| Summary: | Autonomous driving is one of the most actively researched fields in artificial intelligence. The autonomous vehicles are expected to significantly reduce the road accidents and casualties one day when they become sufficiently mature transport option. Currently much effort is focused to prove the concept of autonomous vehicles that is based on a suit of sensors to observe their surroundings. In particular, camera and LiDAR are researched as an efficient combination of sensors for on-line object identification on the road. 2D object identification is an already established field in Computer Vision. The successful application of Deep Learning techniques has led to 2D vision with Human-level accuracy. However, for a matter of improved safety more advanced approaches suggest that the vehicle should not rely on a single class of sensors. LiDAR has been proposed as an additional sensor, particularly due to its 3D vision capability. 3D vision relies on LiDAR captured data to recognize objects in 3D. However, in contrast to the 2D object identifi- cation, 3D object detection is a relatively immature field and still has many challenges to overcome. In addition, LiDARs are expensive sensors, which makes the acquisition of data required for training 3D object recognition techniques expensive tasks as well. In this context, this Master's thesis has the major goal to further facilitate the 3D object identification for autonomous vehicles based on Deep Learning (DL). The specific contributions of the present work are the following. First, a comprehensive overview of the state of the art Deep Learning architectures for 3D object identification based on Point Clouds. The purpose of this overview is to understand how to better approach such a problem in the context of autonomous driving. Second, synthetic but realistic Lidar captured data was generated in the GTA V virtual environment. Tools were developed to convert the generated data into the KITTI dataset format, which has become standard in 3D object detection techniques for autonomous driving. Third, some of the overviewed 3D object identification DL architectures were evaluated with the generated data. Though their performance with the generated data was worse than with the original KITTI data, the models were still able to correctly process the synthetic data without being retrained. The future benefit of this work is that the models can be further trained with home-made data and varying testing scenarios. The implemented GTA V mod has proved to be capable of providing rich, well-structured and compatible datasets with the state of the art 3D object identification architectures. The developed tool is publicly available and we hope it will be useful in advancing 3D object identification for autonomous driving, as it removes the dependency from datasets provided by a third party.
|
|---|