Ship visual trajectory exploitation via an ensemble instance segmentation framework

Maritime traffic safety is of utmost importance in the shipping industry. It is common to avoid potential traffic accident via the support of varied kinematic maritime data (e.g., ship position, ship speed). The study aims to fulfill maritime traffic safety enhancement by extracting multiple ship trajectories from maritime video data. The proposed framework is implemented with three steps. Firstly, the framework obtains initial ship imaging positions by proposing an ensemble YOLOX (You Only Look Once X) with CenterNet model. Secondly, ship contours from each maritime frame are explored with the deep snake module. Thirdly, consecutive ship imaging trajectory is identified with an enhanced Bytetrack multi-ship target tracking algorithm. Experimental results demonstrate that our model outperforms other ship trajectory extraction-like models considering that the identification F-Score (IDF1), IDP (identification Precision), IDR (identification recall), and MOTA (multi-object tracking accuracy) for our proposed framework were 0.965, 0.995, 0.933, and 0.925, respectively. The research findings can help ship crew better aware on-site traffic situations, and thus make optimal ship maneuvering operations for the purpose of enhancing maritime traffic system safety and sustainability. •we propose a novel trajectory fusion framework by integrating YOLOX and CenterNet modules.•we simultaneously track ships from maritime images in a pixel-wise manner with instance segmentation mechanism.•we propose an improved Bytetrack model with the help of irregular polygon detector (i.e., mask of the instance segmentation output) to fulfill ship trajectory data association.