드론 연직 및 경사 영상을 이용한 YOLO 차량 탐지 연구

초록

Recent advances in drone and AI technologies have enabled object detection in transportation, including traffic analysis and parking enforcement. As of the end of June 2024, the cumulative number of registered vehicles in South Korea is 26,134,000, which means there is approximately one personal vehicle for every 1.96 people. This has led to severe issues such as traffic congestion, traffic accidents, illegal parking, and problems with emergency vehicle access. Various studies have been conducted using deep learning–based object detection algorithms and drone imagery, including vehicle detection for parking enforcement and vehicle classification by type. However, most existing studies utilize only one imaging angle, either nadir or oblique, while research employing both perspectives remains limited. Therefore, this study proposes and compares two scenarios for detecting vehicle objects in drone nadir and oblique view images using the YOLOv3 algorithm. Scenario 1 involves detecting vehicles in oblique view images using nadir view images, and Scenario 2 involves detecting vehicles in nadir view images using oblique images. To examine the effects of different oblique view angles, detailed scenarios were analyzed using 60 and 50 degrees. The AP (Average Precision) values for Scenario 1 were 0.36 and 0.32, while Scenario 2 yielded higher AP values of 0.76 and 0.69. The research confirmed that detecting vehicles in oblique view images using nadir view images has limitations, but detecting vehicle objects in nadir view images using oblique view images can be done with high accuracy

키워드