Method for operating an at least partially automated vehicle and vehicle

Abstract

Method for operating an at least partially automated vehicle (50), wherein sensor data (10) are acquired by means of at least one sensor (51), wherein the acquired sensor data (10) are reconstructed by means of a reconstruction method (30), wherein the acquired sensor data (10) and the reconstructed sensor data (11) are each fed to a machine learning-supported perception function (2), wherein a distance (32) is determined between outputs (20, 21) which are each generated by means of the perception function (2) by means of a distance measure (31), wherein the determined distance (32) is compared with at least one predetermined threshold value (25), and wherein at least one safety measure (15) is carried out if the determined distance (32) exceeds the at least one predetermined threshold value (25).

Description

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

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

[0003] A key feature in the development of deep neural networks (training) is purely data-driven parameter fitting without expert intervention: This involves determining the deviation of a neural network's output (for a given parameterization) from a ground truth (the so-called loss). The loss function used here is chosen such that the neural network's parameters depend on it in a differentiable manner. In the gradient descent method, the neural network's parameters are adjusted in each training step depending on the derivative of the deviation (determined from several examples). These training steps are repeated many times until the loss no longer decreases.

[0004] With this approach, the parameters of the neural network are determined without expert assessment or semantically motivated modeling. However, neural networks also have disadvantages. For example, attacks based on adversarial disturbances in the sensor data / input data can lead to misclassification or incorrect semantic segmentation despite the semantically unchanged content of the acquired sensor data. Furthermore, the quality of a neural network is only ever high if the input data originates from the data domain on which the neural network was trained (i.e., in-sample data). If, however, the input data originates from a different data domain (out-of-sample data), the quality of the neural network's output can decrease.

[0005] Particularly in the field of automated driving, where high safety requirements are required, out-of-sample data must be able to be recognized before further processing takes place.

[0006] From the subsequently published DE 10 2018 222 294 A1, a method for data prediction based on an input data item by means of a machine learning method is known, wherein, in parallel to the data prediction, the input data item is mapped to an encoder data item by means of a variational autoencoder, wherein a distance criterion is determined as a function of the encoder data item and an expected distribution, wherein the distance criterion represents a plausibility measure for the data prediction by means of the machine learning method.

[0007] From the subsequently published DE 10 2019 205 085 A1 a method is known for monitoring first input data of an artificial intelligence function which provides a core functionality and was trained with a training data set, by means of a monitoring function which has a reconstruction function which was trained with the training data set, wherein the artificial intelligence function and the monitoring function each receive the same first input data, wherein the reconstruction function maps the first input data to condensed input data which have a smaller dimension than the first input data and reconstructs second input data from the condensed input data, and wherein the monitoring function determines a reconstruction error between the first input data and the second input data.

[0008] From the subsequently published DE 10 2018 218 834 A1, a computer-implemented method for classifying an input signal is known, which is determined in particular as a function of an output signal of a sensor, as to whether or not it has an anomaly, wherein a decision is made as to whether or not the input signal has the anomaly depending on an output signal of an autoencoder to which the input signal is fed, wherein the autoencoder has at least one encoder and at least one decoder, wherein an intermediate variable is determined by means of the encoder as a function of the input signal, and the output signal is determined as a function of this intermediate variable by means of the decoder, wherein the autoencoder provides a plurality of hypotheses for reconstructing the input signal for the input signal fed to it and determines the output signal as a function of this plurality of hypotheses.

[0009] DE 10 2017 116 196 A1 discloses methods and systems for detecting faults in a sensor and for reconstructing an output signal without using the faulty sensor. The method includes: receiving, via a processor, sensor data indicating a measured value of a first sensor; receiving, via a processor, sensor data indicating measured values ​​of a plurality of other sensors; calculating, via a processor, virtual values ​​based on a vehicle model and the sensor data from the plurality of other sensors; calculating, via a processor, a remaining difference between the measured value of the first sensor and the virtual values; detecting, via a processor, whether a fault exists in the first sensor based on the remaining difference; and if a fault of the sensor is detected, generating, via a processor, a control value based on the virtual values ​​instead of the measured value.

[0010] DE 10 2014 118 479 A1 discloses an affective user interface in an autonomous vehicle. It is provided that data relating to autonomous operation of a vehicle can be collected. A first confidence assessment can then be generated regarding whether the vehicle should be operated autonomously, and it can be determined whether the first confidence assessment reaches or exceeds a predetermined threshold. A user interface in the vehicle can provide an indication regarding the termination of autonomous operation of the vehicle if the first confidence assessment reaches or exceeds the predetermined threshold.

[0011] DE 10 2018 200 134 B3 discloses a method for acquiring training data for a driver assistance system for a motor vehicle. A server device specifies a criterion and transmits it to the motor vehicles. If the criterion is met, the vehicles then transmit individual images of the surroundings of the respective motor vehicle captured by a respective on-board camera to the server device. The server device then generates a common image series for training the driver assistance system from the individual images transmitted by various motor vehicles.

[0012] 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, a quilting method and a total variance minimization method for removing adversarial noise in image data are known.

[0013] The invention is based on the object of creating a method for operating an at least partially automated vehicle and a vehicle in which out-of-sample data, in particular caused by adversarial disturbances, can be detected and, after detection, can be reacted to.

[0014] The object is achieved according to the invention by a method having the features of patent claim 1 and a vehicle having the features of patent claim 9. Advantageous embodiments of the invention emerge from the subclaims.

[0015] In particular, a method for operating an at least partially automated vehicle is provided, wherein sensor data are acquired by means of at least one sensor, wherein the acquired sensor data are reconstructed by means of a reconstruction method, wherein the acquired sensor data and the reconstructed sensor data are each fed to a machine learning-supported perception function, wherein a distance is determined between outputs which are each generated by means of the perception function by means of a distance measure, wherein the determined distance is compared with at least one predetermined threshold value, and wherein at least one safety measure is carried out if the determined distance exceeds the at least one predetermined threshold value.

[0016] Furthermore, in particular, a vehicle is created, wherein the vehicle is driven at least partially automatically, comprising at least one sensor, wherein the at least one sensor is configured to capture sensor data; and a control device, wherein the control device is configured to provide a machine learning-supported perception function, to reconstruct the captured sensor data by means of a reconstruction method, to feed the captured sensor data and the reconstructed sensor data respectively to the perception function, to determine a distance between outputs each generated by means of the perception function by means of a distance measure, to compare the determined distance with at least one predetermined threshold value, and to carry out at least one safety measure if the determined distance exceeds the at least one predetermined threshold value.

[0017] The method and the vehicle make it possible to detect out-of-sample data and, following detection, to initiate at least one safety measure. For this purpose, acquired sensor data is reconstructed using a reconstruction process. The acquired sensor data and the reconstructed sensor data are then each fed to the machine-learning-supported perception function. For example, if the sensor data are acquired camera images and the perception function performs semantic segmentation, semantic segmentation is performed once on the original camera image and once on a reconstructed camera image. A distance measure is used to determine the distance between the outputs generated by the perception function. The determined distance is compared with at least one threshold value.If the specified distance exceeds at least one threshold value, this is a sign that out-of-sample data is present. The cause can be either a targeted attack in the form of an adversarial interference contained in the sensor data or sensor data located outside the data domain, for example in the form of so-called "corner cases," i.e., very rarely occurring characteristics of the sensor data or a content represented therein. In any case, the data lies outside the data domain in which the perception function was trained. If at least one threshold value is exceeded, at least one safety measure is implemented to maintain or increase vehicle safety.

[0018] Furthermore, a method for detecting out-of-sample data of a machine-learning-supported perception function, in particular for an at least partially automated vehicle, is also proposed, wherein sensor data acquired by means of at least one sensor are reconstructed by means of a reconstruction method, wherein the acquired sensor data and the reconstructed sensor data are each fed to the perception function, wherein a distance is determined between outputs each generated by means of the perception function by means of a distance measure, wherein the determined distance is compared with at least one predetermined threshold value, and wherein an out-of-sample signal is generated and provided if the determined distance exceeds the at least one predetermined threshold value. The method is carried out in particular by means of a data processing device.

[0019] One advantage of the method and the vehicle is that the detection of adversarial interference can occur regardless of the specific nature of the adversarial interference. This allows even unknown and / or novel adversarial interference to be detected and intercepted. This eliminates the need for (re)training the machine-learning-based perception function, saving time and money while still maintaining a high degree of safety when using the perception function.

[0020] In particular, the perception function is already fully trained. In particular, it is envisaged that (the same) perception function is also used during (partially) automated driving of the vehicle. In this case, it can be provided that an output of the perception function generated, in particular inferred, on the basis of the recorded sensor data is taken into account during (partially) automated driving, as well as that an output of the perception function generated, in particular inferred, on the basis of the reconstructed sensor data is taken into account. In particular, it is envisaged that the perception function provides a function for the (partially) automated driving of the vehicle and / or for driver assistance of the vehicle and / or for environmental detection and / or environmental perception. The perception function can, for example, perform object recognition and / or semantic segmentation on the recorded sensor data.

[0021] A sensor can be, for example, a camera, a stereo camera, a lidar, a radar or an ultrasonic sensor.

[0022] 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 two-dimensional camera images from a camera and / or two-dimensional data from a lidar or radar sensor.

[0023] The reconstruction method reconstructs the acquired sensor data, particularly based on the data domain in which the perception function was trained.

[0024] The distance is determined based on a distance measure. For camera images, the distance measure can be defined, for example, using a Euclidean distance on pixel vectors. For this purpose, a camera image is linearized as a vector. The distance is then determined using a vector norm, such as the L2 norm. In principle, however, other distance measures can also be used, for example, based on other Lp distances or Lp norms. For other types of sensor data, the distance is determined in the same way.

[0025] The at least one threshold value is determined empirically, i.e., in particular with the help of specifically prepared sensor data. For example, adversarial disturbances can be specifically integrated into the sensor data intended for testing in order to determine a resulting distance after the perception function has processed the prepared sensor data and reconstructed prepared sensor data. This procedure can, for example, be performed for several and / or different adversarial disturbances. A threshold value can then be set based on the respective determined distances. In principle, this threshold value can be selected based on the smallest determined distance, a weighted average, or other statistical variables, etc.Furthermore, rarely occurring sensor data, so-called "corner cases," can be specifically generated, for example, by simulating sensor data, to determine at least one threshold value over a specific distance. In principle, more than one threshold value can also be determined. This allows security measures to be selected depending on different threshold values, so that, for example, graduated security measures can be implemented with respect to a specific scope.

[0026] Parts of the control device can be implemented individually or collectively 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 parts are implemented individually or collectively as an application-specific integrated circuit (ASIC).

[0027] A vehicle is, in particular, a motor vehicle. However, in principle, the vehicle can also be any other land, rail, air, water, or space vehicle.

[0028] In particular, it is intended that the process be repeated continuously so that (current) acquired sensor data can be continuously reviewed. This occurs, in particular, in parallel with a regular evaluation of the acquired sensor data, which is performed using the perception function.

[0029] In one embodiment, the machine-learning-supported perception function is provided by a trained neural network. A neural network is, in particular, a deep neural network, in particular a convolutional neural network (CNN). The neural network is trained for the perception function, for example, for the perception of pedestrians or other objects in captured camera images.

[0030] In one embodiment, the reconstruction method comprises quilting and / or total variance minimization and / or smoothing and / or denoising and / or compression of the acquired sensor data. Such methods are known, for example, from Chuan Guo et al., Countering Adversarial Images Using Input Transformations, arXiv:1711.00117v3 [cs.CV], January 25, 2018, https: / / arxiv.org / pdf / 1711.00117.pdf.

[0031] Quilting particularly involves the piecewise replacement of the acquired sensor data, which is also referred to as piecewise reconstruction of the sensor data. For this purpose, the sensor data is divided into several sub-sections. In the case of two-dimensional sensor data (e.g. in the form of camera images), small, particularly rectangular sub-sections (also referred to as patches) are defined. The individual sub-sections are compared with sub-sections, referred to below as sensor data patches, which are stored, for example, in a database. The sensor data patches were generated from the data domain in which the perception function was trained. The comparison is based on a distance measure, which is defined, for example, via a Euclidean distance on image element vectors. For this purpose, a sub-section is linearized as a vector.The distance is then determined using a vector norm, for example, the L2 norm. For quilting, the partial sections are replaced by the nearest or most similar sensor data patch from the database. It can be stipulated that a minimum distance must be maintained or, at least, that there must be no identity between the partial section of the sensor data and the sensor data patch. If the sensor data has a different form or format, the piecewise replacement is performed in a similar manner.

[0032] In one embodiment, the perception function is deactivated as a safety measure. This can prevent the (partially) automated vehicle from being controlled based on the outputs of the perception function, even though these outputs may not be trustworthy or could lead to incorrect vehicle control. In particular, this can prevent an output of the perception function that has been manipulated by an adversarial disturbance in the sensor data from being taken into account during (partially) automated driving.

[0033] In one embodiment, it is provided that a confidence of an output of the perception function is changed as a safety measure. In particular, this makes it possible to change, in particular reduce, the influence of the output of the perception function on a control of the automated driving vehicle. For example, after the at least one threshold value has been exceeded, the output can thereby be taken into account with a lower weighting in decision-making for automated driving. If several threshold values ​​are provided, the strength of consideration can also be selected depending on the respective threshold value, for example via an appropriately selected confidence and / or a weighting dependent thereon. In particular, as the specific distance increases, a confidence or a weighting can be successively reduced.

[0034] In one embodiment, it is provided that, as a security measure, at least one confidence value of acquired sensor data from at least one sensor is changed and / or, as a security measure, sensor data from at least one sensor is no longer considered by the perception function. If the perception function considers sensor data from several sensors, this can influence which sensors are considered by the perception function and with what strength or weighting by selecting the strength or weighting during subsequent processing depending on the respective confidence.For example, if a distance determined for captured camera images exceeds at least one threshold, but a distance determined for lidar data does not exceed the corresponding threshold, a camera's confidence value can be reduced and / or the camera's sensor data can no longer be considered by the perception function (at least temporarily). The perception function then operates (at least temporarily) only on the basis of the lidar data and sensor data from any additional sensors present. If the threshold is again undershot, the camera (or the corresponding sensor) can be considered again, or the confidence of the associated sensor data can be increased again.

[0035] It can be provided that confidence is also changed depending on the sensor data and / or the object. For example, if the sensor data consists of captured camera images, confidence values ​​for objects in a segmentation map created using the perception function for a camera image and a reconstructed camera image can be compared with each other. The confidence values ​​for the individual outputs of the perception function are determined using known methods; in the case of a neural network, for example, using (Monte Carlo) drop-out sampling, in which individual parts of the neural network are randomly deactivated and an output for different deactivations is generated (inferred) multiple times for the same sensor data. The resulting distribution of the output can then be used as a measure of confidence or uncertainty.If the confidence values ​​for the camera image and the reconstructed camera image for the same object differ significantly, a confidence value for this object or segmentation can be reduced. Using this approach, a confidence map or uncertainty map can be created from the determined difference, allowing subsets of the sensor data with different confidence levels to be considered to varying degrees in subsequent processing, for example, by a vehicle control system.

[0036] One embodiment provides for at least one sensor configuration to be changed as a security measure. This allows, for example, a different number of sensors and / or a different combination of sensors to be used. Furthermore, sensor parameters can also be changed. For example, the resolution of a camera or a lidar sensor can be changed.

[0037] In one embodiment, at least one fallback strategy is activated for the automated vehicle as a safety measure. A fallback strategy particularly comprises reconfiguring an environmental perception and / or a control system of the (partially) automated vehicle. A fallback strategy may, for example, comprise falling back to a lower level of automation. A fallback strategy may also comprise activating and / or deactivating and / or reconfiguring at least one sensor and / or a system of the vehicle used for (partially) automated driving. For example, less precise sensors and / or less precise systems may be used. The goal in each case is to maintain the level of safety specified by the fallback strategy in a current situation or to increase safety.

[0038] In one embodiment, it is provided that the automated vehicle is transferred to a safe state as a safety measure. Such a safe state can involve a human driver taking over control. Furthermore, it can also be provided that, to reach the safe state, the vehicle is automatically driven to the side of the road and stopped there.

[0039] It can further be provided that, after the at least one threshold value is exceeded, a warning message is generated and issued and / or transmitted to a recipient. This makes it possible to alert vehicle passengers and / or a vehicle manufacturer and / or third-party service provider to the presence of out-of-sample data. In particular, it is possible for a vehicle manufacturer and / or the third-party service provider to adapt or update the perception function, in particular to retrain it, based on the transmitted warning message.

[0040] Further features of the vehicle design will become apparent from the description of embodiments of the method. The advantages of the device are the same as those of the embodiments of the method.

[0041] The invention will be explained in more detail below using preferred embodiments with reference to the figures. Fig. 1 a schematic representation of an embodiment of the vehicle; Fig. 2 a schematic flow diagram illustrating the procedure for detecting out-of-sample data.

[0042] In Fig. Figure 1 shows a schematic representation of an embodiment of the vehicle 50. The vehicle 50 includes a sensor 51, for example, a camera, with which the surroundings of the vehicle 50 are detected. Furthermore, the vehicle 50 includes a control device 1. The control device 1 is embodied, for example, as a combination of hardware and software, for example, as program code executed on a microcontroller or microprocessor.

[0043] The control device 1 provides a machine-learning-supported perception function 2. The perception function 2 is provided in particular in the form of a trained deep neural network 3. The perception function 2, for example, recognizes objects in the acquired sensor data 10 or performs semantic segmentation, etc., and provides an output 20, which is then fed to a vehicle controller 52, which controls or regulates an actuator system 53 of the vehicle 50, at least also depending on the output 20. Furthermore, the control device 1 executes the method described in this disclosure.

[0044] In Fig. Figure 2 illustrates a core element of an embodiment of the method described in this disclosure in the form of a schematic flowchart, namely the detection of out-of-sample data. For example, it is assumed that the acquired sensor data 10 comprises a camera image.

[0045] The acquired sensor data 10 are reconstructed by means of a reconstruction method 30, for example a quilting method, so that reconstructed sensor data 11 can be provided.

[0046] Furthermore, it can also be provided that the reconstruction method 30 additionally or alternatively comprises a total variance minimization and / or a smoothing and / or a denoising and / or a compression of the acquired sensor data 10.

[0047] The acquired sensor data 10 and the reconstructed sensor data 11 are each fed to the trained perception function 2, in particular to the trained deep neural network 3. The trained perception function 2 generates, in particular infers, an output 20, 21 for both the acquired sensor data 10 and the reconstructed sensor data 11.

[0048] A distance 32 is determined between the outputs 20, 21 using a distance measure 31. For example, an Lp distance or an Lp norm, such as the L2 norm, can be used as the distance measure 31.

[0049] In a verification step 40, the determined distance 32 is compared with a predetermined threshold value 25. If the verification shows that the threshold value 25 is not exceeded, an in-sample signal 60 is generated. If, however, the verification shows that the threshold value 25 is exceeded, an out-of-sample signal 61 is generated, which is sent, for example, to the vehicle control system 52 (see FIG. Fig. 1) is transmitted.

[0050] After the transmission of the out-of-sample signal 61, at least one safety measure 15 is executed, for example by means of the vehicle control 52 ( Fig. 1).

[0051] It may be provided that the perception function 2 is deactivated as a safety measure 15.

[0052] It may be provided that as a security measure 15 a confidence of an output 20 of the perception function 2 is changed.

[0053] Furthermore, it can be provided that as a safety measure 15 at least one confidence value of recorded 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.

[0054] It may be provided that at least one sensor configuration is changed as a safety measure 15. For example, other combinations of sensors 51 (camera, lidar, radar, ultrasound, etc.) may be used and / or parameters of sensors 51 may be changed.

[0055] It may be provided that at least one fallback strategy is activated for the (partially) automated vehicle 50 as a safety measure 15. This occurs, for example, after the out-of-sample signal 61 has been received by the vehicle controller 52.

[0056] It may be provided that the vehicle 50 is transferred to a safe state as a safety measure 15. For this purpose, the vehicle control system 52 controls the vehicle 50 or an actuator 53 of the vehicle 50, for example, such that the vehicle 50 comes to a stop at the side of the road, or the vehicle control system 52 reduces the degree of automation and / or deactivates certain assistance functions of the vehicle 50 that are affected by the out-of-sample data or the associated sensor 51.

[0057] It may be provided that only individual security measures 15 are implemented. However, it may also be provided that several security measures 15 are implemented. List of reference symbols 1 control device 2 Perceptual function 3 deep neural network 10 Sensor data 11 reconstructed sensor data 15 Security measure 20 Output (Sensor Data) 21 Output (reconstructed sensor data) 25 Threshold 30 reconstruction procedures 31 Distance measure 32 Distance 40 verification step 50 vehicles 51 Sensor 52 Vehicle control 53 Actuators 60 In-Sample Signal 61 Out-of-sample signal

Claims

[1] Method for operating an at least partially automated vehicle (50), wherein sensor data (10) are acquired by means of at least one sensor (51), wherein the acquired sensor data (10) are reconstructed by means of a reconstruction method (30), wherein the acquired sensor data (10) and the reconstructed sensor data (11) are each fed to a machine learning-supported perception function (2), wherein a distance (32) is determined between outputs (20, 21) which are each generated by means of the perception function (2) by means of a distance measure (31), wherein the determined distance (32) is compared with at least one predetermined threshold value (25), and wherein at least one safety measure (15) is carried out if the determined distance (32) exceeds the at least one predetermined threshold value (25). [2] Method according to claim 1, characterized bythat the reconstruction method (30) comprises quilting and / or total variance minimization and / or smoothing and / or denoising and / or compression of the acquired sensor data (10). [3] Method according to one of the preceding claims, characterized by that as a safety measure (15) the perception function (2) is deactivated. [4] Method according to one of the preceding claims, characterized by that as a security measure (15) a confidence of an output (20) of the perception function (2) is changed. [5] Method according to one of the preceding claims, characterized by that as a safety measure (15) at least one confidence value of recorded 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). [6] Method according to one of the preceding claims, characterized by that as a safety measure (15) at least one sensor configuration is changed. [7] Method according to one of the preceding claims, characterized by that as a safety measure (15) at least one fallback strategy is activated for the automated vehicle (50). [8] Method according to one of the preceding claims, characterized by that the automated driving vehicle (50) is transferred to a safe state as a safety measure (15). [9] Vehicle (50), wherein the vehicle (50) is driven at least partially automatically, comprising: 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-supported perception function (2), to reconstruct the acquired sensor data (10) by means of a reconstruction method (30), to feed the acquired sensor data (10) and the reconstructed sensor data (11) to the perception function (2), to determine a distance (32) between outputs (20, 21) which are each generated by means of the perception function (2) by means of a distance measure (31), to compare the determined distance (32) with at least one predetermined threshold value (25), and to carry out at least one safety measure (15) if the determined distance (32) exceeds the at least one predetermined threshold value (25). [10] Method for detecting out-of-sample data of a machine learning-based perception function (2), wherein sensor data (10) acquired by at least one sensor (51) are reconstructed by means of a reconstruction method (30), wherein the acquired sensor data (10) and the reconstructed sensor data (11) are each fed to the perception function (2), wherein a distance (32) is determined between outputs (20, 21) which are each generated by means of the perception function (2) by means of a distance measure (31), wherein the determined distance (32) is compared with at least one predetermined threshold value (25), and wherein an out-of-sample signal (61) is generated and provided when the determined distance (32) exceeds the at least one predetermined threshold value (25).

Images