Sea-IoUTracker: A more stable and reliable maritime target tracking scheme for unmanned vessel platforms

Background: As mankind continues to exploit marine resources, there is a growing need to use unmanned surface vessel (USV) platforms for maritime security and surveillance of the operational environment at sea. Satellite remote sensing imagery is now the most common means used for maritime surface security surveillance. Satellite images are susceptible to cloud cover, while camera-equipped USVs can prevent the effects caused by cloud cover. However, there are virtually no vision-based target tracking solutions for the USV perspective. Methods: This paper proposes a tracking solution for multiple target tracking at sea named Sea-IouTracker. It can address the potential problems when the USVs are at sea, such as the target loss, anchor frame drift and even serious Re-ID phenomena due to wind and waves and other environmental effects. We chose Sort as the basic tracking framework and we picked Buffered-IoUs based on two different buffer scales instead of traditional IoUs for matching targets in neighbouring frames. And the innovation of the neighbouring frame matching mechanism lies on that the position of the Buffered-IoU is predicted using a motion prediction model and matched with the next detection frame because both matches are unsuccessful. This method increases the stability of the tracking model and reduces the phenomenon of re-ID. We have also developed a detector that adaptively fuses image features for the tracking scheme. The detector proposed is based on the ASFF (Adaptively Spatial Feature Fusion) detector head, CBAM (Convolutional Block Attention Module) and Yolov7 baseline for maritime target detection. Results and Conclusions: This paper verifies that the Sea-IoUTracker can achieve MOTA = 77.5, ID SW = 1.5% when tracking multiple targets on the MOT17 dataset. We have also proved our detector proposed is superior to other SOTA detectors in accuracy when detecting sea surface targets in the Singapore Maritime Dataset (SMD). We have also verified that the target tracking scheme proposed in this paper outperforms to the current state-of-the-art(SOTA) tracker C-BIoU (Yang et al., 2022) Tracker and ByteTrack based on four tracking video sequences in SMD. In this case, the accuracy is improved by about 2% with respect to the latter, while the ID SW phenomenon is reduced by about 20%.