Marine Image Segmentation for Autonomous Obstacle Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current obstacle detection systems in autonomous ships are inadequate due to inaccuracies in GPS, AIS updates, and radar limitations, making it difficult for autonomous vessels to accurately sense obstacles without visual inspection.
Innovation Solution
A neural network-based method for image segmentation is employed to analyze marine images, identifying obstacles and their distances, which updates the obstacle map and generates control signals for navigation, using a combination of labelling data and error calculation to refine the neural network's output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If ECDIS and GPS are used for obstacle detection, then navigation information is provided, but detection accuracy is reduced due to GPS inaccuracy and AIS update periods
Solution Approach 1:
The patent combines multiple detection systems (ECDIS, GPS, AIS, radar, and camera-based image segmentation) into an integrated obstacle detection system. The neural network processes visual data to identify obstacles and their positions, which are then fused with data from other sensors to compensate for individual system limitations and improve overall detection accuracy.
Solution Approach 2:
The patent introduces an intermediary processing layer (neural network-based image segmentation system) that bridges the gap between raw sensor data and reliable obstacle detection. This intermediary analyzes visual information to provide precise obstacle position data that compensates for the inaccuracies in GPS and AIS systems.
2Area of stationary object
If radar is used for obstacle detection, then detection range is extended, but detection accuracy is reduced due to non-observation regions and noise
Solution Approach 1:
The patent merges radar detection with camera-based visual detection. The radar provides broad coverage and early warning of obstacles, while the camera system with neural network processing provides precise identification and position data, compensating for radar's non-observation regions and noise issues.
Solution Approach 2:
The system uses feedback from the neural network's image segmentation to refine and correct radar detection results. The visual data provides ground truth information that can be used to adjust radar parameters and improve overall detection precision in subsequent cycles.
3Measurement precision
If visual inspection is performed for accurate obstacle detection, then detection accuracy is improved, but automation is reduced requiring human operators
Solution Approach 1:
The patent replaces the mechanical/ human visual inspection system with an automated neural network-based image segmentation system. The neural network processes camera images to automatically identify obstacles, classify them, and determine their positions with high accuracy, enabling autonomous navigation without human operators.
Solution Approach 2:
The neural network system performs self-service by automatically analyzing visual data, making decisions about obstacle detection and classification, and providing output that can be directly used for navigation control, eliminating the need for human intervention in the detection process.
Data Source
AI summary
A method for learning a neural network performed by a computing means wherein the neural network receives a marine image and outputs information about a type and a distance of at least one object included in the marine image is provided. The method comprises: obtaining a marine image including a sea and an obstacle; obtaining a labelling data generated based on the marine image; obtaining an output data by using a neural network, wherein the neural network receives the marine image and outputs the output data; calculating an error value by using the labelling data and the output data; and updating the neural network based on the error value; wherein the labelling data and the output data are determined by a combination of information about a type and a distance of an object.