Bayesian deep learning based semantic segmentation for unmanned surface vehicles in uncertain marine environments

Abstract

Unmanned Surface Vehicles (USVs) face challenges in complex marine environments due to the diversity and unpredictability of obstacles. These challenges are exacerbated by the limited availability of marine semantic segmentation datasets. To address these issues, this work aims to investigate the potential of a Bayesian deep learning-based semantic segmentation approach to improve obstacle recognition and uncertainty estimation. Specifically, Bayesian SegNet is employed to better handle uncertainties arising from environmental changes and model parameters. By estimating uncertainties, USVs can navigate and avoid obstacles more effectively, even in novel environments. Given the scarcity of USV-specific datasets, a stepwise learning strategy is implemented, where training is conducted on the MaSTr1325 dataset, and testing is performed using both the MaSTr1325 and OASIs datasets. This strategy enhances the model’s ability to generalise to real-world scenarios. Experimental results demonstrate that Bayesian SegNet significantly outperforms non-Bayesian models, achieving a 1.3 % increase in precision and a 6.5 % improvement in $F1$ score in marine environments. Additionally, Bayesian SegNet exhibits superior uncertainty estimation and generalisation capabilities, with a notable 39.77 % higher $F1$ score on the OASIs dataset compared to traditional SegNet, highlighting its effectiveness in improving semantic segmentation accuracy in USV navigation tasks.