다중 프레임 레이더 이미지를 활용한 해상 동적 객체 검출 성능 향상

Abstract

Recent deep learning advancements have outperformed traditional methods in tasks like image classification, object detection, and semantic segmentation. In marine settings, various sensors detect environmental features and obstacles: cameras and LiDAR are commonly used for short-range mapping, while RADAR is preferred for long-range detection due to its resilience across large areas and in adverse weather. However, most research and available datasets focus on camera and LiDAR data, with limited data and studies on maritime RADAR. Additionally, RADAR images typically contain high noise and clutter, complicating the development of training datasets. While RADAR is valuable for detecting marine objects, its noise levels demand specialized approaches beyond conventional vision-based detection. Previous work applies semantic segmentation to RADAR images to distinguish objects, noise, and land, but uses only single-frame data. This paper proposes a novel method using sequential RADAR images to enhance object detection in marine environments. By leveraging temporal information and minimizing preprocessing, our approach significantly improves accuracy.