자율주행을 위한 포인트 클라우드 3D 객체 인식에 관한 연구 = Study on Point Cloud Based 3D Object Detection for Autonomous Driving

다국어 초록 (Multilingual Abstract)

For the commercialization of autonomous vehicles, precise perception based on three-dimensional (3D) spatial recognition is imperative. While cameras offer valuable insights, their perception capabilities are inherently limited for comprehensive 3D spatial awareness. Therefore, the integration of LIDAR-based spatial recognition technology is indispensable. This study delved into methods for augmenting point cloud data to maximize the accuracy of LIDAR-based 3D Object Detection. Through this point cloud augmentation approach, techniques such as Jitter, Uniform Sampling, Random Sampling, Scaling, and Translation were employed and analyzed for their impact on detection accuracy. Furthermore, we explored optimal combinations of these techniques to amplify the precision of 3D Object Detection. Experimental outcomes, benchmarked against the KITTI dataset, showcased an improvement in the average precision (AP) by approximately 0.5-0.8. In addition, it was discerned that adopting distinct augmentation techniques, in particular Jitter, for different classes yielded enhanced results.