Procedure for verifying the reliability of AI-based object detection

Abstract

A method is described for validating the function of a trained neural network (TNN) for the detection, segmentation, and classification of objects in the environment (U) of a transport vehicle (61). In this method, sensor data (SD) are acquired from the environment (U) of the transport vehicle (61), and input data (ED) for the TNN are generated from the sensor data (SD). Additionally, feature data (MD) of static objects are determined from the input data (SD) or from the intermediate layer data (ZSD) of the TNN using an auxiliary prediction system (APS). The determined feature data (MD) is compared with field comparison data (FVD), which comprises features of static objects in the environment (U). The reliability of the trained neural network (TNN) is determined from the result of this comparison (VE). A method for generating field comparison data (FVD) is also described.A validation device is also described (50). In addition, a means of transport is described (61).

Description

Procedure for verifying the reliability of AI-based object detection The invention relates to a method for validating the operation of an AI-based system for detecting, segmenting, and classifying objects in the environment of a means of transport. The invention also relates to a method for generating field comparison data. Furthermore, the invention relates to a validation system. Finally, the invention relates to a means of transport. In autonomous or semi-autonomous driving, it is necessary to monitor the vehicle's surroundings to avoid collisions. In the operation of railway vehicles, for example, it occasionally happens that objects, such as people or road vehicles, shopping carts thrown onto the tracks, or even fallen rocks or trees, enter the railway line and thus pose a danger to the safety of rail traffic. In the case of people and road vehicles, they themselves are in serious danger due to the possibility of a collision with a moving railway vehicle. Therefore, it is necessary to detect such objects in time to initiate a braking maneuver on an approaching railway vehicle, so that a collision between the railway vehicle and the detected objects can be avoided. Therefore, detecting potential obstacles of any shape or configuration that could block the tracks is a fundamental requirement for the safety of all types of rail vehicles. Many solutions exist for this problem, with those based on artificial intelligence (AI) being the most promising. However, any AI-based solution is only as good as its training database, the collection and annotation of which are very laborious in rail transport scenarios. In recent years, computer vision approaches based on deep learning have become highly effective compared to traditional approaches, which rely on manual feature analysis. This makes them very promising for addressing the challenges of automated or driverless rail vehicle operation. For complex scenarios requiring particularly sophisticated environmental perception, such as autonomous rail vehicle operation, it is likely that these tasks cannot be accomplished without deep learning-based methods. However, deep learning methods are not entirely error-free and are therefore not completely reliable. One problem is that neural networks exhibit a kind of overconfidence in data that was not part of a training dataset.In this context, the distribution of the data obtained in reality is said to be outside the distribution of the training data. This data is also called OOD data (OOD = out of distribution). Consequently, a confidence measure cannot be used as an indicator of the reliability of a dataset of results from an artificial neural network. In particular, the operation of highly automated, driverless vehicles, such as road or rail vehicles, requires greater sensitivity to the appearance of data outside the training dataset for broad application, so that the system can enter a safe state in such cases. Approaches based on machine learning or deep learning typically present the following shortcomings today: - Typical artificial neural networks do not have inherent mechanisms to determine whether a data point from the input data is included in the training data or not. - Artificial neural networks tend to have a lower success rate and accuracy with OOD data, even when they achieve a high level of confidence in the prediction. Therefore, it is difficult to determine whether a prediction or decision from an artificial neural network is reliable and trustworthy or not. There are several phenomena that can cause the input data to fall outside the distribution of the training data: Such deviation can be caused by a variation in the detected object, for example, by a change in the type of object, by the movement of an object which, for example, causes different postures of pedestrians or vehicles, or by different sizes of objects. A deviation can also be due to a change in the environment or the detection system. For example, the weather can change: it may rain, snow, or be foggy. Lighting conditions can vary, glare can occur, reflections can appear on objects, lenses can become dirty, and sensors can be affected. All these changes can lead to accurate but inaccurate predictions. This can cause, for example, a misinterpretation of the signal status of a traffic light. Publication US 2021 / 0004017 A1 uses high-resolution cartographic data to generate synthetic sensor data intended for autonomous driving vehicles. In publication US 2020 / 0410254 A1, crossing zones in the vicinity of a vehicle are detected and located in real time or near real time using sensors and an artificial neural network. A verification of the performance of an AI-based system for detecting objects around a vehicle is carried out using a fixed AI-based system with the same functionality. Therefore, the objective is to improve the reliability of AI-based sensor data assessment, particularly in the field of autonomous or assisted driving in rail transport, and thereby increase road safety in the automated or semi-automated operation of such systems. This objective is achieved by means of a procedure for validating the operation of an AI-based system for detecting and classifying objects in the environment of a means of transport according to claim 1, a validation system according to claim 9, and a means of transport according to claim 10. In the procedure according to the invention for validating the operation of an AI-based system for detecting, segmenting, and classifying objects in the environment of a means of transport, preferably a railway vehicle, data is collected from sensors in the environment of the means of transport. The means of transport can be, in addition to a railway vehicle, a motor vehicle, an aircraft, a drone, or a bicycle.Autonomous object detection is particularly well-suited for the operation of drones, since these, which by definition do not have a pilot on board, still need the corresponding sensor technology for flight control and navigation. This AI-based system learns to detect objects through a training process using specific training data. Such a machine learning system might include, for example, an artificial neural network. Input data for the trained AI system, such as a trained artificial neural network, is generated from sensor data. This input data can be generated, for example, as an input data vector for the trained AI system based on the sensor data. Furthermore, an auxiliary prediction system determines the feature data of static objects based on the input data or intermediate layer data of the trained AI system.The auxiliary prediction system also comprises a machine learning-based system trained on the same training data as the AI-based object detection system. The auxiliary prediction system may be derived from the trained AI-based system. The feature data determined by the auxiliary prediction system may refer directly to object names or classifications, but may also refer to features extracted from the input data or intermediate layer data. Intermediate layer data refers to the AI-based system data generated between the input and output data of the AI-based system.An artificial neural network can be understood, for example, as composed of several layers. In this case, the intermediate layer data would be that generated in an intermediate layer of the artificial neural network located between the input and output layers. The determined feature data is then compared with field comparison data, which includes features of static objects in the environment. The field comparison data comes from a database and could be, for example, data from a digital map representing the environment of the transport vehicle. The field comparison data includes data relevant to a section of track along which the transport vehicle is traveling. If the transport vehicle is a railway, the data might include masts, signals, signs, beacons, poles, electrical boxes, bridges, switches, and track curves. These objects are assigned known positions.Comparison with field reference data can be performed, for example, by determining the vehicle's position and orientation during transit, for instance, using satellite navigation. The field comparison data is also assigned the geographical positions of the various objects or features, so that, based on the estimated position of a detected object, the field comparison data to be assigned to that object can be determined. The field comparison data constitutes a kind of "ground truth" for the predictions of the auxiliary prediction system. Finally, the reliability of the trained artificial neural network is determined based on the comparison results. The method according to the invention allows for the advantageous detection of safety problems caused by changes in the environment or the detection system, or by variations in environmental conditions, that affect the correct recognition and classification of objects. For example, the weather may change: it may rain, snow, or be foggy. Lighting conditions may vary, glare may occur, reflections may appear on objects, lenses may become dirty, and sensors may be affected. Furthermore, based on validation, a defective sensor in the means of transport used for object detection and identification can be detected. All these changes are detected by the method according to the invention. If a sensor is defective, the driver may receive an alert or be required to exercise greater vigilance while driving.As will be explained later, data from the faulty sensor can also simply be weighted more weakly or not weighted at all in object identification, in order not to skew the result. The validation system according to the invention includes a sensor data interface for receiving sensor data from the environment of a means of transport. Part of the validation system according to the invention is also an AI-based system, preferably a trained artificial neural network, for detecting and classifying an object within the environment of the means of transport. The validation system according to the invention also includes an auxiliary prediction system for processing input data or data from intermediate layers of the AI-based system, for example, a trained artificial neural network. The auxiliary prediction system is designed to determine characteristics of static objects. The auxiliary prediction system has been trained using the same data as the AI-based system.Part of the validation system according to the invention is also a comparison unit for comparing the determined characteristics with field comparison data, which includes characteristics of static objects in the environment of the means of transport. The validation system according to the invention also includes a validation unit for determining the reliability of the AI-based system's operation based on the comparison results. The validation system according to the invention shares the advantages of the procedure according to the invention for validating the operation of an AI-based system for the detection, segmentation, and classification of objects in the environment of a means of transport. In the method of the invention for generating field comparison data, sensor data is collected from the environment of a transport vehicle, and features are extracted from this data using a feature extractor, preferably an intermediate layer of the AI-based system. The extracted features are then stored as field comparison data along with position data. In this method, the field comparison data is advantageously generated while the transport vehicle is in operation. This is advantageous when field comparison data for a section of the railway network is unavailable. In such a case, a database can be created from this field comparison data while the vehicle is in motion. Since the transport vehicle typically travels a section of the route repeatedly, the resulting field comparison data can then be used for validation.In this case, the currently extracted features can be compared with those previously stored as field comparison data. This can be done, for example, by fitting a mixed Gaussian model to the stored features. The probability that the currently extracted features come from the same distribution as the stored features is then determined. If the probability is low, there is a high risk that the currently captured data falls outside the training data. The means of transport according to the invention includes a sensor unit for capturing sensor data, preferably image data, from the environment of the means of transport. Furthermore, the means of transport according to the invention comprises the validation system according to the invention. In addition, the means of transport according to the invention includes a control system for controlling the driving behavior of the means of transport based on a validation result from the validation system. The means of transport according to the invention shares the advantages of the validation system according to the invention. Some components of the invention's validation system may be configured, for the most part, as software components. This refers, in particular, to the sensor data interface, the AI-based trained system, the auxiliary prediction system, the comparison unit, and the validation unit. In principle, however, these components can also be implemented, at least in part, particularly for especially fast calculations, as software-assisted hardware, such as FPGAs or similar devices. Likewise, the necessary interfaces, for example, when solely for data transfer from other software components, can be configured as software interfaces. However, they can also be configured as hardware interfaces, controlled by appropriate software. A software-based implementation has the advantage that existing computer systems in a means of transport, after possible expansion with additional hardware elements such as a sensor unit, can be easily adapted via a software update to function as described in the invention. In this respect, the objective is also achieved by means of a corresponding software product, which can be loaded directly into a storage system of said computer system, with program sections for executing the steps of the procedure according to the invention that can be performed by software when the software program is running on the computer system. This software product may include, in addition to the software program, additional components such as documentation and / or other components, as well as hardware components such as hardware keys (dongles or adapters, etc.) for the use of the software For transport to and / or storage within the computer system, a computer-readable medium can be used, such as a USB flash drive, a hard drive, or other portable or embedded data storage device, on which sections of the computer program that can be read and executed by a computer unit are stored. The computer unit may, for example, consist of one or more microprocessors working together or similar components. The dependent claims, as well as the description that follows, respectively contain particularly advantageous configurations and developments of the invention. In this respect, the claims of one category of claims may be developed, in particular, analogously to the dependent claims of another category of claims and their descriptive portions. Furthermore, within the scope of the invention, the various features of different embodiments and claims may also be combined to give rise to new embodiments. In one configuration of the procedure according to the invention for validating the operation of an AI-based system for detecting and classifying objects in the environment of a means of transport, the field comparison data comprises a digital map of the environment of the means of transport, in which characteristics of static objects in the environment of the means of transport are stored. The environment preferably comprises the surroundings of a section of track along which the means of transport travels. Geographic positions are assigned to the static objects. If the means of transport detects an object, it can, provided its own position and the orientation of the sensors are known, determine on the digital map the field comparison data to be assigned to the detected objects.Digital maps of the areas through which a means of transport circulates are usually already available and, therefore, can be used for the comparison operation according to the invention with minimal adaptation effort. Preferably, the auxiliary prediction system comprises an artificial neural network trained on the same training data as the AI-based system, which, preferably, also comprises a trained artificial neural network. If an input dataset falls outside the training data, the auxiliary prediction system will also have difficulty correctly classifying objects or their associated features based on that input data, and a deviation will be determined when comparing it to the field comparison data. The auxiliary prediction system can also include an artificial neural network derived from the AI-based system, which in this case also includes a trained artificial neural network. Advantageously, the auxiliary prediction system behaves analogously, or at least very similarly, to the AI-based system, which contributes to obtaining particularly meaningful comparison or validation results. Alternatively, or additionally, the auxiliary prediction system can also include a support vector machine. Therefore, the auxiliary prediction system can also be based on a different design algorithm instead of an artificial neural network. The simplicity of this algorithm lies in its lower data requirements for learning and faster predictions; moreover, due to its lower hardware requirements, it is also suitable for use on less powerful computer systems than those needed for AI-based algorithms. It is particularly preferable that the auxiliary prediction system be identical to the AI-based system. In this variant, it is advantageous to be able to dispense with an additional AI-based system. Instead, the intermediate layer data or the results data from the AI-based system are directly compared with the field comparison data. Furthermore, the output data from the auxiliary prediction system can be of the same type as the output data from the AI-based system. Advantageously, field comparison data can be obtained or supplemented directly using the AI-based system. The sensor technology used for data acquisition may include an image capture unit for capturing two-dimensional or even three-dimensional image data of the surrounding area of ​​a means of transport. In the case of three-dimensional image data, this can be captured directly using a stereoscopic camera or generated by a 2D image capture unit via SAR (Synthetic Aperture Radar). Alternatively or additionally, radar sensors, infrared sensors, LIDAR sensors, or similar sensors may also be part of the sensor technology. In this context, validation according to the invention is used to determine which types of sensors are currently particularly unreliable, for example, due to unfavorable weather or lighting conditions for specific sensor types.In this case, reliable object detection can continue based on data from some of the sensors. Alternatively, the data from the various sensors can be weighted based on a validation result. This makes object detection and identification more reliable and independent of varying environmental influences. In a variant of the procedure according to the invention for validating the operation of an AI-based system for detecting and classifying objects in the environment of a means of transport, a validation is carried out when the means of transport starts from a departure point, such as a stop, train station, or sorting yard or station, in order to determine whether there are any defective sensors or whether the weather conditions are unfavorable for at least some of the sensors. Advantageously, this validation can also be performed in advance if digital mapping material is lacking for a route to be traveled, provided that the conditions at the departure point are known.In the event of invalidity, one can also opt for increased monitoring of the transport system's operation or for a type of safety mode in which the transport system operates at low speed, in order to account for any detected adverse environmental conditions. The invention is explained in more detail below, with reference to the accompanying figures and based on exemplary embodiments. In these: FIG.1 shows a flowchart illustrating a procedure for validating the operation of a neural network trained for the detection and classification of objects in the environment of a railway vehicle, according to an example of an embodiment of the invention, FIG.2 shows a schematic representation of a trained neural network, FIG.3 shows a schematic representation of an auxiliary prediction system; FIG.4 shows a schematic representation of a digital map that provides field comparison data; FIG.5 shows a validation system according to an example of an embodiment of the invention; FIG.6 shows a railway vehicle according to an example of an embodiment of the invention. Figure 1 shows a flowchart 100 illustrating a procedure for validating the operation of a trained neural network (TNN) (shown in Figure 2) intended for the detection and classification of objects in the environment of a railway vehicle 61. The railway vehicle 61 travels along a railway track through an environment U (see Figure 6) and, in step 1.I, collects SD sensor data from its environment using sensors arranged on the railway vehicle 61. From the SD sensor data, input data (ED) is generated, for example in the form of input vectors, which are fed into the trained neural network TNN. In step 1.II, based on the ED input data or on ZSD intermediate layer data which are based on the ED input data, an auxiliary prediction system APS (shown in FIG. 3) generates MD feature data of static objects from the trained TNN neural network. In step 1.III, the determined MD feature data are compared with the FVD field comparison data. The field comparison data includes features of static objects in the rail vehicle's environment. If the FVD field comparison data is found to be identical to the MD feature data or to deviate only by a predetermined threshold value, indicated in FIG. 1 with "y", then in step 1.IV the object detection is deemed valid. This is indicated in FIG. 1 with "V = 1". That is, if in step 1.III a match is found between the MD feature data and the FVD field comparison data, the validity (V) of the TNN trained on the basis of this comparison result in step 1.IV is confirmed. Conversely, if in step 1.III. It is determined that the FVD field comparison data differ from the MD characteristic data of the APS auxiliary prediction system, which is indicated in FIG.1 with "n", in step 1. V. A reduced validity of the trained artificial neural network TNN is assumed, which is symbolized in FIG.1 with "V = 0". In this case, it can be assumed that the ED input data of the trained artificial neural network TNN differs from the original training data of said artificial neural network TNN. Figure 2 shows a schematic representation of a trained artificial neural network (TNN) according to an exemplary embodiment of the invention. The trained artificial neural network has a plurality of convolutional layers (FS) for feature determination. Data points (DP) are input to the TNN. Feature maps (MK) or intermediate layer data (ZSD) are output as output data from the convolutional layers. The intermediate layer data (ZSD) is generated by an intermediate layer (ZS) of the trained artificial neural network (TNN) and is used as input data for the auxiliary prediction system (APS). Prediction data from the neural network (NV) is output as output data from the TNN. Figure 3 shows a schematic representation of an auxiliary prediction system (APS) according to an exemplary embodiment of the invention. The APS receives intermediate output data (ZSD) from the trained artificial neural network (TNN) as input data and applies a model-derived (MD) model, which, for example, may be AI-based, to this data. The output data generated by the APS is AV prediction data, which refers, for example, to recognized features or object classifications. Figure 4 shows a schematic representation of a DMRA digital map providing FVD field comparison data. Field comparison data links, for example, a PD position of a railway vehicle, determined, for example, by a satellite navigation system, with a set of FVD field comparison data in the vicinity of that PD position. The FVD field comparison data can include, for example, extracted features or classified objects located in the vicinity of the determined PD position. On the left of the image, for example, a mast and a sign can be seen, serving as comparison objects. Figure 5 schematically represents a validation system 50 according to an exemplary embodiment of the invention. The validation system 50 comprises a sensor data interface 51 for receiving sensor data SD from an environment U of a railway vehicle 61. The sensor data SD is aggregated into a data point DP, which is fed into a trained artificial neural network (TNN). The trained artificial neural network (TNN) is configured to detect and classify an object in the environment of the railway vehicle 61. As described in relation to Figure 2, this TNN has a plurality of convolutional layers with which the data point DP is processed. Instead of the prediction data NV from the TNN, intermediate layer data ZSD is transmitted to an auxiliary prediction system (APS), which is also part of the validation system 50.This auxiliary prediction system (APS) processes the intermediate layer data (ZSD) from the trained artificial neural network (TNN) and the detected objects, classifying these objects as feature data (MD). The MD-classified objects are transmitted to a comparison unit (52), which compares the classified objects with field comparison data (FVD) from a digital map (DMRA). For example, a plurality of objects have been plotted and named on the DMRA digital map. Consequently, their geographic position is also known. If the position of the railway vehicle is known, a comparison of the object classifications with the digital cartographic data or the field comparison data (FVD) from the DMRA digital map can be easily performed.The result of the VE comparison is transmitted to a validation unit 53, which decides, based on the result of the VE comparison, whether the trained TNN neural network is sufficiently reliable and whether, in particular, the DP data point of the neural network's input data is within the training dataset or whether, possibly due to a change in environmental conditions, such input data deviates considerably from the training data. Figure 6 illustrates a schematic representation 60 of a railway vehicle 61 according to an exemplary embodiment of the invention. The railway vehicle 61, schematically represented on the left in Figure 6, is shown. Railcar 61 travels from left to right in the direction of the arrow on a railway track GK and captures three-dimensional image data as SD sensor data from a surrounding area U located in front of the railcar 61. For this purpose, the railcar 61 includes a camera unit 62 in its front section. SD sensor data is captured from an environment U of the railcar 61, which also includes a mast P. The SD sensor data is transmitted to a validation system 50, which is also part of the railcar 61 and has the structure illustrated in FIG. 5. The validation system 50 then determines, based on the SD sensor data, whether the captured sensor data falls within the training data range of a TNN artificial neural network used for object recognition, or whether the environmental conditions have changed in such a way that object detection and classification are no longer reliable.The validation result V is transmitted to a control system 63 integrated into the rail vehicle 61, which, if the reliability of the SD sensor data is not guaranteed, triggers an appropriate response. For example, this could involve a braking maneuver combined with a warning that the rail vehicle 61 can only continue to be driven by a train driver. Alternatively, the weighting of the SD sensor data during object evaluation and classification can also be modified, or other sensors with potentially more reliable data can be used, in order to continue enabling safe driving based on the sensor data. Finally, it should be noted once again that the procedures and devices described above are merely examples of preferred embodiments of the invention and that those skilled in the art may modify them without leaving the scope of the invention as defined in the claims. For the sake of completeness, it should also be noted that the use of the indefinite articles "a" or "an" does not preclude the features in question from being present multiple times. Similarly, the term "unit" does not preclude it from consisting of several components, which, where applicable, may also be spatially distributed.

Claims

1. A procedure for validating the operation of an AI-based system (TNN) for the detection, segmentation, and classification of objects in an environment (U) of a means of transport (61), comprising the following steps: - Capture of sensor data (SD) from the environment (U) of the means of transport (61), - Generation of input data (ED, DP) for the AI-based system (TNN) from the sensor data (SD), - Generation of feature data (MD) of static objects from the input data (ED, DP) or intermediate layer data (ZSD) of the AI-based system (TNN) by means of an auxiliary prediction system (APS), - Comparison of the determined feature data (MD) with field comparison data (FVD), comprising features of static objects in the environment (U), - Determination of the reliability (V) of the AI-based system (TNN) from the result of the comparison (VE),wherein functional validation is used to determine which types of sensors are currently particularly unreliable, and reliable object detection is then carried out based on a subset of the sensors, or the data from the various sensors is weighted based on the validation result during object detection.

2. A method according to claim 1, wherein the field comparison data (FVD) comprises a digital map (DMRA) of the environment (U) of the transport means (60), in which features of static objects in the environment (U) of the transport means (61) are stored.

3. A method according to claim 1 or 2, wherein the environment (U) comprises the surroundings of a section of track (GK) along which the transport means (61) travels.

4. A method according to any one of the preceding claims,wherein the auxiliary prediction system (APS) comprises an artificial neural network that has been trained with the same training data as the AI-based system (TNN).

5. A method according to any of the preceding claims, wherein the auxiliary prediction system (APS) comprises an artificial neural network derived from the AI-based system (TNN).

6. A method according to any of the preceding claims, wherein the auxiliary prediction system (APS) comprises a support vector machine.

7. A method according to any of the preceding claims, wherein the auxiliary prediction system (APS) is identical to the AI-based system (TNN).

8. A method according to any of the preceding claims,in which the output data (AV) of the auxiliary prediction system (APS) are of the same type as the output data (NV) of the AI-based system (TNN).

9. Validation system (50) comprising: - a sensor data interface (51) for receiving sensor data (SD) from an environment (U) of a transport medium (61), - an AI-based system (TNN) for detecting and classifying an object in the environment (U) of the transport medium (61), - an auxiliary prediction system (APS) for processing input data (ED, DP) or intermediate layer data (ZSD) from the AI-based system (TNN), the auxiliary prediction system (APS) being designed to determine features (MD) of static objects, - a comparison unit (52) for determining a comparison result (VE) based on the determined features (MD) and field comparison data (FVD), comprising features of static objects in the environment (U) of the transport medium (61),- a validation unit (53) for determining the reliability (V) of the AI-based system (TNN) from the comparison result (VE), using functional validation to determine which types of sensors are currently particularly unreliable, and then proceeding with reliable object detection based on a subset of the sensors, or weighting the data from the various sensors based on the validation result during object detection.

10. Transport means (61), comprising - a sensor unit (62) for capturing sensor data (SD) from the environment of the transport means (61), - a validation system (50) according to claim 9,- a control system (63) for controlling the transport means (61) based on a validation result (VE) from the validation system (50).

11. Computer program product comprising a computer program that can be directly loaded into a memory unit of a control system (63) of a transport means (61), with program sections for executing all the steps of a procedure according to any one of claims 1 to 8, when the computer program is executed in the control system (63).

12. Computer-readable medium on which program sections executable by a processing unit are stored to carry out all the steps of the procedure according to any one of claims 1 to 8, when the program sections are executed by the processing unit.

Images