Search

Visual multiobject tracking (MOT) on the maritime surface based on unmanned surface vessels (USVs) is a crucial technology for realizing intelligent maritime surveillance. However, videos obtained from the maritime surface often encounter challenges such as scene blurring, oscillations of USV observation platforms, and nonlinear movements of objects, significantly affecting the tracking performance. In response to these challenges, this article builds upon ByteTrack to introduce the maritime object motion tracker (MOMT), employing YOLOv7 as the object detector. Diverging from ByteTrack, MOMT proposes a new metric function and incorporates Gaussian cascade matching for a more refined handling of low-score and unmatched bounding boxes. In addition, an augmented observation Kalman filter (KF) module is developed to diminish the effects of nonlinear motion in objects. Specifically, first, in establishing the relationship between adjacent frames, the Fuse-distance metric function is proposed by synthesizing information on intersection over union (IoU) distance, object category, and size, to determine the optimal matching between detection boxes and KF prediction boxes. Second, the introduction of the Gaussian cascade matching module better addresses the issue of frequent identity switches (IDSs) caused by significant shaking of the USV observation platform. In addition, to cope with the impact of nonlinear object motion, the combination of camera motion compensation (CMC) and the observation-centric KF is employed to mitigate estimation errors in the KF. Finally, MOMT is compared with other state-of-the-art tracking algorithms on the Singapore Maritime Dataset, achieving competitive performance with 46.0 multiple object tracking accuracy (MOTA) and 66.7 IDF1, and operating in real time at a speed of 44.6 frames per second (FPS). The experimental results demonstrate that MOMT can achieve real-time and effective MOT in complex maritime scenarios, providing technical support for intelligent maritime surface monitoring.

The time-varying noise covariance matrices with respect to a linear Gaussian system are inaccurate, which may make the estimation accuracy of some filtering algorithms less than expected. To tackle the aforementioned issue, a modified variational robust filter is presented. In this algorithm, a sliding-window-based variational filter is employed, which improves the algorithm’s accuracy and efficiency via the latter state to adjust the previous state and avoids the iterations of fixed point. The inverse-Wishart distribution is considered as the observation noise covariance’s prior distribution and is capable of estimating the observation noise covariance and the system state via the variational Bayesian technique. The Sage-Husa filter is employed for estimating the state noise covariance, and a modified monitoring strategy for real-time adjustment of the state noise covariance is developed to ensure the matrix’s positive semidefinite. An adaptive factor is constructed for balancing the predicted and observation states to improve the algorithm’s robustness. Simulation and experiment results show that the proposed filtering algorithm performs better than some existing filtering algorithms with regards to convergence and accuracy.