A Novel Brain-Like Navigation Based on Dynamic Attention with Modified Unet

In cognitive navigation system, animals show an inborn ability of spatial representations and correct self-positioning errors at every fired cell. Inspired by navigation mechanism of animals, we propose a novel strategy to improve the navigation accuracy of brain-like navigation based on UAV. Firstly, we employs encoder-decoder structure based on Unet to solve semantic segmentation tasks. Unet are able to encoder detailed information of images by constantly pooling and upsampling operations with less training parameters, while it often ignores high-level spatial information. Hence, we propose “dynamic attention with modified Unet” structure, which learns high-level information maintaining less training parameters. Specifically, multi-scale atrous convolutions are adopted in dynamic modules between encoder and decoder to extract features at different resolution. Secondly, the pixels with maximum probability segmentation are extracted, and they will be mapped to satellite map to obtain actual position coordinate of UAV. Finally, positioning errors are corrected at each place cells in the brain-like navigation of UAV. Our results show that proposed segmentation model improve performance by 9.64% compared with conventional Unet, and the positioning accuracy is improved by 90.52%.