Revolutionizing Maritime Safety: AI-Driven Ship Detection and Tracking

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Sep 15 2023

In a paper published in the journal Applied Sciences, researchers unveiled a highly accurate ship detection and tracking model for marine applications to overcome data scarcity issues. Using few-shot learning and innovative transfer learning from a highway vehicle dataset has significantly improved inland ship detection.

The model utilizes Shuffle Attention and smaller anchor boxes, leading to swift training. The model achieves an impressive accuracy of 84.9% at an intersection-over-union threshold of 0.5 (mAP=0.5) with a training dataset comprising only 585 images. This performance surpasses existing methods, promising advancements in intelligent ships and marine monitoring systems. 

Background

The foundation of intelligent shipping lies in autonomous navigation. It primarily focuses on situational awareness, especially in the detection and tracking of high-risk, collision-prone target ships. Artificial intelligence (AI) and high-speed processors have propelled computer vision-based ship object detection and tracking. However, achieving precision in these models often requires copious, well-labeled data, posing challenges for inland ships due to limited samples, time-consuming labeling, extended training periods, and limited generalization.

The utilization of deep learning algorithms and Synthetic Aperture Radar (SAR) imagery has made significant progress in the prior investigations focused on ship object detection and tracking. This progress is exemplified by the introduction of advanced models like You Only Look Once (YOLOX) and enhanced CenterNet networks, both of which have played a pivotal role in enhancing detection accuracy. The application of ship tracking has seen both traditional and deep learning-based methods addressing challenges like data quality and ship scale variations. Remarkably, the Multi-Object Tracking (MOT) problem has garnered significant attention; however, challenges such as prolonged occlusion for slower vessels still require resolution.

Proposed Method

Comprehensive Multi-Object Detection for Inland Ships: This paper presents a comprehensive approach for multi-object detection and tracking of inland ships that aims to overcome challenges associated with small datasets, lengthy training times, and limited generalization. The methodology combines the YOLOv5s architecture, the Shuffle Attention mechanism, and transfer learning strategies. Smaller anchor boxes are introduced to enhance the detection accuracy for smaller objects. This achieves improved performance for buoys and similar targets. Furthermore, a novel Alpha-IoU loss function is introduced to enhance the precision of bounding box localization. The parameter α is fine-tuned to enhance robustness, particularly in scenarios involving smaller datasets.

The effectiveness of this methodology is demonstrated through experiments by showcasing its applicability in challenging scenarios with limited training data.

Furthermore, the implementation of this approach encompasses various enhancements, including Mosaic data augmentation, adaptive anchor box calculation, and adaptive image scaling inspired by YOLOv4. These modifications contribute to the adaptability of the model and overall robustness. Overall, this framework presents a holistic solution to the detection and tracking of inland ships to address the critical limitations associated with small datasets and offer improved accuracy and generalization.

Inland Ship Multi-Object Tracking: The DeepSORT algorithm builds upon the Sort object tracking method by enhancing it with additional features such as Hungarian matching, Kalman filtering, and a deep learning model for ReID (Re-Identification). This combination allows for real-time object tracking focusing on appearance features to minimize ID switching and improve tracking continuity. In ship tracking, DeepSORT is seamlessly integrated with the YOLOv5 model to achieve reliable and continuous tracking results.

The algorithm begins by processing video frames through YOLOv5 to extract object locations and confidence scores. It then employs Kalman filtering to predict and update object positions. Cascade and Intersection over Union (IoU) matching policies associate tracks with detections that ensure robust tracking. Finally, Kalman filtering updates are applied to correct and refine the matched tracks, ultimately completing the tracking process.

The DeepSORT algorithm forms a robust framework for ship detection and tracking in maritime scenarios. Its effective combination of traditional tracking techniques with deep learning models enhances real-time tracking. This is achieved by incorporating matching policies and Kalman filtering, which minimize ID switching and ensure tracking continuity across frames. This approach demonstrates its practical utility in ship-tracking applications that provide simplicity and cost-effectiveness while delivering accurate results.

Experimental Analysis

The impact of three key factors on our ship object detection framework is examined: the Shuffle Attention mechanism, transfer learning strategy, and smaller anchor boxes. These experiments are conducted on a laptop with specific hardware and software configurations. The images are categorized into seven classes, representing different types of ships found along the Yangtze River. The sensitivity analysis demonstrates that the inclusion of Shuffle Attention, transfer learning, and smaller anchor boxes results in improved object detection accuracy and specific ship categories.

Furthermore, the framework is compared with other state-of-the-art models to showcase its superior performance in terms of accuracy and processing speed while maintaining compact model parameters. Multi-object detection and tracking experiments are also conducted under different visibility conditions that show the effectiveness of the framework in tracking cargo ships and buoys. However, challenges arise in poor visibility that leads to lower accuracy. This analysis comprehensively explains the ship object detection framework's strengths and limitations across various scenarios.

Conclusion

In summary, this paper introduces a transfer-learning-based approach for ship object detection and tracking in low-data scenarios. The YOLOv5 network is enhanced with Shuffle Attention and smaller anchors. The transfer learning method is also applied using the UA-DETRAC high-speed dataset. The DeepSORT tracking algorithm is incorporated to achieve accurate vessel detection and tracking with limited data. The DeepSORT tracking algorithm is incorporated to achieve accurate vessel detection and tracking with limited data. The model is fine-tuned with a Yangtze River vessel dataset.

These contributions encompass improvements to the YOLOv5 network, the creation of a low-data tracking method, and the application of transfer learning. As a result, the approach achieves an impressive accuracy of 84.9% (mAP=0.5) with just 585 training images. Future research avenues may involve improving ship detection accuracy in extreme weather conditions using LiDAR data and addressing challenges related to rapid ID switching during ship crossings.