-the-art baseline methods in pixel-wise comparisons to the Ground Truth, both considering RMSE with an average of 5.83 (baseline: 9.07 and 9.72) and MAE with 1.10 (baseline: 2.78 and 2.21) in meter considering the Livox Avia sensor. To showcase the model's efficacy, we present our solution's impact through a study on semantic segmentation. Using completed depth data instead of raw Lidar data as input for the segmentation can considerably improve the Intersection over Union (IoU) score from 0.72 (raw depth data as input) to 0.79 (completed depth data as input). Additionally, in an ablation study we examine the importance of the used features and losses.