Maritime Situation Awareness Using Neural Obstacle Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current obstacle detection systems in autonomous ships, such as ECDIS and radar, are limited by GPS inaccuracy, AIS update periods, and noise, making it difficult to accurately sense obstacles without visual inspection, which hinders the development of fully autonomous navigation.
Innovation Solution
A neural network-based method for image segmentation is employed to analyze marine images, identifying types and distances of objects, and update obstacle maps dynamically, incorporating camera data and ship position information to generate control signals for navigation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If ECDIS and radar are used for obstacle detection, then automated navigation is enabled, but detection accuracy deteriorates due to GPS inaccuracy, AIS update periods, and noise
Solution Approach 1:
The patent combines multiple sensing systems (ECDIS, radar, AIS, and camera-based vision systems) into an integrated obstacle detection system. By merging these different sensing modalities, the system compensates for the limitations of individual sensors - GPS inaccuracy is offset by visual positioning, AIS update delays are compensated by radar and camera data, and radar noise is filtered by cross-validation with other sensors. This multi-sensor fusion enables both automated navigation and high detection accuracy.
Solution Approach 2:
The patent introduces an image processing and neural network analysis system as an intermediary between the physical environment and the navigation control system. The camera captures visual data, which is then processed through image segmentation and object recognition algorithms to generate accurate obstacle detection results. This intermediary processing layer transforms raw sensor data into reliable navigation information, improving both automation capability and detection precision.
2Measurement precision
If visual inspection is performed to accurately detect obstacles, then detection accuracy improves, but operational complexity increases and autonomous operation becomes difficult
Solution Approach 1:
The patent implements self-service through autonomous image processing and automated obstacle recognition. The neural network system automatically captures images, processes them through segmentation algorithms, identifies obstacles, and generates navigation decisions without human intervention. This self-service capability maintains high detection accuracy while reducing operational complexity, as the system performs all visual inspection tasks autonomously without requiring manual oversight.
Solution Approach 2:
The patent replaces manual visual inspection with an automated computer vision system. Instead of human operators visually examining the environment, the system uses cameras coupled with neural network-based image processing to automatically detect and classify obstacles. This substitution of mechanical/human visual inspection with automated electronic sensing and processing reduces operational complexity while maintaining or improving detection accuracy.
3Reliability
If multiple sensor systems are integrated for comprehensive obstacle detection, then detection reliability improves, but system complexity increases
Solution Approach 1:
The patent segments the overall obstacle detection task into distinct functional modules: ECDIS for electronic chart data, radar for radio wave-based detection, AIS for identification information, and camera-based vision systems for visual detection. Each sensor system operates as an independent module with specialized processing, and their results are integrated through a fusion algorithm. This segmentation approach improves reliability through multi-sensor validation while managing complexity by keeping each module independent and well-defined.
Data Source
AI summary
A method for situation awareness is provided. The method comprises: preparing a neural network trained by a learning set, wherein the learning set includes a plurality of maritime images and maritime information including object type information which includes a first type index for a vessel, a second type index for a water surface and a third type index for a ground surface, and distance level information which includes a first level index indicating that a distance is undefined, a second level index indicating a first distance range and a third level index indicating a second distance range greater than the first distance range; obtaining a target maritime image generated from a camera; and determining a distance of a target vessel based on the distance level index of the maritime information being outputted from the neural network which receives the target maritime image and having the first type index.