Position determination by means of neural networks
A method for training an artificial neural generator network for the generating of synthetic landmark images is provided, in which landmark images are extracted from at least one training map as training data, which form a first training data set, and the generator network as a generator of a generative adversarial network learns with the aid of the first training data set and with the aid of real, non-annotated image data, recorded by a sensor device, to generate synthetic landmark images which are suited to reproducing the probability distribution underlying the first training data set. The invention also relates to a method for training an artificial neural localization network by the trained generator network and a method for determining a position of a mobile unit with at least one sensor device and at least one environment map by the trained localization network.
1 . A method for training an artificial neural generator network that generates synthetic landmark images, comprising:
extracting landmark images from at least one training map as training data, which form a first training data set, and
training the artificial neural generator network as a generator of a generative adversarial network using the first training data set and real, non-annotated image data, recorded by a sensor device, to generate synthetic landmark images which are suited to reproducing a probability distribution underlying the first training data set,
wherein an input of the artificial neural generator network comprises at least one image of the sensor device as real non-annotated image data.
2 . The method according to claim 1 , wherein the at least one training map comprises a first known landmark and/or an object class indication of a first known landmark.
3 . The method according to claim 1 , wherein the at least one training map is three-dimensional or two-dimensional.
4 . A method for training an artificial neural localization network, comprising:
performing a method for training an artificial neural generator network that generates synthetic landmark images, by extracting landmark images from at least one training map as training data, which form a first training data set, and wherein the artificial neural generator network as a generator of a generative adversarial network learns using the first training data set and using real, non-annotated image data, recorded by a sensor device, to generate synthetic landmark images which are suited to reproducing a probability distribution underlying the first training data set;
after performing the method for training of the artificial neural generator network, generating synthetic landmark images as training data, forming a second training data set; and
wherein the artificial neural localization network learns using the second training data set and using the at least one training map to determine a relative posture of an image of the second training data set relative to an image generated from the training map.
5 . The method according to claim 4 , wherein the artificial neural localization network is a convolutional neural network and/or is trained by a supervised machine learning method.
6 . A method for determining a position of a mobile unit with at least one sensor device and at least one environment map, including an initial position estimation, the method comprising:
performing a method for training an artificial neural generator network that generates synthetic landmark images, by extracting landmark images from at least one training map as training data, which form a first training data set, and wherein the artificial neural generator network as a generator of a generative adversarial network learns using the first training data set and using real, non-annotated image data, recorded by a sensor device, to generate synthetic landmark images which are suited to reproducing a probability distribution underlying the first training data set;
performing a method for training an artificial neural localization network, by, after performing the method for training of the artificial neural generator network, generating synthetic landmark images as training data, forming a second training data set; and wherein the artificial neural localization network learns using the second training data set and using the at least one training map to determine a relative posture of an image of the second training data set relative to an image generated from the training map; and
after the images recorded by the sensor device are processed by the artificial neural generator network and results are generated for the determination of a relative posture by the artificial neural localization network, performing an updating of an initial position of the mobile unit by a comparison between the initial position estimation and the relative posture from the artificial neural localization network.
7 . The method according to claim 6 , wherein when determining the position of the mobile unit an initial estimation of the position of the mobile unit is made by an odometric method.
8 . The method according to claim 6 , wherein when determining the position of the mobile unit an initial estimation of the position of the mobile unit is taken into account by at least one other sensor device.
9 . The method according to claim 6 , wherein the at least one environment map is three-dimensional or two-dimensional.
10 . The method according to claim 6 , wherein a position of the mobile unit comprises values for coordinates and/or angle of orientation.