Fast Ship Tracking Algorithm for Remote Sensing Video Based on Background Data Mining and Adaptive Selection

Ship tracking for remote sensing video has strong background characteristics. A single network structure hardly adapts to various environmental characteristics, where the tracking in the ocean background needs to have a strong real-time performance, while it tends to pursue the accuracy of detection when there exists land interference in the scene. This paper proposes a self-selecting neural network SSNN to comprehensively utilize the superiority of deep neural network and traditional tracking algorithm based on correlation filtering. In our network, video sequences are classified into tow categories according to the background information through the pre-classification module every fixed number of frames. MOSSE is introduced to pursue the tracking speed under the ocean background, and improved Siamese-RPN tracks the video sequence under the complex background with high precision. We introduce template bags to achieve the balance of training bounding boxes and learning matching methods in Siamese-RPN. Innovatively, category branch in improved SOLOv2 is decomposed into coarse branch and fine branch to filter interference information from land. To weaken tracking loss caused by scene change and strengthen the integration between sub-networks, we set multiple loss return mechanisms by comparing the degree of difference between adjacent bounding boxes. The proposed algorithm achieves a tracking speed of 207 fps while reaching 0.426 EAO tested on our dataset. Experimental results show that SSNN outperforms advanced trackers combined accuracy and speed.