These are neural modules for estimating the position and orientation of a vehicle by combining 1) GNSS measurements with 2) available signals on in-vehicle sensors and 3) occasional observations of surrounding landmarks.
The neural modules combine flexibly into a localization filter, adapting to the available signals. The filter is composed of a neural predictive model, which uses on-board sensors, and a neural corrective model that fuses the prediction with data from GNSS and sparse position observations. Neural networks have features that make them more efficient and accurate than classical position estimation solutions: they have a physics-inspired architecture that makes them interpretable and reduces the number of parameters, thereby reducing (eliminating) the risk of overfitting; they require a limited number of examples to train; and they have no need for sensor calibration. They are flexible by being able to accept different sensor configurations; they are easily converted to independent C++ code.
Addressing target audiences and expressing needs
- Business Angels
- Venture Capital
- Other type of Investment
Automotive OEM (original equipment manufacturers) and suppliers to the automotive industry.
- Research and Technology Organisations
- Private Investors
R&D, Technology and Innovation aspects
The technology has been demonstrated (TRL about 8). Needs investors, adopters, and deployment on production vehicles.
The neural network architecture is modular, flexible, and customizable. It can be easily fine-tuned to different types of sensors/vehicles with little effort.
The neural network is customizable at two levels: hyperparameters-modular configuration, and training. The latter requires very little training data because the networks have very few efficient parameters.
The solution improves the quality of vehicle control and driver assistance functions (including self-driving).
