Method for creating and deploying an enhanced environmental map

Abstract

Procedure for creating and deploying an enhanced environmental map for use in vehicles, comprising the following steps: - Receiving (S1) sensor data from sensors of at least one vehicle in a backend server; - Evaluating (S2) the sensor data in the backend server; - Defining (S3) location-dependent sensor data by assigning the sensor data to a position in an environment map; - Determining (S4) the detection accuracies of the sensors based on the location-dependent sensor data; - Entering (S5) the detection accuracies as sensor models into the environment map at the corresponding positions of the location-dependent sensor data to create an extended environment map; - Deploying (S6) the extended environment map to at least one vehicle.

Description

[0001] The invention relates to a method for creating and providing an extended environmental map.

[0002] It is already known from the state of the art that sensor data processing, for example, for localization, but also for data fusion and tracking, requires so-called sensor models. These models model sensor properties relevant to data processing, such as the accuracy of a measurement in the radial or azimuthal direction or the detection probability, i.e., the probability with which a received measurement represents a real target / object or with which a real target / object is overlooked or not detected by the sensor.

[0003] From DE 10 2016 202 317 A1, a method is known for controlling vehicle functions by a driver assistance system of a vehicle, comprising the steps of: providing sensor data relating to a vehicle environment of the vehicle by a plurality of vehicle sensors of the vehicle; classifying the vehicle sensors into different vehicle sensor type classes; creating an occupancy grid representing the vehicle environment of the vehicle; and controlling vehicle functions of the vehicle by the driver assistance system using the created occupancy grid.

[0004] WO 2020 / 060308 A1 discloses a method for controlling a vehicle performed by an electronic device, the method comprising: transmitting profile information of the vehicle to an external server; receiving accurate map data; and controlling the vehicle to perform autonomous driving using the received map data.

[0005] WO 2021 / 218 388 A1 describes a method for generating a high-precision map, a localization method and a device used to improve the localization accuracy of a high-precision map in a specific environment.

[0006] It is an object of the present invention to provide a method by means of which an extended environment map can be created and provided, thereby improving the accuracy and robustness of sensor detections.

[0007] This object is solved by the subject matter of independent claim 1. Further advantageous embodiments and embodiments are the subject matter of the subclaims.

[0008] Initial considerations were that sensor models today are generally configured once, for example, by determining an optimal parameter set over a certain number of measurements. In simple models, the same model is used for each measurement cycle. However, there are also approaches that adapt a model to specific scenarios, since, for example, the probability of a false positive detection is higher in a tunnel than on a "clear" road. A location-dependent adaptation of sensor models is not known.

[0009] Especially for localization and the static environment or free space, a large number of measurement points from the vehicle's surroundings are used. Following the example of the tunnel, the parameters of the sensor model depend heavily on the environment.

[0010] Since the described properties can usually be assigned to a fixed position in the world, the idea of ​​this invention is to learn a location-dependent sensor model and store it as an additional layer in a map.

[0011] In the following, the combination of detection probability and measurement accuracy will be summarized under the term detection accuracy.

[0012] According to the invention, a method for creating and providing an extended environment map for use in vehicles is proposed, comprising the following steps: - Receiving sensor data from sensors of at least one vehicle in a backend server; - Evaluating the sensor data in the backend server; - Specifying location-dependent sensor data by assigning the sensor data to a position in an environmental map; - Determining the detection accuracy of the sensors based on the location-dependent sensor data; - Entering the detection accuracies as sensor models into the environment map for the corresponding positions of the location-dependent sensor data to create an extended environment map; - Providing the extended surrounding map to at least one vehicle.

[0013] The sensor data can comprise raw sensor data or a raw data-like representation. It is also conceivable for data already classified in the vehicle to be transmitted to the backend server. Preferably, data from a large number of vehicles, e.g., a fleet, is transmitted to the server and evaluated there. This is particularly advantageous because it allows individual sensor-related detection inaccuracies to be identified.

[0014] When evaluating the sensor data, for example, objects in the vehicle's surroundings can be identified from the sensor data. When defining the location-dependent sensor data, the evaluated sensor data is assigned to positions on the surrounding map. The positions of the sensor data can be located on the map, for example, by assigning a GPS location. Another option, either alternatively or cumulatively, would be to enter the sensor data, for example, into an occupancy grid. The relative position of the recording vehicle could also be stored here. When determining the detection accuracy of the sensors based on the location-dependent sensor data, the knowledge of the environment from the location-dependent sensor data can be used to determine how likely a certain detection is or what accuracy can be expected in a specific environment. For example, vegetation generally leads to poorer detection accuracy.Building walls can be measured with very good detection accuracy, whereas parked vehicles, for example, provide very good measurement points, but appear in different locations depending on the parking situation and thus only lead to medium detection accuracies for localization, for example. Furthermore, it is particularly advantageous to learn the sensor models in the backend, since the required detection accuracy can only be determined inadequately during a single pass due to systematic errors. A parked vehicle, for example, looks like a very precise target during a single pass. Only after many passes with different parked vehicles in slightly different positions can the systematic errors be identified and good models for detection accuracy determined.

[0015] Entering the detection accuracies as sensor models is advantageous because in this way sensor models can be entered for different sensor types and made available for the corresponding sensors.

[0016] The resulting enhanced environment map is then made available to vehicles via download. There, the factors can be read by the localization, tracking, and fusion algorithms and used in internal signal processing to contribute to improved accuracy and robustness of the obtained position and object quality.

[0017] In a preferred embodiment, the sensor models are entered into the environmental map as an additional map layer. For example, the sensor models can be entered in an additional layer for each cell in an occupancy grid. This layer would, in principle, be evaluated by the fusion algorithms during the processing of a new sensor measurement, instead of, for example, statically stored values.

[0018] Particularly preferably, the sensor models include a factor for accuracy in the radial, azimuthal, and elevation directions, as well as probabilities for false positive and false negative detections. This is advantageous because these factors and probabilities can be directly incorporated into the sensor recordings at the corresponding positions, contributing to higher detection accuracy.

[0019] Furthermore, in a preferred embodiment, the location-dependent sensor data is determined based on localization and / or tracking and / or fusion of the sensor data. This can be performed in the vehicle or in the backend server. If executed in the vehicle, the already processed data would be transmitted to the server. If executed in the backend server, the respective vehicle would transmit raw data or raw-data-like representations to the server.

[0020] In a further preferred embodiment, fusion filters, static or dynamic environmental occupancy maps, or Kalman filters are used for localization and / or tracking and / or fusion. During localization, tracking, or fusion, the fusion filters, e.g., particle filters, static or dynamic environmental occupancy maps (occupancy grid), or Kalman filters, generate data, e.g., regarding the possibility of association or the accuracy of the position measurement by comparing the prediction of the internal filter state with the current measurement. From this data, a measure of the detection accuracy depending on the environment can be determined.This data can be provided directly by the vehicle or, as already mentioned, generated by executing the respective procedures in the backend server. The advantage of the backend is that, regardless of the specific vehicle type, the data processing is the same, there are fewer restrictions regarding computing time, and access to all internal states is easy.

[0021] Particularly preferred is the determination of location-dependent sensor data by classifying the sensor data with the help of a neural network. Classification using a neural network preferably takes place on the backend server, as there are no or fewer restrictions on computing power and time. Using CNNs, a semantic classification of the environment into streets, vegetation, buildings, parking spaces, etc. is performed in the raw data or on intermediate representations such as the Dynamic Grid. This classification can be used to advantageously determine detection accuracy. This method can also be combined with determination from internal data. For example, a convolutional neural network (CNN) or a recurrent neural network (RNN) can be used as the neural network, with the networks having been pre-trained with appropriately classified data.

[0022] Further advantageous embodiments and embodiments are shown in the drawings, in which: Fig. 1: a schematic representation of a method according to an embodiment of the invention.

[0023] Fig.1 shows a schematic representation of a method according to an embodiment of the invention. The following steps are carried out in the method for creating and providing an extended environmental map in vehicles. In step S1, sensor data from sensors of at least one vehicle is received in a backend server. Sensor data from a plurality of vehicles is preferably transmitted here. In a step S2, the sensor data is evaluated in the backend server. In step S3, location-dependent sensor data is determined by assigning the sensor data to a position in the environmental map. In step S4, the detection accuracies of the sensors are determined based on the location-dependent sensor data. In step S5, the detection accuracies are entered as sensor models in the environmental map at the corresponding positions of the location-dependent sensor data in order to create an extended environmental map.Finally, in step S6, the extended environment map is provided to at least one vehicle.

Claims

[1] A method for creating and providing an extended environment map for use in vehicles, comprising the following steps: - receiving (S1) sensor data from sensors of at least one vehicle in a backend server; - Evaluating (S2) the sensor data in the backend server; - Determining (S3) location-dependent sensor data by assigning the sensor data to a position in an environmental map; - Determining (S4) the detection accuracies of the sensors based on the location-dependent sensor data; - Entering (S5) the detection accuracies as sensor models into the environment map for the corresponding positions of the location-dependent sensor data to create an extended environment map; - Providing (S6) the extended environment map to at least one vehicle. [2] Method according to claim 1, characterized bythat the sensor models are entered into the environmental map as an additional map layer. [3] Method according to claim 1, characterized by that the sensor models include a factor for the accuracy in radial, azimuthal and elevation directions as well as probabilities for false positive and false negative detections. [4] Method according to claim 1, characterized by that the determination of the location-dependent sensor data is based on localization and / or tracking and / or fusion of the sensor data. [5] Method according to claim 4, characterized by that fusion filters, static or dynamic environmental occupancy maps or Kalman filters are used for localization and / or tracking and / or fusion. [6] Method according to claim 1, characterized by that the determination of the location-dependent sensor data is carried out by classifying the sensor data with the help of a neural network.

Images