Development of deep learning based perception module for autonomous driving

Autonomous vehicles are rapidly evolving and are shaping the future vision of mobility with their potential to provide safe, comfortable, and sustainable transportation solutions. The fundamental software and hardware architecture of autonomous vehicles consists of an integrated structure including sensor hardware, data processing infrastructure, and decision-making algorithms. In this architecture, the integration of sensors such as LiDAR, cameras, and RADARs, combined with high-performance computing units, directly affects the accuracy of environmental perception. Among these components, the perception layer is one of the most critical, as it accurately and quickly detects objects in the environment, forming the basis of driving decisions. The performance of this layer is shaped by factors such as sensor resolution, the number of channels, data processing techniques, and the effectiveness of learning algorithms. However, deep learning-based 3D object detection models often suffer significant performance degradation in different scenarios due to domain shift between training and testing datasets. In this study, to address the domain shift problem, six modern domain adaptation methods (UDA, Bi3D, CLUE, ST3D, SESS, and Uni3D) were systematically evaluated and compared. In addition to publicly available datasets such as Waymo and nuScenes, a high-resolution real-world dataset collected using the KARSAN prototype autonomous minibus and a custom simulation dataset with various LiDAR configurations were utilized. Experimental results revealed that SESS and Uni3D significantly reduce performance losses caused by domain shift and achieve high overall accuracy. Furthermore, the effects of LiDAR resolution and channel count on detection performance were comprehensively analyzed, highlighting the contributions of sensor configurations to system performance. In conclusion, this thesis presents a comprehensive domain adaptation benchmark that contributes to the development of safe, generalizable autonomous driving perception systems. The findings presented in this study are expected to provide significant benefits for both academic research and industrial applications.