Derin öğrenme algoritmaları ile 3 boyutlu veri işleme ve donanım gerçeklemesi

Over the past decade, the rapid development of artificial intelligence technologies has led to deep neural networks becoming an integral part of many data processing tasks across various domains. DNNs have transformed the data processing field by automatically learning representations and features directly from the data, outperforming traditional methods, which are based on mathematical modeling with hand-crafted features. One of the areas that have seen the most significant advancements in computer vision since vision data is complex and hard to process with classical hand-crafted methods. In addition to 2D image processing, deep neural networks have proven effective for 3D data processing, which is critical for applications such as autonomous driving, robotics, and virtual reality, which rely on accurate 3D perception to understand spatial relationships and object detection in dynamic environments. In this work, we studied deep neural networks for 3D data processing, especially for point cloud data generated by LiDAR sensors. We also addressed the challenges of point cloud processing, such as irregular structure and high dimensionality. To overcome these issues, we explored 3D scene understanding through 2D image data, specifically tackling the depth estimation and semantic segmentation problems. We also explored fusion methods to effectively leverage the information coming from different modalities. Furthermore, recognizing edge devices' power and latency constraints, we investigated effective hardware implementation and acceleration techniques for designed neural networks.