A Real-Time Method for Time-to-Collision Estimation from Aerial Images

Area of Science:

• Maritime safety
• Computer Vision
• Artificial Intelligence

Background:

• Container ships require skilled pilots for navigation in complex maritime environments.
• Accurate prediction of potential collisions is crucial for maritime safety.

Purpose of the Study:

• To develop a deep learning-based method for estimating time-to-collision (TTC) between vessels using aerial drone video data.
• To enhance the accuracy and robustness of TTC estimation in dynamic maritime settings.

Main Methods:

• Utilized deep learning for object detection, segmentation, and tracking from real-time drone video.
• Employed virtually generated data for feature optimization.
• Incorporated rotated bounding box representations from semantic segmentation for improved accuracy.

Main Results:

• Achieved precise, robust, and efficient TTC prediction for dynamic objects from a top-view perspective.
• Demonstrated a low mean error (0.358 s) and standard deviation (0.114 s) in worst-case scenarios.
• Visualized collision estimates using intuitive color-changing arrows.

Conclusions:

• The proposed deep learning approach effectively predicts vessel collisions using drone imagery.
• The method offers a reliable solution for enhancing maritime safety through advanced video analysis.