Fast Segmentation Algorithm of USV Accessible Area Based on Attention Fast Deeplabv3
Liang Cheng, Rui Xiong, Jia‐Rong Wu, Xuemei Yan, Chunli Yang, Yongjie Zhang, Yunze He
Abstract
Accessible area identification is of great significance to the intelligent navigation of unmanned surface vessels (USVs). This study focuses on the segmentation of accessible areas and obstacles using an artificial intelligence system named “CS40P” operating on Nvidia’s Jetson AGX Xavier edge computing platform. This article analyzes the limitations of many mainstream semantic segmentation methods and proposes attention fast DeeplabV3 (AF-DeeplabV3). First, this work employs fast feature extraction (FFE) to save computational costs and reduce inference time. Then, multiscale attention atrous spatial pyramid pooling (MsAASPP) is introduced to gather feature information at various depths. Finally, subtraction spatial attention (SSA) within the attention feature pyramid (AFP) is used to enhance the effective use of edge feature information. We collected a substantial amount of raw data through maritime USV navigation experiments and, through the innovative use of human-machine collaborative annotation and mosaic data augmentation methods, constructed a passable area dataset containing 11 991 images. Notably, the proposed model achieves outstanding results, with the mean intersection over union (MIoU) reaching 89.70%, surpassing DeeplabV3+ by 4.51%. Furthermore, the inference speed on PC reaches 47.62 fps, which is 15.47 fps faster than DeeplabV3+, and the image inference speed reaches 24 fps on NVIDIA.