The aim of the DataSOW project for the development of autonomous inland navigation was to carry out tests on the Spree-Oder waterwayin order to collect extensive data on infrastructure objects along the route. This data was then used to train and optimise an AI module. This provided an important basis for the navigation of autonomously operating inland vessels. The Federal Ministry of Digital Affairs and Transport funded the project as part of the mFUND innovation initiative with a total of 42,780 euros over a period of one year.
Inland navigation The inland navigation sector is facing great challenges concerning the implementation of autonomously operating inland vessels. This is due to the very complex environmental situation in regard to inland waterways. Our project DataSOW aims to create a data and algorithm basis for the recognition of the infrastructure on inland waterways. So, we will support the development of autonomously operating inland vessels.
As a first step, image data of infrastructure components will be recorded using industrial standard cameras. The test runs for the acquisition will be carried out on the Spree-Oder-Waterway in cooperation with the Spree Havel Waterways and Shipping Authority (WSA). The tests take place over a cycle of one year under different lighting and weather conditions as well as in different vegetation periods. The collected data will be compiled in a dataset and used for the development of an AI module. The AI module is to provide targeted detection, classification and direction data from the camera system for assistance and control systems (e.g. navigation) in real time.
What Does Artificial Intelligence Mean For DataSOW?
- Complex algorithms for dealing with given decision-making and evaluation problems whose solution path has been learned beforehand
- Visual computing: AI learns from existing images what a certain object looks like and then transfers this rule to new images
How Does the AI Module Work?
- Neural network consisting of several layers (mostly filter and aggregation layers)
- Filter layer: extraction of primary features
- Aggregation layer: condensation of primary features
- Processing of the input data via filter and aggregation layers enables classification
How is the AI Module Trained?
- adapt pre-trained neural network to own use case, i.e. the recognition of infrastructure components along the SOW
- To do this, the last layer is replaced and the neural network is trained with new data
- Combining several networks to process navigation signs with numbers
What is necessary for the training?
- high-quality training data
- Data volume and variation: over a yearly cycle, extensive image material is acquired under different light and weather conditions and in different vegetation periods
- Avoidance of misinterpretation