도심 자율주행을 위한 국내 주행 환경이 반영된 3D 라이다 데이터 셋 구축 및 검증 프로세스 설계
Design of Validating Process and 3D LiDAR Dataset Reflected Domestic Urban Driving Environments for Autonomous Driving
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
In complex urban driving environments, unlike highways and roads exclusive to automobiles, various dynamic objects such as motorcycles, bicycles, buses, trucks, and pedestrians coexist on the roads. Securing a high-quality dataset for such urban environments is crucial for autonomous driving. Therefore, we designs a process for collecting and annotating a high-quality dataset tailored to the domestic urban environment using 3D LiDAR. Subsequently, we construct our dataset and validate it using the Pillar-based CenterPoint model. To achieve this, we conducted training and evaluation under the same conditions as comparison experiments with public datasets. Our dataset, like the H3D and A3D (Korean) datasets, can be evaluated across 7 classes, and through experiments, we demonstrate performance comparable to existing public datasets. We emphasize the importance of designing processes for data collection, filtering, and annotation to construct a dataset reflecting the domestic driving environment, as well as the importance of designing evaluation processes to validate this dataset. Currently, the dataset construction process is ongoing, and it will be made publicly available upon completion.
Abstract
In complex urban driving environments, unlike highways and roads exclusive to automobiles, various dynamic objects such as motorcycles, bicycles, buses, trucks, and pedestrians coexist on the roads. Securing a high-quality dataset for such urban environments is crucial for autonomous driving. Therefore, we designs a process for collecting and annotating a high-quality dataset tailored to the domestic urban environment using 3D LiDAR. Subsequently, we construct our dataset and validate it using the Pillar-based CenterPoint model. To achieve this, we conducted training and evaluation under the same conditions as comparison experiments with public datasets. Our dataset, like the H3D and A3D (Korean) datasets, can be evaluated across 7 classes, and through experiments, we demonstrate performance comparable to existing public datasets. We emphasize the importance of designing processes for data collection, filtering, and annotation to construct a dataset reflecting the domestic driving environment, as well as the importance of designing evaluation processes to validate this dataset. Currently, the dataset construction process is ongoing, and it will be made publicly available upon completion.