AodeMar: Attention-Aware Occlusion Detection of Vessels for Maritime Autonomous Surface Ships
Ning Wang, Yuanyuan Wang, Yuan Feng, Yi Wei
Abstract
For maritime autonomous surface ships (MASS), challenges exist in visual detection of occluded marine vessels since invisible occlusion is required to be inferred from locally unoccluded features which are weakly related to the entirety. In this paper, an attention-aware occlusion detection scheme of marine vessels, termed AodeMar, is originated from the viewpoint of MASS transportation. To this end, a position enhancement module is created by virtue of residual connections and coordinate attentions such that high-level semantics and spatial feature dependencies can be efficiently exploited, respectively, thereby accurately locating bounding boxes. Moreover, a multi-scale feature semantics correlation block is devised by combining spatial pyramid pooling and swin transformer-based self-attention encoder in order that the classification ability can be fertilized in both global and local sense. Experiments and comparisons show that the proposed AodeMar outperforms typical approaches including Faster R-CNN, SSD and YOLO series in terms of detection accuracy and robustness.