METHOD FOR TESTING THE RELIABILITY OF AI-BASED OBJECT DETECTION

Description

[0001] The invention relates to a method for validating the function 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 device. Finally, the invention relates to a means of transport.

[0002] In autonomous or semi-autonomous driving, the vehicle's surroundings must be monitored to prevent collisions. For example, in rail operations, objects such as people, road vehicles, shopping carts thrown onto the tracks, boulders, or fallen trees occasionally end up on the track and pose a danger to rail safety. In the case of people and road vehicles, the potential for collision with a moving train also puts them at high risk. Therefore, such objects must be detected in time to initiate braking for an approaching train, thus preventing a collision.

[0003] Therefore, detecting potential obstacles of any shape or form that could block the tracks is a safety-critical necessity for all types of rail vehicles. Many approaches exist to address this problem, with artificial intelligence (AI)-based solutions being the most promising. However, any AI-based solution is only as good as its training database, which is very complex to compile and annotate for rail traffic scenarios.

[0004] In recent years, deep learning-based computer vision approaches have achieved high effectiveness compared to classical approaches based on manual feature analysis, and are therefore very promising for addressing the challenges in the field of automated or driverless operation of rail vehicles. For complex scenarios requiring particularly sophisticated environmental perception, such as driverless rail operation, it is foreseeable that these tasks cannot be solved without deep learning-based methods. However, deep learning-based methods are not entirely error-free and therefore not completely reliable. One problem is that neural networks exhibit a kind of overconfidence for data that was not part of a training dataset.In this context, it is said that the distribution of data occurring in reality lies outside the distribution of the training data. This data is also referred to as OOD data (OOD = out of distribution = outside the distribution of the training data). Therefore, a confidence measurement cannot be used as a measure of the reliability of a result dataset from an artificial neural network.

[0005] Especially in the operation of highly automated driverless vehicles, such as road vehicles or rail vehicles, increased sensitivity to the occurrence of data outside the training dataset is required for widespread application, so that the system can be brought to a safe state in such a case. Currently, approaches based on machine learning or deep learning typically exhibit the following shortcomings: Typical artificial neural networks lack inherent mechanisms to determine whether a data point in the input data is included in the training data; artificial neural networks typically exhibit lower hit rates and accuracy with OOD data, even when they achieve high confidence in their predictions.

[0006] Therefore, it is difficult to decide whether a prediction or decision of an artificial neural network is trustworthy and reliable or not.

[0007] There are various phenomena that can cause input data to lie outside the data distribution of the training data: Such a deviation can occur due to a variation of the detected object, for example by changing the object type, by the movement of an object, which results in different poses of pedestrians or vehicles, or different sizes of objects.

[0008] A deviation can also be caused by changes in the environment or the detector system. For example, the weather can change; it can rain, snow, or be foggy. Lighting conditions can change, glare can occur, reflections from objects can appear, lenses can become dirty, and sensors can malfunction.

[0009] All these changes can lead to confident but incorrect predictions. For example, this can lead to a misinterpretation of the signal state of a traffic light.

[0010] In US 2021 / 0 004 017 A1, high-resolution map data is used to generate synthetic sensor data for autonomous vehicles.

[0011] In US 2020 / 0410254A1, intersection areas in the vicinity of a vehicle are detected and located in real time or near real time using sensors and an artificial neural network. The application tests the functionality of an AI-based system for object detection of a vehicle against a stationary, identically functioning AI-based system.

[0012] The task, therefore, is to improve the reliability of AI-based sensor data evaluation, particularly in the field of autonomous or assisted driving in rail transport, and thus to increase the traffic safety of the automated or semi-automated operation of such systems.

[0013] This problem is solved by a method for validating the function of an AI-based system for detecting and classifying objects in the environment of a means of transport according to claim 1, a validation device according to claim 9, and a means of transport according to claim 10. In the inventive method for validating the function of an AI-based system for detecting, segmenting, and classifying objects in the environment of a means of transport, preferably a rail vehicle, sensor data from the environment of the means of transport are acquired. The means of transport can include, in addition to a rail vehicle, a motor vehicle or an aircraft, a drone, or a bicycle. Autonomous object detection is particularly suitable for the operation of drones, since drones, which by definition do not have a pilot on board, require the corresponding sensors for flight control and navigation anyway.

[0014] Such an AI-based system learns to detect objects using specific training data as part of a training process. This machine learning-based system can, for example, include an artificial neural network. Input data for the AI-based trained system, such as a trained artificial neural network, is generated based on sensor data. This input data can, for example, be generated as an input data vector for the AI-based trained system based on the sensor data. Additionally, feature data of static objects is determined by an auxiliary prediction system based on the input data or on intermediate layer data from the trained AI-based system. This auxiliary prediction system also includes a machine learning-based system that has been trained with the same training data as the AI-based system for object detection.The auxiliary prediction system can be derived from the trained AI-based system. The feature data determined by the auxiliary prediction system can directly relate to the labels or classification of objects, or it can relate to extracted features from the input data or intermediate layer data. Intermediate layer data refers to data generated by the AI-based system between its input and output data. For example, an artificial neural network can be understood as consisting of multiple layers; in this case, intermediate layer data would be data 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 comprises features of static objects in the environment.The field comparison data originates from a database and can, for example, be data from a digital map depicting the transport vehicle's surroundings. This field comparison data comprises information relevant to the section of track the transport vehicle is currently traveling through. If the transport vehicle is a rail vehicle, the data can include masts, signals, signs, balises, poles, electrical boxes, bridges, switches, and track curves. Known positions are assigned to these objects. Comparison with the field comparison data can be achieved, for example, by determining the transport vehicle's ego position and orientation during its journey, perhaps using satellite navigation.The field comparison data also includes the geographic positions assigned to each object or feature, so that the field comparison data corresponding to a detected object can be determined based on its estimated position. This field comparison data forms 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.

[0015] Advantageously, the method according to the invention can detect safety problems caused by changes in the environment or the detector system, or by changes in environmental conditions, which impair correct object detection and classification. For example, the weather can change; it can rain, snow, or be foggy. Lighting conditions can change, glare can occur, reflections can appear on objects, lenses can become dirty, and sensors can be affected by defects. Furthermore, based on the validation, a defective sensor in the means of transport, which is 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 be alerted or required to increase their monitoring of the vehicle's operation. As explained later, sensor data from the defective sensor can also be weighted less heavily or not at all during object identification to avoid distorting the result.

[0016] The validation device according to the invention has a sensor data interface for receiving sensor data from the environment of a means of transport. Part of the validation device according to the invention is also an AI-based system, preferably a trained artificial neural network, for detecting and classifying an object in the environment of the means of transport. The validation device according to the invention also has an auxiliary prediction system for processing input data or intermediate layer data of the AI-based system, for example, a trained artificial neural network. The auxiliary prediction system is configured to determine features of static objects. The auxiliary prediction system is trained with the same data as the AI-based system.Part of the validation device according to the invention is also a comparison unit for comparing the determined features with field comparison data, which includes features of static objects in the environment of the means of transport. The validation device according to the invention also includes a validation unit for determining the reliability of the function of the AI-based system based on the comparison result. The validation device according to the invention shares the advantages of the method according to the invention for validating the function of an AI-based system for the detection, segmentation, and classification of objects in the environment of a means of transport.

[0017] In the inventive method for generating field comparison data, sensor data from the environment of a means of transport are acquired, and features are extracted from the sensor data by a feature extractor, preferably an intermediate layer of the AI-based system. Furthermore, the extracted features are stored as field comparison data together with position data. In this method, field comparison data is advantageously generated during the operation of the means of transport. This is beneficial when no field comparison data exists for a section of a rail network. In such cases, a database of such field comparison data can be built up while the vehicle is in motion. Since the means of transport typically travels a section of track repeatedly, the field comparison data thus acquired can subsequently be used for validation.Here, newly extracted features can be compared with those features previously stored as field comparison data. This can be done, for example, by fitting a Gaussian mixture model to the stored features. The probability that the newly 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 newly acquired data lies outside the training data.

[0018] The transport means according to the invention comprises a sensor unit for acquiring sensor data, preferably image data, from the environment of the transport means. Furthermore, the transport means according to the invention includes the validation device according to the invention. The transport means according to the invention also comprises a control device for controlling the driving behavior of the transport means depending on a validation result of the validation device. The transport means according to the invention shares the advantages of the validation device according to the invention.

[0019] Some components of the validation device according to the invention can be predominantly designed in the form of software components. This applies in particular to the sensor data interface, the trained AI-based system, the auxiliary prediction system, the comparison unit, and the validation unit.

[0020] In principle, these components can also be partially implemented as software-supported hardware, such as FPGAs or similar devices, especially when particularly fast calculations are required. Similarly, the necessary interfaces, for example, when it is only a matter of transferring data from other software components, can be implemented as software interfaces. However, they can also be implemented as hardware interfaces controlled by suitable software.

[0021] A largely software-based implementation has the advantage that even computer systems already present in a means of transport can be easily retrofitted with a software update to operate according to the invention, after possible addition of further hardware elements, such as a sensor unit. In this respect, the problem is also solved by a corresponding computer program product with a computer program that can be directly loaded into a storage device of such a computer system, containing program sections to execute the software-implemented steps of the method according to the invention when the computer program is executed in the computer system.

[0022] Such a computer program product may, in addition to the computer program itself, include additional components such as documentation and / or additional components, including hardware components such as hardware keys (dongles, etc.) for using the software.

[0023] For transport to the computer system's storage device and / or for storage on the computer system, a computer-readable medium, such as a memory stick, a hard drive, or other portable or permanently installed data carrier, can be used, on which the program sections of the computer program that can be read and executed by a computer unit are stored. The computer unit can, for example, include one or more cooperating microprocessors or similar components.

[0024] The dependent claims and the subsequent description each contain particularly advantageous embodiments and further developments of the invention. In particular, the claims of one claim category may also be further developed analogously to the dependent claims of another claim category and their descriptive parts. Furthermore, within the scope of the invention, the various features of different embodiments and claims may also be combined to form new embodiments.

[0025] In one embodiment of the inventive method for validating the function 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 features of static objects in the environment of the means of transport are stored. The environment preferably comprises the area surrounding a track section on which the means of transport is moving. Geographic positions are assigned to the static objects. If the means of transport detects an object, it can determine field comparison data in the digital map that can be assigned to the detected objects, provided its own position and the direction of the sensors are known. Digital maps of the areas through which a means of transport travels are usually already available and can therefore be used for the comparison operation according to the invention with minimal adaptation effort.

[0026] 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 a set of input data lies outside the training data, the auxiliary prediction system will also have problems correctly classifying objects or features associated with objects based on this input data, and a deviation will be detected when comparing it with the field comparison data.

[0027] 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, contributing to particularly meaningful comparison and validation results.

[0028] Alternatively or additionally, the auxiliary prediction system can also include a support vector machine. Thus, instead of an artificial neural network, the auxiliary prediction system can also be based on a differently designed algorithm. The simplicity of this algorithm lies in the fact that it requires less data for training and delivers faster predictions. Furthermore, due to its lower hardware requirements, it is also suitable for use on less powerful computer systems than AI-based algorithms.

[0029] The preferred method is to use an auxiliary prediction system identical to the AI-based system. This variant offers the advantage of eliminating the need for an additional AI-based system. Instead, intermediate layer data or results data from the AI-based system are directly compared with the field comparison data.

[0030] The input data for the auxiliary prediction system can also be of the same type as the input data for the AI-based system. Advantageously, the field comparison data can be directly obtained or supplemented using the AI-based system.

[0031] The sensor system used for acquiring sensor data can include an image acquisition 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 by a stereo camera or generated by a 2D image acquisition unit using so-called SAR imaging (SAR = Synthetic Aperture Radar). Alternatively or additionally, radar sensors, infrared sensors, lidar sensors, or similar devices can also be part of the sensor system. The validation according to the invention is used in this context to determine which type of sensors are currently particularly unreliable, for example, due to unfavorable weather conditions or lighting conditions for specific sensor types. In this case, reliable object detection can be continued based on sensor data from some of the sensors.Alternatively, the sensor data from different sensors can be weighted based on a validation result. This makes the detection and identification of objects more reliable and less dependent on variable environmental influences.

[0032] In one variant of the inventive method for validating the function of an AI-based system for detecting and classifying objects in the environment of a means of transport, a validation is performed at the start of the means of transport at a starting point, such as a stop, a station, a railway depot, or a marshalling yard, in order to determine whether sensors are defective or whether weather conditions are unfavorable for at least some of the sensors. Advantageously, such a validation can also be carried out in advance if digital map data for a route to be traveled is lacking and the conditions at the starting point are known.Even if validity is lacking, increased attention may be paid to monitoring the operation of the means of transport, or a kind of safety mode may be chosen in which the means of transport is operated at a low speed, in order to take account of the identified adverse environmental conditions.

[0033] The invention is explained in more detail below with reference to the accompanying figures and exemplary embodiments. The figures show: FIG 1 a flowchart illustrating a method for validating the function of a trained neural network for detecting and classifying objects in the environment of a rail vehicle according to an embodiment of the invention, FIG 2 a schematic representation of a trained neural network, FIG 3 a schematic representation of an auxiliary forecasting system, FIG 4 a schematic representation of a digital map which provides field comparison data, FIG 5 a validation device according to an embodiment of the invention, FIG 6 a rail vehicle according to an embodiment of the invention.

[0034] In FIG 1 A flowchart 100 is shown, which illustrates a procedure for validating the function of a trained neural network TNN (in FIG 2 (shown) illustrates the detection and classification of objects in the environment of a rail vehicle 61. The rail vehicle 61 travels on a rail track through an environment U (see FIG 6 ) and in step 1.I, sensor data SD from the environment is acquired by sensors arranged on the rail vehicle 61. Based on the sensor data SD, input data ED, for example in the form of input vectors, is generated and fed into the trained neural network TNN.

[0035] In step 1.II, feature data MD of static objects of the trained neural network TNN are generated by an auxiliary prediction system APS (in) based on the input data ED or on intermediate layer data ZSD that are based on the input data ED. FIG 3 (shown) generated.

[0036] In step 1.III, the determined characteristic data MD are compared with field comparison data FVD. The field comparison data comprises characteristics of static objects in the vicinity of the rail vehicle. If it is determined that the field comparison data FVD is identical to the characteristic data MD or differs only by a predetermined threshold value, which is then FIG 1 If the object is marked with "y", then step 1.IV determines that the object detection is valid. This is in FIG 1 marked with "V = 1". This means that if, in step 1.III, a match between the feature data MD and the field comparison data FVD is found, the validity V of the trained artificial neural network TNN is confirmed in step 1.IV based on this comparison result. However, if, in step 1.III, it is determined that the field comparison data FVD differs from the feature data MD of the auxiliary prediction system APS, which in FIG 1 If marked with "n", a reduced validity of the trained artificial neural network TNN is assumed in step 1.V, which is in FIG 1 is symbolized by "V = 0". In this case, it can be assumed that the input data ED of the trained artificial neural network TNN differs from the original training data of this artificial neural network TNN.

[0037] In FIG 2 Figure 1 shows a schematic representation of a trained artificial neural network (TNN) according to an embodiment of the invention. The trained artificial neural network has a plurality of convolutional layers (FS) for determining features. Data points (DP) are inputted to the neural network TNN as input data. Feature maps (MK) or intermediate layer data (ZSD) are output as output data from the convolutional layers (FS). The intermediate layer data (ZSD) are generated by an intermediate layer (ZS) of the trained artificial neural network TNN and are used as input data for the auxiliary prediction system (APS). Network prediction data (NV) are output as output data from the neural network TNN.

[0038] In FIG 3 Figure 1 shows a schematic representation of an auxiliary prediction system (APS) according to an embodiment of the invention. The APS receives the intermediate output data (ZSD) of the trained artificial neural network (TNN) as input data and applies a model (MD), which may itself be AI-based, to this data. The output data generated are prediction data (AV) from the APS, which relate, for example, to recognized features or classifications of objects.

[0039] In FIG 4 The diagram shows a schematic representation of a digital map (DMRA) that provides field comparison data (FVD). The field comparison data links, for example, the position (PD) of a rail vehicle, determined using a satellite navigation system, with a group of field comparison data (FVD) located near that position (PD). The field comparison data (FVD) can include, for example, extracted features or classified objects located in the vicinity of the determined position (PD). On the left of the image, for example, a pole and a sign can be seen, which are suitable as comparison objects.

[0040] In FIG 5 A validation device 50 according to an embodiment of the invention is shown schematically. The validation device 50 comprises a sensor data interface 51 for receiving sensor data SD from the environment U of a rail vehicle 61. The sensor data SD are summarized into a data point DP, which is input 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 rail vehicle 61. As described in connection with FIG 2 As described, such an artificial neural network (TNN) has a plurality of convolutional layers with which the data point (DP) is processed. Instead of the prediction data (NV) of the artificial neural network (TNN), intermediate layer data (ZSD) are transmitted to an auxiliary prediction system (APS), which is also part of the validation unit (V50). This APS processes the intermediate layer data (ZSD) of the trained artificial neural network (TNN) and detected objects, classifying these objects as feature data (MD). The classified objects (MD) are transmitted to a comparison unit (52), which compares the classified objects with field comparison data (FVD) from a digital map (DMRA). For example, the digital map (DMRA) contains a plurality of objects that are plotted and named. Thus, their geographic position is also known.If the position of the rail vehicle is known, a comparison of the object classifications with digital map data or field comparison data (FVD) of the digital map DMRA can easily be performed. The comparison result VE is transmitted to a validation unit 53, which decides, based on the comparison result VE, whether the trained neural network TNN is sufficiently reliable and, in particular, whether the data point DP of the neural network's input data lies within the set of training data or whether, possibly due to a change in environmental conditions, this input data deviates significantly from the training data.

[0041] In FIG 6 A schematic representation 60 of a rail vehicle 61 according to an embodiment of the invention is illustrated. The one on the left in FIG 6 The schematically depicted rail vehicle 61 travels from left to right in the direction of the arrow on a track GK and acquires three-dimensional image data as sensor data SD from an area in front of the rail vehicle 61. U. For this purpose, the rail vehicle 61 includes a camera unit 62 in its front area. Sensor data SD from an environment U of the rail vehicle 61, which also includes a mast P, are acquired. The sensor data SD are transmitted to a validation device 50, which is also part of the rail vehicle 61 and which is located in FIG 5The illustrated setup is shown. The validation unit 50 now determines, based on the sensor data SD, whether the acquired sensor data is within the training data range of an artificial neural network (TNN) used for object recognition, or whether the environmental conditions may have changed to such an extent that object detection and classification are no longer reliable. The validation result V is transmitted to a control unit 63 integrated into the rail vehicle 61, which triggers an appropriate response if the sensor data SD is not reliable. For example, this could involve a braking maneuver combined with a message that the rail vehicle 61 may only be operated by a driver.Alternatively, the weighting of sensor data (SD) during evaluation and object classification can be changed, or a switch can be made to other sensors whose sensor data may be more reliable, in order to further enable safe sensor-based driving.

[0042] Finally, it should be noted once again that the methods and devices described above are merely preferred embodiments of the invention and that the invention can be varied by a person skilled in the art without departing from the scope of the invention, insofar as it is defined by 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 possibility that the features in question may be present multiple times. Likewise, the term "unit" does not preclude the possibility that it consists of several components, which may also be spatially distributed.

Claims

1. Method for validating the function of an AI-based system (TNN) for detecting, segmenting and classifying objects in an environment (U) of a transport means (61), having the steps: - detecting sensor data (SD) from the environment (U) of the transport means (61), - generating input data (ED, DP) for the AI-based system (TNN) based on the sensor data (SD), - generating characteristic data (MD) of static objects based on the input data (ED, DP) or intermediate layer data (ZSD) of the AI-based system (TNN) by way of an auxiliary prediction system (APS), - comparing the ascertained characteristic data (MD) with field comparison data (FVD), which comprises characteristics of static objects in the environment (U), - ascertaining the reliability (V) of the AI-based system (TNN) based on the comparison result (VE), wherein validating the function is used to ascertain which type of sensors are currently especially unreliable, and reliable object recognition is continued based on a part of the sensor or a weighting of the sensor data of different sensors is performed based on the validation result with an instance of object recognition.

2. 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 map, characteristics of static objects in the environment (U) of the transport means (61) are stored.

3. Method according to claim 1 or 2, wherein the environment (U) comprises the periphery of a track section (GK) on which the transport means (61) moves.

4. Method according to one of the preceding claims, wherein the auxiliary prediction system (APS) comprises an artificial neural network which was trained with the same training data as the AI-based system (TNN).

5. Method according to one of the preceding claims, wherein the auxiliary prediction system (APS) comprises an artificial neural network which was derived from the AI-based system (TNN) .

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

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

8. Method according to one of the preceding claims, wherein the output data (AV) of the auxiliary prediction system (APS) is of the same type as the output data (NV) of the AI-based system (TNN).

9. Validation facility (50), having - a sensor data interface (51) for receiving sensor data (SD) from an environment (U) of a transport means (61), - an AI-based system (TNN) for detecting and classifying an object in the environment (U) of the transport means (61), - an auxiliary prediction system (APS) for processing input data (ED, DP) or intermediate layer data (ZSD) of the AI-based system (TNN), wherein the auxiliary prediction system (APS) is designed to ascertain characteristics (MD) of static objects, - a comparison unit (52) for ascertaining a comparison result (VE) based on the ascertained characteristics (MD) and the field comparison data (FVD), which comprises characteristics of static objects in the environment (U) of the transport means (61), - a validation unit (53) for ascertaining the reliability (V) of the AI-based system (TNN) based on the comparison result (VE), wherein validating the function is used to ascertain which type of sensors are currently especially unreliable, and reliable object recognition is continued based on a part of the sensor or a weighting of the sensor data of different sensors is performed based on the validation result with an instance of object recognition.

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

11. Computer program product with a computer program, which can be loaded directly into a memory unit of a control facility (63) of a transport means (61), with program sections for performing all steps of a method according to one of claims 1 to 8, when the computer program is executed in the control facility (63).

12. Computer-readable medium, on which program sections which can be executed by a computing unit are stored for carrying out all steps of the method according to one of claims 1 to 8, when the program sections are executed by the computing unit.