Semantic Segmentation in Marine Environment: Using 2D spherical projection and convolutional neural networks

Abstract

The marine environment is constantly changing and therefore it can be difficult to control a vessel under these conditions. In this thesis, a method is proposed to interpret the surroundings to aid easy and safe travels on water. This is done through semantic segmentation by transforming 3D point clouds to 2D images, using a projection-based method. The transformation enables training with convolutional neural networks to achieve a fast and high performance network. The above method is successfully implemented in the marine environment and the results show that fewer classes are preferable to reach a high accuracy of the model. The features from the environment was unbalanced, which was compensated for by implementing a loss function that weighted the underrepresented classes higher. The model increased in performance for the minority classes. Furthermore, the real-time semantic segmentation was slower compared to the sensors update-time but there are possibilities to reduce the prediction time in future work. Precipitation was hard to detect due to low amounts of annotated data but the other surroundings could be detected in harsh weather conditions either way. The results show promising outcome for future implementation.