Knowledge-Driven Semantic Segmentation for Waterway Scene Perception

Semantic segmentation as one of the most popular scene perception techniques has been studied for autonomous vehicles. However, deep learning-based solutions rely on the volume and quality of data, and knowledge from specific scene might not be incorporated. A novel knowledge-driven semantic segmentation method is proposed for waterway scene perception. Based on the knowledge that water is irregular and dynamically changing, a life time-of feature (LToF) detector is designed to distinguish water region from surrounding scene. Using a Bayesian framework, the detector as the likelihood function is combined with U-Net-based semantic segmentation to achieve an optimized solution. Finally, two public datasets and typical semantic segmentation networks, FlowNet, DeepLab, and DVSNet are selected to evaluate the proposed method. Also, the sensitivity of these methods and ours to the dataset is discussed.