A Self-Selective Correlation Ship Tracking Method for Smart Ocean Systems

Xu Kang, Bin Song, Jie Guo, Xiaojiang Du, Mohsen Guizani2019 · DOI

TL;DR

This paper introduces a novel "self-selective correlation filtering" (BRCF) method for robust ship tracking in complex marine environments. By addressing limitations of traditional methods like boundary effects and scale changes, BRCF achieves significantly higher accuracy and faster processing speeds, making it well-suited for "Smart Ocean" applications.

What Problem Does This Paper Address?

Tracking ships in marine environments presents unique challenges due to the dynamic nature of the sea, weather conditions, and the sheer scale of operations. Traditional object tracking methods, particularly those based on correlation filters, often struggle with:

  • Boundary Effects: When a target is near the edge of an image or during significant shifts, the circular nature of correlation filters can lead to incorrect feature extraction and tracking errors.
  • Scale Variation: Ships can appear at vastly different sizes due to their distance from the camera. Traditional methods often require separate scale estimation or struggle to adapt to rapid changes in target size and aspect ratio.
  • Background Interference: Cluttered or similar-looking backgrounds can confuse trackers, leading to false positives or loss of the target.

These issues are critical for "Smart Ocean" systems, which rely on accurate real-time monitoring for maritime security, resource management, and intelligent navigation.

Methods

The proposed Self-Selective Correlation Filtering (BRCF) method tackles these problems with a multi-pronged approach:

  • Self-Selective Model with Hard Negative Samples Mining: This is a core innovation. Instead of using all negative samples, BRCF intelligently selects "hard" negative samples (those that are difficult to classify correctly). This reduces the boundary effect and improves the classifier's ability to distinguish targets from challenging backgrounds.
  • Adaptive Multi-Response Fusion: The system integrates multiple feature types (e.g., HOG, color histogram, local binary pattern histogram). A self-selective model then adaptively fuses the responses from these different feature extractors, enhancing robustness against varying environmental conditions and target appearances. Kullback-Leibler (KL) divergence is used to evaluate the reliability of each feature's prediction, guiding the fusion process.
  • Bounding Box Regression with Scale Pre-estimation:
    • Scale Pre-estimation: Key point matching (using SURF features) is employed to roughly estimate the target's scale between frames. This helps anticipate changes in target size.
  • Bounding Box Regression: A dedicated regression model is trained to precisely adjust the bounding box. This allows for more flexible adaptation to target deformation and scale changes, going beyond simple rectangular adjustments.

Datasets

The paper utilizes a custom dataset collected by the authors from their laboratory's "marine traffic dataset." This dataset consists of ten video sequences with a resolution of 1920x1080 and a frame rate of 25 fps, featuring ships sailing into or out of bays. The authors state that this dataset was used for evaluating the performance of their proposed method against existing ones.

Key Insights and Ideas

  • "Hard" Negative Samples are Key: The insight that focusing on difficult-to-classify negative samples is more effective than using all of them for training is crucial. This directly addresses the boundary effect and improves discriminative power.
  • Feature Fusion for Robustness: Combining diverse features (HOG, color, LBP) and adaptively weighting them based on their reliability (via KL divergence) significantly boosts tracking performance in challenging scenarios.
  • Two-Stage Scale Handling: The combination of a fast scale pre-estimation step (key point matching) followed by a precise bounding box regression step offers an efficient and effective way to handle scale variations and target deformations.
  • Adaptability Beyond Rectangles: The bounding box regression's ability to adjust aspect ratios and handle deformations is a significant advantage over methods that rely on fixed-shape bounding boxes.

Key Results

The BRCF method demonstrated superior performance compared to traditional methods like KCF and DSST on their marine traffic dataset:

  • Accuracy: BRCF achieved an average success rate and precision over 8% higher than DSST. The average overlap ratio was also higher, and the average distance between the predicted and ground-truth bounding box was significantly lower (19.85 pixels for BRCF vs. 28.41 for DSST).
  • Speed: BRCF processed nearly 22 frames per second faster than DSST, achieving almost real-time performance.
  • Robustness: The method effectively handled object size changes and background interference, as evidenced by its ability to maintain tracking even when DSST and KCF lost the target in challenging sequences.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while BRCF is more adaptable, it might still struggle with extreme deformations or rapid, unpredictable changes.
  • The performance comparison is primarily against KCF and DSST on their specific dataset. Generalizability to other types of marine objects or environments might require further validation.
  • Future Work: The authors suggest that the proposed method can effectively deal with object size changes and background interference, implying that further improvements in these areas are possible. The real-time processing capability also opens doors for integration into more complex "Smart Ocean" systems.

Summary updated

BibTeX
@misc{kang2019self,
  author = {Xu Kang and Bin Song and Jie Guo and Xiaojiang Du and Mohsen Guizani},
  title = {{A Self-Selective Correlation Ship Tracking Method for Smart Ocean Systems}},
  year = {2019},
  doi = {10.3390/s19040821},
}
Figure