Radar-gestuurde Monoculaire Diepteschatting en 3D-objectdetectie voor Autonoom Rijden

With the growth of sensing devices such as RGB cameras and LiDAR sensors, 3D image processing on multi-modality sensing data serves as an essential foundation for autonomous driving. These sensing devices and corresponding computer vision algorithms are necessary for a self-driving car to observe the surrounding environment and act accordingly. By virtue of the recent development of neural networks, sensing technologies, such as denoising, compression, localization, 3D object detection, 3D reconstruction, have improved significantly these years. However, sensing data from multiple modalities has its limitations, which will restrict the performance of neural network models. In this thesis, we aim to study 3D image processing and develop state-of-the-art algorithms by exploiting and fusing the semantic and geometry information in different modalities for autonomous driving.