Computer-implemented method for operating an ego vehicle and ego vehicle

Abstract

The invention relates to a computer-implemented method for operating an ego vehicle (1), comprising an operating system and a plurality of components for operating the ego vehicle (1), wherein the components comprise at least one sensor system (2) with one or more image sensors for capturing the surroundings as individual images, comprising the steps of: - providing a machine learning model ensemble (4) comprising a plurality of trained machine models, wherein each machine learning model is trained, validated and / or tested in its domain for classification, localization and / or tracking of objects and / or image segmentation in the individual images, and providing a respective operating system configuration for each machine learning model, wherein a machine learning model used in the ego vehicle (1) represents an original operating machine learning model with its operating system configuration for a domain,- Estimating an aleatoric and / or an epistemic uncertainty for individual images or sequences of individual images (...),

Claims

[1] Computer-implemented method for operating an ego vehicle (1), the ego vehicle (1) comprising a configured operating system and several components for operating the ego vehicle (1), wherein the components comprise at least one sensor system (2) with one or more image sensors for capturing the surroundings as individual images characterized by the steps: - Providing a machine learning model ensemble (4), wherein the machine learning model ensemble (4) comprises a plurality of trained machine models, wherein each machine learning model is trained, validated and / or tested in its domain for classification, localization and / or tracking of objects and / or image segmentation in the individual images, and providing a respective operating system configuration for each machine learning model, wherein a machine learning model used in the ego vehicle (1) represents an original operating machine learning model with its operating system configuration for a recognized domain, - Estimating an aleatoric uncertainty and / or estimating an epistemic uncertainty for individual images or image sequences of individual images on which the original machine learning model operates in a deployment phase with respect to its domain as a measure of the performance of the machine learning model with regard to the classification, localization and / or tracking of objects and / or image segmentation, where a high uncertainty corresponds to a low performance and a low uncertainty to a high performance, - Perceiving a high uncertainty above a given threshold as an operation of the ego vehicle (1) outside the domain underlying the currently operating original machine learning model and recognizing the new domain, - Selecting a new machine learning model from the machine learning model ensemble (4) which is associated with the detected new domain and replacing the original operating machine learning model as the new replaced machine learning model or adding the new machine learning model to the original operating machine learning model. [2] Computer-implemented method according to claim 1, characterized by that the new replaced machine learning model or the added new machine learning model with the original operating machine learning model thus becomes the original machine learning model in use. [3] Computer-implemented method according to claim 1 or 2, characterized by that the original operating system configuration is adapted to the new machine learning model or to the combination of the new machine learning model with the original operating machine learning model. [4] Computer-implemented method according to claim 3, characterized by that the operating system configuration is accomplished with respect to a resolution of the individual images recorded by the image sensors of the sensor system (2). [5] Computer-implemented method according to claim 3 or 4, characterized by that the ego vehicle (1) comprises several computing components for evaluating the captured individual images and the operating system configuration is carried out with regard to the computing components. [6] Computer-implemented method according to one of the preceding claims, characterized by that each machine learning model in the machine learning model ensemble (4) is assigned its own threshold. [7] Computer-implemented method according to one of the preceding claims, characterized by that each machine learning model is assigned a different threshold with respect to aleatoric uncertainty and epistemic uncertainty. [8] Computer-implemented method according to one of the preceding claims, characterized by that epistemic uncertainty and / or aleatoric uncertainty are determined in real time. [9] Computer-implemented method according to one of the preceding claims characterized by that the epistemic uncertainty is determined using a Monte Carlo dropout procedure. [10] Computer-implemented method according to one of the preceding claims, characterized by that the machine learning models are trained as random forest and / or ensemble neural networks. [11] Computer-implemented method according to one of the preceding claims, characterized by that the uncertainties are determined by the last n individual images, where n is an adjustable number, or are determined for each k-th individual image, where k is an adjustable number. [12] Ego vehicle (1) comprising a correspondingly configured operating system and several components for operating the ego vehicle (1), wherein the components comprise at least one sensor system (2) with one or more image sensors for capturing the surroundings as individual images, characterized by , that a storage unit is provided in which a machine learning model ensemble (4) is stored, wherein the machine learning model ensemble (4) comprises a plurality of trained machine models, wherein each machine learning model is trained, validated and / or tested in its domain for classification, localization and / or tracking of objects and / or image segmentation in the individual images, and wherein each machine learning model has an operating system configuration, further comprising a control unit (5) which is designed to apply a corresponding machine learning model as an originally operating machine learning model with its operating system configuration for a recognized domain and further comprising a configurator (6) which is designed to estimate the aleatoric uncertainty and / or to estimate an epistemic uncertainty for individual images,on which the original machine learning model operates in a deployment phase with respect to its domain as a measure of the performance of the machine learning model with regard to the classification, localization and / or tracking of objects and / or image segmentation, wherein a high uncertainty corresponds to a low performance and a low uncertainty corresponds to a high performance, and which is further designed to perceive a high uncertainty above a predetermined threshold value as an operation of the ego vehicle (1) outside the domain underlying the currently operating original machine learning model and which is further designed to recognize the new domain, and wherein the configurator (6) is further designed to select a new machine learning model from the machine learning model ensemble (4) which is assigned to the recognized new domain,and replacing the original operating machine learning model as a new replaced machine learning model or adding the new machine learning model to the original operating machine learning model. [13] Ego vehicle (1) according to claim 12, characterized by that the sensor system (2) has at least one or more image sensors for detecting the surroundings of the ego vehicle (1). [14] Ego vehicle (1) according to claim 13, characterized by that the configurator (6) is designed to adapt at least the original operating system configuration of the sensor system (2) to the replaced new machine learning model or the combination of the new machine learning model with the original operating machine learning model. [15] Ego vehicle (1) according to claim 14, characterized bythat the operating system configuration comprises at least one resolution of the individual images recorded by the image sensors of the sensor system (2), wherein the resolution comprises a sampling rate, an image resolution and / or a signal level.

Images