Introduction
Maritime surveillance is a critical component of "Smart Ocean" systems, yet tracking ships at sea presents unique challenges for computer vision. Traditional Correlation Filter (CF) methods often suffer from the boundary effect, where circular shifts used for training create unrealistic samples, leading to model drift. Furthermore, ships frequently undergo significant changes in scale and aspect ratio as they maneuver, which standard fixed-ratio trackers fail to capture. To address these issues, researchers have developed the Box Regression Correlation Filter (BRCF), a self-selective tracking algorithm designed for high accuracy and real-time performance in complex maritime environments.
Methodology
The BRCF framework introduces three primary innovations to enhance tracking robustness and efficiency.
Hard Negative Samples Mining
To mitigate the boundary effect, the model moves away from using all circular shifts. Instead, it selectively mines "hard" negative samples from the local search region. By focusing on samples that are difficult for the classifier to distinguish from the target, the model strengthens its discriminative power against complex backgrounds.
Adaptive Multi-Feature Fusion
The tracker utilizes a combination of diverse features: Histogram of Oriented Gradients (HOG) for shape, Color Histograms (CH) for appearance, and Local Binary Patterns (LH) for texture. The integration of these features is governed by an adaptive mechanism using Kullback–Leibler (KL) divergence. By measuring the reliability of each feature's prediction in real-time, the system dynamically adjusts their weights, favoring the most reliable feature under current environmental conditions (e.g., prioritizing HOG over CH in cases of severe glare).
Two-Step Scale Estimation
Accurate bounding box adjustment is achieved through a two-step process:
- Pre-estimation: Speeded Up Robust Features (SURF) keypoint matching between consecutive frames provides an initial estimate of scale changes.
- Box Regression: A dedicated regression model fine-tunes the bounding box, allowing for independent adjustments to width and height. This enables the tracker to adapt to changes in the ship's aspect ratio as it turns or tilts.
Experimental Setup
The BRCF method was evaluated using a Marine Traffic Dataset consisting of 10 to 20 high-definition (1920 × 1080) video sequences captured at 25 FPS. These sequences represent diverse and challenging scenarios, including ships entering or exiting bays, background interference from waves, and significant variations in distance from shore-based and ship-borne cameras.
Results and Discussion
Comparative analysis against industry-standard trackers, specifically Kernelized Correlation Filter (KCF) and Discriminative Scale Space Tracking (DSST), demonstrated the superiority of the BRCF approach.
- Accuracy and Precision: BRCF achieved a success rate and precision over 8 percentage points higher than DSST on the marine traffic dataset.
- Real-Time Performance: On a standard Intel i7-4790 CPU, BRCF reached an average processing speed of 44.98 FPS, nearly 22 FPS faster than DSST. This efficiency stems from the sparse selection of hard samples, which reduces the computational load.
- Robustness: Qualitative evaluations showed that BRCF maintains a stable lock on targets even during long-range movement or heavy background clutter, where KCF and DSST often fail.
The use of KL divergence for reliability estimation proves highly effective for handling dynamic maritime environments. Furthermore, breaking the "rigid box" constraint through regression allows the model to accurately track the ship's silhouette throughout various maneuvers.
Conclusion
The BRCF algorithm provides a significant advancement in maritime ship tracking by combining selective sample mining with adaptive feature fusion and flexible scale estimation. While highly effective, the method's reliance on SURF keypoints suggests potential limitations in extremely low-visibility conditions (e.g., heavy fog). Future work will focus on evaluating the system on larger, global datasets and optimizing the algorithm for deployment on specialized edge hardware like NVIDIA Jetson modules for use in autonomous maritime platforms.
Summary updated