Method for operating an at least partially automated vehicle and vehicle

Abstract

The invention relates to a method for operating an autonomous vehicle (50), wherein sensor data (10) are detected by means of at least one sensor (51), wherein the detected sensor data (10) are reconstructed by means of a reconstruction method (30), wherein the detected sensor data (10) and the reconstructed sensor data (11) are each supplied to a machine-learning-based perception function (2), wherein a distance (32) is determined between outputs (20, 21) respectively generated by means of a distance metric (31), wherein the determined distance (32) is compared with at least one predefined threshold value (25), and wherein at least one safety measure (15) is executed if the determined distance (32) exceeds the at least one predefined threshold value (25). The invention also relates to a vehicle (50) and a method for detecting out-of-sample data of a machine-learning-based perception function (2).

Description

Technical Field

[0001] The invention relates to a method for operating an at least partially automated vehicle and to a vehicle. Background Art

[0002] Machine learning (e.g., based on neural networks) has enormous potential for application in modern driver assistance systems and automated vehicles. Deep neural network-based functions process sensor data (e.g., from cameras, radar sensors, or lidar sensors) to derive relevant information. This information includes, for example, the type and position of objects in the vehicle's environment, their behavior, or the geometry or topology of lanes.

[0003] A key feature in the development (training) of deep neural networks is purely data-driven parameter adaptation without expert intervention: Here, the deviation (the so-called loss) of the output of the neural network (for a given parameterization) from the ground truth is determined. The loss function used is selected in such a way that the parameters of the neural network are differentiably dependent on it. Within the scope of the gradient descent method, the parameters of the neural network are adapted in each training step based on the derivatives of the deviation (determined over multiple examples). These training steps are repeated a large number of times until the loss no longer decreases.

[0004] In this process, the parameters of the neural network are determined without expert evaluation or semantically motivated modeling. However, neural networks also have drawbacks. Thus, attacks based on adversarial perturbations in sensor data / input data can, for example, lead to misclassification and incorrect semantic segmentation, despite the semantically unchanged content of the detected sensor data. Furthermore, the quality of the neural network is consistently high only when the input data originates from the data domain on which the neural network was trained (i.e., in the case of in-sample data). Conversely, if the input data originates from a different data domain (out-of-sample data), the quality of the neural network's output can degrade.

[0005] Particularly within the scope of automated driving, in which high safety requirements are placed, it must be possible to detect out-of-sample data before further processing.

[0006] A quilting method and a total variance minimization method for eliminating adversarial interference in image data are known from Chuan Guo et al., Countering Adversarial Images Using InputTransformations, arXiv:1711.00117v3 [cs.CV], 25. Jan. 2018, https: / / arxiv.org / pdf / 1711.00117.pdf. Summary of the Invention

[0007] The object of the present invention is to provide a method for operating an at least partially automated vehicle and a vehicle in which out-of-sample data, in particular caused by adversarial interference, can be detected and reacted to after detection.

[0008] According to the invention, this object is achieved by a method having the features of patent claim 1 and a carrier having the features of patent claim 9. Advantageous embodiments of the invention are derived from the dependent claims.

[0009] In particular, a method for operating an at least partially automated vehicle is provided, wherein sensor data are detected by means of at least one sensor, wherein the detected sensor data are reconstructed by means of a reconstruction method, wherein the detected sensor data and the reconstructed sensor data are respectively supplied to a perception function based on machine learning, wherein a distance is determined between outputs respectively generated by means of a distance metric, wherein the determined distance is compared with at least one preset threshold value, and wherein at least one safety measure is implemented if the determined distance exceeds at least one preset threshold value.

[0010] In addition, a vehicle is provided in particular, wherein the vehicle travels at least partially automatically, the vehicle comprising: at least one sensor, wherein the at least one sensor is configured to detect sensor data; and a control device, wherein the control device is configured to provide a perception function based on machine learning, reconstruct the detected sensor data by means of a reconstruction method, supply the detected sensor data and the reconstructed sensor data to the perception function, respectively, to determine a distance between outputs respectively generated by means of a distance metric, compare the determined distance with at least one preset threshold value, and perform at least one safety measure if the determined distance exceeds at least one preset threshold value.

[0011] The method and vehicle enable the identification of out-of-sample data and the implementation of at least one safety measure after such identification. To this end, the detected sensor data is reconstructed using a reconstruction method. The detected sensor data and the reconstructed sensor data are then fed to a machine learning-based perception function. If the sensor data is, for example, a detected camera image, and the perception function performs, for example, semantic segmentation, semantic segmentation is performed on the original camera image and on the reconstructed camera image. A distance metric is used to determine the distance between the outputs generated by the perception function, and the determined distance is compared with at least one threshold. If the determined distance exceeds at least one threshold, this indicates the presence of out-of-sample data. This can be caused by a targeted attack in the form of adversarial perturbations contained in the sensor data, or by sensor data that lies outside the data domain, such as so-called "corner cases," i.e., rarely occurring representations of the sensor data or the content depicted therein. In any case, the data lies outside the data domain in which the perception function was trained. If at least one threshold is exceeded, at least one safety measure is implemented to maintain or increase the safety of the vehicle.

[0012] Furthermore, a method for identifying out-of-sample data for a machine learning-based perception function, in particular for at least partially automated vehicles, is proposed. Sensor data detected by means of at least one sensor are reconstructed using a reconstruction method. The detected sensor data and the reconstructed sensor data are each supplied to the perception function. A distance is determined between outputs generated by the perception function using a distance metric. The determined distance is compared with at least one predefined threshold value, and an out-of-sample signal is generated and provided if the determined distance exceeds the at least one predefined threshold value. The method is implemented, in particular, using a data processing device.

[0013] The advantage of this method and vehicle is that adversarial interference can be detected independently of its specific manifestation. This makes it possible to identify and intercept unknown and / or new adversarial interference. This eliminates the need for complex (re)training of machine learning-based perception functions, saving time and costs while maintaining a high level of safety when using perception functions.

[0014] The perception function is particularly already fully trained. It is particularly provided that the (same) perception function is also used during (partially) automated driving of the vehicle. In this case, it can be provided that both outputs generated, in particular inferred, based on detected sensor data of the perception function are taken into account during (partially) automated driving, and outputs generated, in particular inferred, based on reconstructed sensor data are taken into account. It is particularly provided that the perception function provides functions for (partially) automated driving of the vehicle and / or for driver assistance of the vehicle and / or for environmental detection and / or environmental perception. For example, the perception function can perform object recognition and / or semantic segmentation on the detected sensor data.

[0015] The sensor may be, for example, a camera, a stereo camera, a lidar, a radar or an ultrasonic sensor.

[0016] The sensor data of the at least one sensor can in principle be one-dimensional or multi-dimensional, in particular two-dimensional. For example, the sensor data can be a two-dimensional camera image and / or two-dimensional data of a lidar sensor or a radar sensor.

[0017] The reconstruction method reconstructs the detected sensor data based on, among other things, the data domain in which the perception function was trained.

[0018] The distance is determined based on a distance metric. For camera images, the distance metric can be defined, for example, via the Euclidean distance on the image element vector. To this end, the camera image is linearized as a vector. The distance is then determined using the vector norm, for example, the L2 norm. However, other distance metrics can also be used, for example based on other Lp distances or Lp norms. Distance determination for other types of sensor data is generally performed in the same manner.

[0019] At least one threshold value is determined in particular empirically (i.e. in particular with the aid of sensor data prepared in a targeted manner). For example, antagonistic interference can be specifically integrated into the sensor data provided for the test so that the distance resulting from this can be determined by the perception function after processing the prepared sensor data and the reconstructed prepared sensor data. For example, this process can be performed for multiple and / or different antagonistic interferences. Subsequently, a threshold value can be determined based on the distances determined respectively. In principle, this threshold value can be selected based on a determined minimum distance, a weighted average, or other statistical parameters. Furthermore, sensor data that occurs very rarely (so-called "extreme cases") can also be generated in a targeted manner, for example, by simulations of sensor data, so that at least one threshold value can be determined via the distance determined for this. In principle, more than one threshold value can also be determined. This enables the selection of safety measures based on different threshold values, so that, for example, graded safety measures can be implemented in terms of scope.

[0020] The components of the control device can be constructed individually or jointly as a combination of hardware and software, for example as program code executed on a microcontroller or microprocessor. However, it can also be provided that the components are constructed individually or jointly as an application-specific integrated circuit (ASIC).

[0021] The vehicle is in particular a motor vehicle. However, the vehicle can in principle also be another land vehicle, a rail vehicle, an air vehicle, a water vehicle or a space vehicle.

[0022] In particular, it is provided that the method is repeated continuously so that the (current) detected sensor data can be checked continuously. This is carried out in particular in parallel with a regular evaluation of the detected sensor data, which is performed by means of the perception function.

[0023] In one embodiment, the machine learning-based perception function is provided by means of a trained neural network. The neural network is in particular a deep neural network, in particular a convolutional neural network (CNN). The neural network is trained for a perception function, for example, for the perception of pedestrians or other objects detected in a camera image.

[0024] In one embodiment, the reconstruction method includes quilting and / or total variance minimization and / or smoothing and / or denoising and / or compression of the detected sensor data. Such methods are known, for example, from Chuan Guo et al., "Countering Adversarial Images Using Input Transformations," arXiv:1711.00117v3 [cs.CV], 25 Jan. 2018, https: / / arxiv.org / pdf / 1711.00117.pdf.

[0025] Quilting includes, in particular, a one-by-one replacement of the detected sensor data, which is also referred to as one-by-one reconstruction of the sensor data. The sensor data is divided into a plurality of sub-segments for this purpose. In the case of two-dimensional sensor data (for example in the form of camera images), smaller, in particular rectangular sub-segments (also referred to as patches) are defined for this purpose. The individual sub-segments are compared with sub-segments, which are subsequently referred to as sensor data patches (which are stored, for example, in a database). The sensor data patches are generated from the data domain in which the perception function has been trained. The comparison is based on a distance metric, which is defined, for example, by the Euclidean distance on the image element vector. For this purpose, the sub-segments are linearized as vectors. The distance is then determined via the vector norm, for example, via the L2 norm. In order to quilt, the sub-segments are replaced by the closest or most similar sensor data patches in the database. It can be provided that a minimum distance must be maintained, or at least that there must be no similarity between the sub-segments and the sensor data patches in the sensor data. If the sensor data has other forms or other formats, the one-by-one replacement is performed in a similar manner.

[0026] In one embodiment, the perception function is deactivated as a safety measure. This prevents the vehicle from being controlled based on outputs of the perception function during (partially) automated driving, even though these outputs may be untrustworthy or could lead to erroneous control of the vehicle. In particular, this prevents outputs of the perception function that are manipulated by adverse interference in the sensor data from being taken into account during (partially) automated driving.

[0027] In one embodiment, the confidence level of the output of the perception function is varied as a safety measure. This allows, in particular, the influence of the output of the perception function on the automated vehicle to be varied, in particular to be reduced. For example, after exceeding at least one threshold, the output can be considered with less weight in the decision regarding automated driving. If multiple thresholds are provided, the intensity of the consideration can be selected, for example, via the correspondingly selected confidence level and / or the weighting dependent thereon, also depending on the respective threshold. In particular, the confidence level or weighting can be continuously reduced as the determined distance increases.

[0028] In one embodiment, it is provided that, as a safety measure, at least one confidence value of the detected sensor data of at least one sensor is changed and / or that, as a safety measure, the sensor data of at least one sensor is no longer considered by the perception function. If the sensor data of multiple sensors are considered by the perception function, this can influence which sensors are considered by the perception function with what intensity or weight, by selecting the intensity or weight in subsequent processing according to the corresponding confidence level. For example, if the distance determined for the detected camera image exceeds at least one threshold value, while the distance determined for the lidar data does not exceed the relevant threshold value, the confidence value of the camera can be lowered and / or the sensor data of the camera can be (at least temporarily) no longer considered by the perception function. The perception function then operates solely based on the lidar data and, if necessary, the sensor data of further sensors. If the threshold value is again exceeded, the camera (or the corresponding sensor) can be considered again or the confidence level of the relevant sensor data can be increased again.

[0029] It can be provided that the confidence level is also varied here depending on the sensor data and / or the object. For example, if the sensor data is a captured camera image, the confidence values ​​of objects in the camera image and the segmentation map of the reconstructed image generated using the perception function can be compared. The confidence values ​​of the individual outputs of the perception function are determined using methods known per se, such as (Monte Carlo) dropout sampling in a neural network, in which various parts of the neural network are randomly deactivated and different deactivated outputs are generated (inferred) multiple times for the same sensor data. The resulting distribution of the outputs can then be used as a measure of confidence or uncertainty. If the confidence values ​​of the camera image and the reconstructed camera image of the same object deviate significantly from each other, the confidence value of the object or segmentation can be reduced. This process can generate a confidence map or uncertainty map from the correspondingly determined differences, so that subsets of sensor data with different confidence levels can be considered with varying degrees in subsequent processing, for example by the vehicle control system.

[0030] In one embodiment, as a safety measure, the configuration of at least one sensor is modified. This allows, for example, the use of a different number of sensors and / or a different combination of sensors. Furthermore, sensor parameters can be modified. For example, the resolution of a camera or lidar sensor can be changed.

[0031] In one embodiment, it is provided that, as a safety measure, at least one fallback strategy is activated for the vehicle for automated driving. The fallback strategy comprises, in particular, a reconfiguration of the environmental perception and / or the control of the (partially) automated vehicle. For example, the fallback strategy may comprise a fallback to a lower level of automation. The fallback strategy may also comprise the activation and / or deactivation and / or reconfiguration of at least one sensor of the vehicle and / or a system for (partially) automated driving. For example, less precise sensors and / or less precisely operating systems may be used. In any case, the aim is to maintain a predetermined safety measure in the current situation by the fallback strategy, or to increase safety.

[0032] In one embodiment, as a safety measure, the autonomous vehicle is moved to a safe state. This safe state may include a human driver taking over control. Furthermore, it may be provided that, to achieve the safe state, the vehicle is autonomously driven to the side of the road and parked there.

[0033] Furthermore, it can be provided that, after at least one threshold value is exceeded, a warning message is generated and output and / or transmitted to a recipient. This allows the presence of out-of-sample data to be brought to the attention of vehicle passengers and / or vehicle manufacturers and / or third-party service providers. In particular, it is possible for vehicle manufacturers and / or third-party service providers to adapt or update, in particular retrain, the perception function based on the transmitted warning message.

[0034] The further features of the embodiment of the carrier are derived from the description of the embodiment of the method. The advantages of the device are the same as those in the embodiment of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention will be explained in more detail below with reference to the accompanying drawings using preferred embodiments.

[0036] Figure 1 A schematic diagram showing an embodiment of a carrier is shown;

[0037] Figure 2 A schematic flow chart illustrating a method for detecting out-of-sample data is shown. DETAILED DESCRIPTION

[0038] Figure 1 shows a schematic diagram of an embodiment of a vehicle 50. Vehicle 50 includes a sensor 51 (e.g., a camera) with which the surroundings of vehicle 50 are detected. Vehicle 50 also includes a control device 1. Control device 1 is designed, for example, as a combination of hardware and software, for example, as program code executed on a microcontroller or microprocessor.

[0039] Control device 1 provides a machine learning-based perception function 2. Perception function 2 is provided, in particular, in the form of a trained deep neural network 3. Perception function 2, for example, recognizes objects in detected sensor data 10 or performs semantic segmentation, and provides an output 20, which is then supplied to vehicle control 52, which also controls or regulates actuators 53 of vehicle 50 at least as a function of output 20. Furthermore, control device 1 implements the method described in this disclosure.

[0040] exist Figure 2 In the embodiment of the present disclosure, the core elements of the method described in the present disclosure, namely the identification of out-of-sample data, are illustrated in the form of a schematic flow chart. For example, it is assumed that the detected sensor data 10 comprises a camera image.

[0041] The detected sensor data 10 are reconstructed with the aid of a reconstruction method 30 (for example a quilting method), so that reconstructed sensor data 11 can be provided.

[0042] Furthermore, it may be provided that reconstruction method 30 additionally or alternatively includes minimizing the total variance and / or smoothing and / or denoising and / or compressing detected sensor data 10 .

[0043] The detected sensor data 10 and the reconstructed sensor data 11 are each supplied to a trained perception function 2, in particular a trained deep neural network 3. The trained perception function 2 generates, in particular infers, an output 20, 21 for both the detected sensor data 10 and the reconstructed sensor data 11.

[0044] A distance 32 is determined between the outputs 20, 21 using a distance metric 31. For example, the Lp distance or the Lp norm, for example the L2 norm, can be used as the distance metric 31.

[0045] In the checking step 40, the determined distance 32 is compared with a predefined threshold value 25. If the checking shows that the threshold value 25 is not exceeded, an in-sample signal 60 is generated. On the other hand, if the checking shows that the threshold value 25 is exceeded, an out-of-sample signal 61 is generated, which is transmitted, for example, to the vehicle control device 52 (see Figure 1 ).

[0046] After the transmission of the out-of-sample signal 61, at least one safety measure 15 is implemented, for example by means of the vehicle control device 52 ( Figure 1 ).

[0047] It can be provided that, as a safety measure 15 , the sensing function 2 is deactivated.

[0048] Provision may be made, as a safety measure 15 , to change the confidence level of the output 20 of the perception function 2 .

[0049] It can further be provided that as a safety measure 15 at least one confidence value of the detected sensor data 10 of at least one sensor 51 is changed and / or as a safety measure 15 sensor data 10 of at least one sensor 51 are no longer taken into account by the perception function 2 .

[0050] Provision may be made to change at least one sensor configuration as a safety measure 15. For example, other combinations of sensors 51 (camera, lidar, radar, ultrasound, etc.) may be used and / or parameters of the sensors 51 may be changed.

[0051] Provision may be made to activate at least one fallback strategy for the (partially) automated driving of the vehicle 50 as a safety measure 15 . This occurs, for example, after the out-of-sample signal 61 has been received by the vehicle control unit 52 .

[0052] It can be provided that the vehicle 50 is transferred to a safe state as a safety measure 15. To this end, the vehicle control device 52 controls the vehicle 50 or the actuators 53 of the vehicle 50, for example, so that the vehicle 50 stops at the roadside, or the vehicle control device 52 reduces the degree of automation and / or deactivates certain auxiliary functions of the vehicle 50 that are dependent on the out-of-sample data or the associated sensors 51.

[0053] It can be provided that only a single safety measure 15 is carried out. However, it can also be provided that a plurality of safety measures 15 are carried out.

[0054] Reference Signs List

[0055] 1 Control device

[0056] 2 Perception function

[0057] 3 Deep Learning Networks

[0058] 10 Sensor Data

[0059] 11 Reconstructed sensor data

[0060] 15 Safety Measures

[0061] 20 Output (sensor data)

[0062] 21 Output (reconstructed sensor data)

[0063] 25 threshold

[0064] 30 Reconstruction Methods

[0065] 31 Distance Metrics

[0066] 32 Distance

[0067] 40 Inspection Steps

[0068] 50 Vehicles

[0069] 51 sensors

[0070] 52 Vehicle Control Device

[0071] 53 Actuator

[0072] 60 In-sample signal

[0073] 61 Out-of-sample signal.

Claims

1. A method for operating an at least partially automated vehicle (50), in, detecting sensor data (10) by means of at least one sensor (51), wherein the detected sensor data (10) are reconstructed by means of a reconstruction method (30), wherein the detected sensor data (10) and the reconstructed sensor data (11) are supplied to a machine learning-based perception function (2), respectively, wherein a distance (32) is determined between the outputs (20, 21) respectively generated by means of the perception function (2) by means of a distance metric (31), wherein the determined distance (32) is compared with at least one predefined threshold value (25), and wherein, if the determined distance (32) exceeds the at least one predetermined threshold value (25), at least one safety measure (15) is executed, A plurality of different threshold values ​​(25) are preset, and a safety measure (15) is selected according to the different threshold values ​​(25).

2. The method according to claim 1, characterized in that The reconstruction method (30) comprises quilting and / or total variance minimization and / or smoothing and / or denoising and / or compression of the detected sensor data (10).

3. The method according to any one of the preceding claims, characterized in that As a safety measure (15), the sensing function (2) is deactivated.

4. The method according to any one of the preceding claims, characterized in that As a safety measure (15), the confidence level of the output (20) of the sensing function (2) is changed.

5. The method according to any one of the preceding claims, characterized in that As a safety measure (15), at least one confidence value of the detected sensor data (10) of at least one sensor (51) is changed and / or as a safety measure (15), the sensor data (10) of at least one sensor (51) is no longer taken into account by the perception function (2).

6. The method according to any one of the preceding claims, characterized in that As a safety measure (15), at least one sensor configuration is changed.

7. The method according to any one of the preceding claims, characterized in that As a safety measure (15), at least one rollback strategy for the automated vehicle (50) is activated.

8. The method according to any one of the preceding claims, characterized in that As a safety measure (15), the automated vehicle (50) is transferred to a safe state.

9. A carrier (50), wherein: The vehicle (50) is at least partially automated and comprises: at least one sensor (51), wherein the at least one sensor (51) is configured to detect sensor data (10), and A control device (1), wherein the control device (1) is configured to provide a machine learning-based perception function (2), reconstruct the detected sensor data (10) by means of a reconstruction method (30), supply the detected sensor data (10) and the reconstructed sensor data (11) to the perception function (2), determine a distance (32) between outputs (20, 21) respectively generated by means of the perception function (2) by means of a distance metric (31), compare the determined distance (32) with at least one predefined threshold value (25), and execute at least one safety measure (15) if the determined distance (32) exceeds the at least one predefined threshold value (25), A plurality of different threshold values ​​(25) are preset, and a safety measure (15) is selected according to the different threshold values ​​(25).

10. A method for identifying out-of-sample data of perceptual functions (2) based on machine learning, in, Reconstructing sensor data (10) detected by means of at least one sensor (51) by means of a reconstruction method (30), wherein the detected sensor data (10) and the reconstructed sensor data (11) are supplied to the perception function (2), respectively, wherein a distance (32) is determined between the outputs (20, 21) respectively generated by means of the perception function (2) by means of a distance metric (31), wherein the determined distance (32) is compared with at least one predefined threshold value (25), and wherein, if the determined distance (32) exceeds the at least one predetermined threshold value (25), an out-of-sample signal (61) is generated and provided, There are multiple different thresholds (25) set.

Images