Altitude-Based Automatic Tiling Algorithm for Small Object Detection

Altitude-Based Automatic Tiling Algorithm for Small Object Detection

초록

In this paper, a method is proposed to perform consistent detection of small objects across all altitudes using UAVs. Unlike traditional tiling algorithms that detect small objects with a fixed number of tiles, our method dynamically adjusts the number of tiles based on the altitude and the size of the bounding box. This approach determines the proper number of tiles for detecting small objects at each altitude. Real-time object detection was conducted on the AI board mounted on the UAV, and for this purpose, the model was optimized using a lightweighting process. Using the Visdrone dataset, the performance of small object detection at each altitude was tested in real-world conditions. Compared to the conventional YOLOv5 and fixed-number tiling algorithms, our method consistently detected small objects at all altitudes.

키워드

Deep Learning; Object Detection; Tiling; Small Object; UAV
제목
Altitude-Based Automatic Tiling Algorithm for Small Object Detection
제목 (타언어)
Altitude-Based Automatic Tiling Algorithm for Small Object Detection
저자
최문규; 신수용
DOI
10.7840/kics.2024.49.3.424
발행일
2024-03
저널명
한국통신학회논문지
권
49
호
3
페이지
424 ~ 434