Year: 2023
Type: conference-paper
Abstract: USV (Unmanned Surface Vehicle) have high requirements for object detection accuracy and real-time performance during sea patrol. YOLO, which combines accuracy and speed, is the preferred object detection algorithm in industrial applications. However, deploying on an embedded platform with limited energy and computing power is challenging. For this problem, this paper proposes a lightweight object detection algorithm suitable for deployment in USV. Firstly, from the perspective of network structure, the lightweight object detection algorithms widely used in the industry in recent years are sorted out. Then describe the MS-COCO, SeaShips dataset, and algorithm evaluation metrics; Secondly, by comparing the performance of various lightweight algorithms on MS-COCO, combined recognition accuracy, speed, and model size conclude that YOLOv5n is a more suitable lightweight object detection algorithm for deployment on embedded platforms. Finally, YOLOv5n was tested with the SeaShips dataset, and experimental results show that although the number of parameters is only 1.9M, the AP0.5 reaches 93.9%, and the floating-point computation is 4.2G FLOPs.
Authors: Xin Zhang, Chenfan Huo
Institution: Xi’an University of Posts and Telecommunications
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4399658695
