| CPC G06N 3/08 (2013.01) [B60Q 9/00 (2013.01); B60R 16/0231 (2013.01); G06N 3/04 (2013.01)] | 16 Claims |
|
1. A method for training a machine learning model for determining quality grades of data sets from each of a plurality of sensors, the plurality of sensors being configured to generate surroundings representations, wherein each of the plurality of sensors is configured to sense an environment using a respective sensing modality for generating a respective one of the data sets and the quality grade is a measure of an accuracy with which the respective one of the sensors that has generated the respective data set is to perform its sensing of the environment using the respective sensor's respective sensing modality in a current state, the method comprising the following steps:
providing a multitude of training data sets of each of the plurality of the sensors, wherein the multitude of training data sets correspond to a multitude of the surroundings representations, the multitude of training data sets including at least two data sets generated by two different ones of the plurality of sensors simultaneously sensing overlapping fields of view, the respective sensing modalities of the two different ones of the plurality of sensors being different than each other;
providing attribute data of ground truth objects, which are labels describing characteristics of the multitude of surroundings representations;
determining a quality grade of each respective training data set of each respective sensor of the plurality of sensors using a metric, the metric comparing at least one variable, which is determined using the respective training data set, with at least one attribute datum of at least one associated one of the ground truth objects of a respective one of the surroundings representations; and
training the machine learning model using the multitude of training data sets of each of the plurality of the sensors and the respectively determined quality grades to, in the future, determine quality grades of future data sets from the plurality of sensors without use of ground truth objects;
wherein:
the training includes an automatic iterative modification of the machine learning model, the multitude of training data sets being a single input data set upon which the iterative modification of the machine learning model is performed in combination; and
the training is performed such that, after the training, the machine learning model is configured to operate on the future data sets, obtained simultaneously by the two different ones of the plurality of sensors as a single machine learning model input, in combination to assign respective quality grades to each of the future of data sets, with different ones of the future data sets, included as the single machine learning model input, affecting the quality grade assignments of one another by the trained machine learning model.
|