Subject

Depth estimation is a fundamental task for scene understanding in autonomous driving and robotics navigation. While learning-based supervised approaches for monocular depth have achieved good performance in outdoor scenarios [1], the use LiDAR sensors provide the most accurate depth information. However, the depth maps generated by these devices are sparsely distributed compared to the ones obtained from RGB images. This sparsity significantly impacts the performance of LiDAR-based applications. Image-guided methods for predicting dense depth maps from sparse LiDAR data using lidar-camera fusion showed significant improvement in results compared to the conventional depth-only techniques [2, 3].
Recently, due to the common use of radar sensors in the automotive and robotics industry, researchers attempted to address the problem using radar and camera fusion [4]. Still, using radars for depth completion is not yet thoroughly explored due to the increased sparsity and noises in the measurements. However, radars are proven to be more robust in harsh environments and generally, they are more cost-effective and used in the automotive/robotics industry.

Kind of work

In this study, we will investigate existing radar-camera completion approaches. In addition, we will investigate a self-learning fusion approach for scene reconstruction in adverse surrounding conditions.