Vehicle data system and method for determining useful or communicable vehicle data from ambient capture sensors - Patents.com
The vehicle data system with neural networks and autoencoders optimizes data collection from real-world scenarios, addressing edge cases and reducing computational load to enhance AI development in vehicles.
Patent Information
- Application Number
- JP2024531112
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-12-14
- Filing Date
- 2022-12-12
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-12-12
AI Technical Summary
Conventional methods for collecting data for vehicle AI development are limited to scenarios encountered during testing, failing to capture edge cases and requiring significant computational resources for data selection, which is not feasible for mass-produced vehicles.
A vehicle data system utilizing a surrounding detection system and a monitoring system with trained artificial neural networks, including an autoencoder, to identify and transfer useful data efficiently, minimizing computational load and capturing edge cases.
The system effectively identifies and transfers useful data from real-world scenarios, enhancing AI development by incorporating edge cases while reducing computational requirements, thus improving the robustness of vehicle detection systems.
Smart Images

Figure 2024541489000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to vehicle data systems and methods for determining communicable vehicle data from surrounding capture sensors, which may be employed, for example, in the context of improving systems for assisted or automated driving. [Background technology]
[0002] Conventional techniques involve training networks for sensing functions or vehicle functions on training data acquired, for example, by test vehicles during development, which is limited to data from scenarios encountered at that time. To achieve a sufficient coverage of road traffic scenarios, data collection during the actual use of the vehicle is advantageous, as this allows a wide diversity of data to select training data from.
[0003] Data collection with test vehicles during development also includes scenarios that are collected at this time. However, during operation of series production vehicles, further scenarios occur that cannot be covered sufficiently or hardly at all during development. These are in particular "edge cases" (meaning borderline cases, special cases, individual cases, or special cases that occur so rarely that they are not usually recorded by test vehicles or are recorded only randomly). To develop an artificial intelligence (KI) that can also handle edge cases, data collection during the operation of the KI helps. In order to reduce the number of data transferred to further development, data selection in the vehicles is necessary. Here, the computational load required for the data selection must be kept to a minimum due to the limited budget for the systems installed in the vehicles.
[0004] That is, for the development of KI algorithms for ADAS and AD in mass-produced vehicles, a method that requires only a small amount of computing time to evaluate whether a road traffic scenario is useful would be advantageous.
[0005] A general method for obtaining training data from a production vehicle is described for individual sensors in WO2020 / 056331A1. An artificial neural network is provided within the vehicle for evaluating the sensor data. A trigger classification means is applied to intermediate results from the neural network to determine a classification score for the sensor data. A decision is made whether or not to forward at least a portion of the sensor data via the computer network based at least in part on the classification score. In the event of a positive decision, the sensor data is forwarded and used to generate training data. [Prior art documents] [Patent documents]
[0006] [Patent Document 1] WO2020 / 056331A1 Summary of the Invention [Problem to be solved by the invention]
[0007] It is an object of the present invention to provide an optimized method for efficiently collecting useful data from multiple vehicles. [Means for solving the problem]
[0008] One aspect relates to identifying useful examples and edge cases from the vehicle fleet for optimization of data-based algorithms and / or machine learning systems and methods using data.
[0009] The vehicle data system of the present invention includes a surroundings detection system and a monitoring system within the vehicle. The perimeter detection system is adapted to receive and evaluate input data X of perimeter acquisition sensors. The input data X is evaluated by a first trained artificial neural network K, and as a result of this evaluation perimeter detection data Y' is output. The monitor system then feeds the same input data X from the peripheral capture sensors to a second trained artificial neural network K. R and outputs the reproduced data X' as a result of this evaluation. If the monitoring system detects a difference in the reconstructed data X' with respect to the input data X that exceeds a threshold (→ a candidate edge case), the monitoring system prompts the transfer of the input data X to an independent data unit. The independent data unit can be, for example, an external server, a cloud memory, or the backbone of a V2X system. V2X (vehicle-to-X) refers to a communication system or telematics system for the vehicle to communicate with other participants.
[0010] In one embodiment, the second artificial neural network K R is an autoencoder. The described "non-safety-critical" (monitoring) system is implemented as an auto-encoder. It is based on the same input data as the detection algorithms or functions described above. Due to its working principle, the auto-encoder offers significant advantages compared to other possible approaches.
[0011] We illustrate this in an example where the sensor data is image data: If an image is input to an autoencoder, the encoder tries to reproduce it at the output. Autoencoders can be trained using only the input signal, without additional labels. This not only has the advantage that no additional effort is required in terms of annotation, but also offers the crucial advantage that the error associated with the input data can be quantified at any time (even for unknown data). However, a suitable thresholding is recommended. Autoencoders can be used on the entire image or on a small excerpt.
[0012] Thus, the autoencoder is able to measure its own error for each possible input signal. In machine learning methods, higher error values are generally expected for input data that do not match the training data well. Since the autoencoder and the detector are based on the same data, the following advantages are obtained: unknown or uncertain scenarios identified by the autoencoder imply that such scenarios are not well represented in the training data, and a broad coverage of traffic scenarios is useful for the feature development.
[0013] According to one embodiment, the monitoring system calculates a score that estimates the usefulness of the input data X based on the difference.
[0014] In an embodiment, a first artificial neural network K is trained using predefined training data, where reference output data Y_1, Y_2, ..., Y_n are used for input data X_1, X_2, ..., X_n, respectively. By adjusting the weights of the first neural network K, a first error function, which indicates the difference between the output of the first neural network K for the input data X_1, X_2, ..., X_n and the corresponding reference output data Y_1, Y_2, ..., Y_n, is minimized.
[0015] According to one embodiment, a second artificial neural network K R is the second artificial neural network K RThe weights of the vectors are adjusted so that a second error function, which indicates the difference between the reproduced data X' and the input data X of the peripheral acquisition sensor, is minimized.
[0016] In an embodiment, additionally, in addition to the input data X of the perimeter acquisition sensors, meta-information is also transmitted, which corresponds to one or more of the following groups of information: latest software version, calculated score of the monitoring system, GPS data, date, time, vehicle identification number (FIN=Fahrzeug-Identifikations-Number) and cloud data allowing to grasp the scene and / or vehicle situation. On this basis, for example, the unfolding scene can be faithfully reproduced (simulation with the same software version). This information acquisition allows unknown scenes and edge cases to be selected and directly fed into the development process of the vehicle detection or surroundings detection. This allows a continuous quality assurance process to be established. With each development step, more and more useful data is fed into the system.
[0017] Furthermore, the maturity of the software can be derived from the number of data transmissions received: the less the number of data transmission processes due to lack of prediction accuracy, the higher the maturity of the software.
[0018] According to one embodiment, the perimeter detection system is configured to receive input data X from a number of perimeter acquisition sensors and evaluate them collectively. This corresponds to a multi-sensor setup. A multi-sensor system has the advantage that it can increase the safety of detection algorithms for road traffic by validating the detections of several surrounding acquisition sensors. A multi-sensor system can be used for example with the following candidates: - one or more cameras, - one or more radars, - one or more ultrasound systems; - one or more riders; - and / or one or more microphones It can be any combination of.
[0019] In this case, a multi-sensor system consists of at least two sensors. The data acquisition of one of these sensors s at a time t can be denoted as D_(s,t). These data D can be images and / or audio recordings, angle, distance, speed measurements, and reflections from several objects in the periphery.
[0020] In one embodiment, the monitor system is configured to process perimeter acquisition sensor input data X in parallel with the perimeter detection system.
[0021] According to one embodiment, the monitoring system is integrated into a perimeter detection system as an additional detector head, the monitoring system and the perimeter detection system using a common encoder.
[0022] In one embodiment, the input data of the perimeter capture sensor is image data, and the monitor system is configured to reconstruct the entire image or a partial excerpt of the image, and can further estimate and output a value of the reconstruction error.
[0023] According to one embodiment, the monitoring system is configured to determine and output an uncertainty measure, the uncertainty measure being a value that gives how certain the monitoring system is in outputting the reproduced data X'.
[0024] In one embodiment, the monitoring system is configured to consider the consistency of the reconstruction data X' over time.
[0025] In some implementation variations, a distinction is additionally made as to whether the difference in the monitor system's reconstructed data X' from the input data X is continuous or only present for a limited period of time.
[0026] In one embodiment, both the perimeter detection system and the monitoring system are configured for over the air updates.
[0027] A further object of the present invention relates to a method for determining vehicle data worth transmitting from a perimeter capture sensor, the method comprising the steps of: evaluating input data X of the perimeter capture sensor by a perimeter detection system implemented using a trained first artificial neural network; outputting the perimeter detection data Y' determined during the evaluation; A second trained artificial neural network, K R evaluating the same input data X of the perimeter capture sensor by the monitoring system implemented using outputting the reproduced data X' determined during the evaluation; and a step of instructing transmission of the input data X to an independent data unit when a difference between the reproduced data X' and the input data X that exceeds a threshold is identified;
[0028] The evaluation of the usefulness of the traffic scenarios can be further improved by considering a multi-sensor setup, its evolution over time, and an estimation of the reliability of predicting the network output. Furthermore, the short computation time is advantageous for embedded systems. These points can be addressed, for example, by choosing less computationally intensive monitor systems or autoencoder approaches and extending them for multi-sensor systems with reliability estimation and longitudinal validation.
[0029] The following examples and figures are explained in more detail. Figure Description: [Brief description of the drawings]
[0030] [Figure 1] Figure 1 shows a vehicle equipped with surrounding sensors and a data system. [Diagram 2]FIG. 2 shows the peripheral acquisition sensors, data system and independent data unit. [Diagram 3] FIG. 3 shows an embodiment of a classification system and a monitor system as a peripheral detection system. [Figure 4] Figure 4 shows a vehicle capturing an unknown scene / situation. [Diagram 5] FIG. 5 shows a grayscale image showing overexposure due to incorrect exposure control of an on-board camera, as well as [Figure 6] FIG. 6 shows a grayscale image of a dashcam that contains a work of art that may lead to unpredictable situations when evaluated. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0031] FIG. 1 shows a vehicle 2 equipped with a perimeter acquisition sensor 1 and a data system 10. The depicted perimeter acquisition sensor 1 corresponds to a camera arranged inside the vehicle 2, for example in the area of the rear-view mirror, which acquires the vehicle's surroundings through the windshield of the vehicle 2. Further perimeter acquisition sensors 1 of the vehicle 2 can be, for example, camera sensors, radar sensors, lidar sensors and / or ultrasonic sensors (not shown). All perimeter acquisition sensors 1 record data containing information about the surroundings of the vehicle 2. These data are transmitted to and processed by the vehicle's data system 10. The data system can include a control unit, for example a control unit for assisted or automated driving (ADAS or AD, e.g. Automated Driving Control Unit, ADCU), which determines detections from the sensor data and has some knowledge of the vehicle's surroundings and the traffic situation. Based on these detections, in assisted driving, a warning can be issued to the driver or only a limited intervention in the vehicle control can be made, whereas in automated driving, the ADCU can take over the entire control of the vehicle 2.
[0032] FIG. 2 shows a perimeter acquisition sensor 1 , a data system 10 and a separate data unit 20 .
[0033] The data system 10 is electrically connected to at least one surroundings capture sensor 1, e.g. image capture means, inside the vehicle 2. The image capture means can be a front camera of the vehicle. The front camera serves as a sensor for capturing the surroundings in front of the vehicle. Based on the signals or image data of the front camera, the surroundings of the vehicle 2 are detected. Based on these surroundings detections, ADAS or AD functions, such as, for example, lane recognition, lane keeping support, road sign recognition, speed limit assistant, traffic participant recognition, collision warning, emergency braking assistant, distance keeping control, construction site assistant, Autobahn Pilot, cruising driver function and / or autopilot, are provided by the ADAS / AD control unit. The image capture means typically includes an optical system or lens and an image capture sensor, for example a CMOS sensor.
[0034] The data or signals captured by the perimeter capture sensor 1 are transmitted to an input interface 12 of a data system 10. The data are processed in the data system 10 by a data processor 14, which includes a perimeter detection system 16 as well as a monitor system 15. The perimeter detection system 16 can include a first artificial neural network, e.g. a CNN. In addition to pure detection, the perimeter detection system 16 can generate a comprehensive perimeter awareness and situational understanding, e.g. a trajectory prediction of the vehicle 2 as well as of other objects or traffic participants in the periphery of the vehicle 2. The detections of the perimeter detection system 16 can be safety-critical, since the actions and warnings of the ADAS or AD system of the vehicle depend on them. The monitor system 15, on the other hand, is not safety-critical, since its main task is to monitor the perimeter detection system 16 and decide whether to transmit the data to an independent data unit 20. The monitor system 15 can include a second artificial neural network, e.g. an auto-encoder. The data system 10 or data processor 14 may include one or more hardware acceleration means for the artificial neural network so that the artificial neural network can process data in real time within the vehicle.
[0035] If the detection of the monitor system 15 differs from the detection of the surrounding detection system 16, for example exceeding a threshold value, the data system 10 performs wireless transmission of data via the output interface 18 to an independent data unit 20 (cloud, backbone, infrastructure, etc.).
[0036] FIG. 3 shows an embodiment of a classification system and a monitor system as a perimeter detection system.
[0037] The classification system K classifies objects based on, for example, the sensor data X of the peripheral capture sensor 1. In addition to the classification system K, an independent and additional second monitor system K R will also be introduced.
[0038] 3a shows a flow chart of an embodiment of a method for training a machine learning classification system K and a monitoring system. During training, both systems are trained with the same sensor data X. To train a classification system K, which may be, for example, a decision tree learning system, a support vector machine, a learning system based on regression analysis, a Bayesian network, a neural network, or a convolutional neural network, training input data X and training target values Y are provided. Using the classification system K, output data Y' are generated from the training input data X. R The autoencoder creates reproduction data X' that is similar to the training input data X from the training input data X. The goal of training the classification system K is for the output data Y' to be as similar as possible to the training target value Y without overfitting. From the generated output data Y' and the training target values Y, a first error function Loss is used to determine the remaining differences between the output data Y' and the training target values Y. These differences are used to adapt the parameters of the classification system K, for example via backpropagation. This is repeated until a predetermined degree of agreement is achieved or until signs of overfitting appear. Monitor System K R is based on an autoencoder and does not require any further annotations other than the actual sensor data X. R The goal of training is for the reproduced data X' to be as similar as possible to the training input data X. From the reproduced data X' and the training input data X, the remaining differences between the reproduced data X' and the training input data X are determined using a second error function Loss2. These differences are then fed to the monitor system K via, for example, backpropagation. R This is iteratively performed until a predetermined degree of agreement is achieved or until signs of overfitting appear. Monitor System K R It should be noted that the training of x is, for example, performed after the training of the classification system K, but the same input data X must be used. Auto-encoder or monitor system R At any time, the output can be compared to the original sensor data and a quantified error or quantified uncertainty U can be calculated (using a metric). The difference of the reconstructed data X' with respect to the input data X is determined by the metric. The difference value thus determined quantifies the uncertainty U of the output data Y'.
[0039] FIG. 3b shows a classification system K and a monitor system K according to the embodiment of FIG. 3a. R 1 shows a flowchart of a method for using the method of the present invention, which allows the quantification of the uncertainty U of the output data Y′ of the classification system K. This principle involves the use of a monitor system K within the application (inference). R As such, it has the advantage that the metrics can be used to measure the difference of the reproduced data X' with respect to the input signal or data X at any time.
[0040] Not only the classification system K, but also the monitor system K R Since both systems were developed based on the same data, they have similar defects. In the case of classification system K, the defect in output data Y' cannot be measured in the application, but in the case of monitor system K, R It can be identified from the relationship between
[0041] For example, the input data X may be image data of an image, and the output data Y′ may be classification data corresponding to an object depicted on the image. R If a reproduced image X' similar to the original image X can be created using the R This suggests that similar input data X was already present during the training of the classification system K, and therefore the uncertainty U of the output data Y' is low. However, if the reproduced image X' is significantly different from the original image X, the monitor system K R This suggests that similar input data X was not used when training the classification system K, and therefore the uncertainty U of the output data Y' is large. A large uncertainty U suggests that the input data were edge cases.
[0042] 4 shows diagrammatically a vehicle 2 that has captured an unknown scene / situation. An unusual object 42, here an elephant, is present in the capture area 44 of the vehicle's surrounding capture sensors. That is, when using the trained system, it will encounter situations that were not fully taken into account during development because the training data did not include elephants as unusual objects 42. Such extreme situations are a problem for the monitor system K. R4 , for example by an exclamation point "!" (Caution) at 46. The data captured by the perimeter capture sensor is then used to further develop the detection system K. For this purpose, the data can be wirelessly transmitted 48, for example via telematics means, to an external data unit, for example a server or cloud 50. Within the scope of a possible development process, the data stored in the cloud is used in order to be able to be prepared for such extreme cases in the future also with optimized training.
[0043] The system identifies many areas of the training data that are under-represented, some of which may be due to limitations of the surroundings capture sensors in specific driving situations, but may also simply be due to unusual surroundings scenarios.
[0044] Fig. 5 shows a schematic grayscale image of an on-board camera showing overexposure due to incorrect exposure control. Classification of such images, which are too bright and have low contrast, is not successful in many cases. Such situations occur frequently in similar driving situations, for example when exiting a tunnel, due to limitations in exposure control. The reason is the nature of the sensor.
[0045] Figure 6 shows a schematic greyscale image of the unexpected effect of modern art placed on islands between roads (e.g. safety zones). The artwork shown is the "Traffic Light Tree" located at 5TG, Trafalgar Way, Blackwall, London, UK. This modern art piece is made up of a number of traffic lights that do not serve to control traffic. Camera-based traffic light recognition would be unable to attribute any information about traffic control to this piece of art.
[0046] Further aspects and embodiments are described below:
[0047] The transition from assisted to autonomous driving is a technological hurdle. Autonomous systems must be able to handle complex scenarios at will, which may not be covered by simulations or test drives. The goal here is for the surroundings capture for vehicle functions to work reliably at all times and in as many scenarios as possible – ideally, all scenarios.
[0048] To solve this problem, we propose an additive monitoring system that automatically collects this complex and useful data from real road traffic for development, i.e., non-safety-critical systems. However, since the "useful" data may change with software versions, the system also needs to be updatable and the raw data needs to be version-controlled. The details of the total system are as follows:
[0049] A vehicle function that one wishes to implement or monitor is extended by adding an independent second monitoring system based on the same data, i.e., which is also adjusted when an adjustment is made in the vehicle function.
[0050] - During operation, the monitoring system then continuously monitors the coverage of the current situation on the basis of the data covered by the system, and if significant differences are detected, for example via threshold control, the data is captured for development and passed on for further development of vehicle functions. To this end, the monitoring system can evaluate the entire scene captured by one or more sensors or excerpts from the scene, for example individual objects.
[0051] - In addition to the raw data, this data package also contains meta information such as the latest software version, a score on the usefulness of the data calculated by the monitoring system, GPS information, date and time, FIN (Fahrzeug-Identifikations-Number) or connected-vehicle-cloud data. Based on this, the unfolding scenes can be faithfully reproduced (simulation with the same software version). This information acquisition makes it possible to select unknown scenes and edge cases and feed them directly into the development process of vehicle functions. This allows a continuous quality assurance process to be established. With each development step, more and more useful data is fed into the system.
[0052] Furthermore, the maturity of the software can be derived from the number of data transmissions received.
[0053] Expansion to multi-sensor / multi-network setups
[0054] The proposed monitoring system here considers a sensor setup for vehicles in the context of assisted and autonomous driving. It can optionally be extended to a multi-sensor setup. A multi-sensor system has the advantage that it can increase the safety of detection algorithms for road traffic by verifying the detection by several sensors. A multi-sensor system can, for example, be equipped with the following candidates: - one or more cameras, - one or more radars, - one or more ultrasound systems; - one or more riders; - and / or one or more microphones It can be any combination of.
[0055] A multi-sensor system consists of at least two sensors. The data acquisition of one of these sensors s at a time t is denoted hereafter as D_(s,t). These data D can be images and / or audio recordings, angle, distance, speed measurements, and reflections from several objects in the periphery.
[0056] Operation mode
[0057] The described monitoring system (or "non-safety critical system") is implemented as an auto-encoder. It is based on the same data as the detection algorithm or function described above. The auto-encoder can be realized in different ways. One example is to use an auto-encoder in parallel with the detection algorithm, another example is to use it as an additional detection head, i.e. to use a common encoder for detection and auto-encoder. The following system is also conceivable, i.e. a monitoring system that starts after the perimeter detection has been performed. Due to its working principle, autoencoders offer significant advantages over other approaches. Let us illustrate this using the example of image data:
[0058] If an image is input to an autoencoder, the encoder tries to reproduce it at the output. Autoencoders can be trained using only the input signal, without additional labels. This not only has the advantage that no additional effort is required in terms of annotation, but also offers the crucial advantage that the error associated with the input data can be quantified at any time (even for unknown data). However, a suitable thresholding is recommended. Autoencoders can be used on the entire image or on a small excerpt.
[0059] Thus, the autoencoder can measure its own error for each possible input signal. In machine learning methods, higher error values are generally expected for input data that do not match the training data well. Since the autoencoder and the detector are based on the same data, the following advantages are obtained: unknown or uncertain scenarios identified by the autoencoder imply that such scenarios are not well represented in the training data, and a broad coverage of traffic scenarios is useful for the feature development.
[0060] Extension with uncertainty measures
[0061] In addition to the output of the autoencoder estimating the network's reconstruction certainty, it is also useful to have a measure of the certainty on which the autoencoder makes its decisions. This so-called certainty measure complements the autoencoder output and is especially useful when considering the fusion of autoencoder outputs across different sensors and / or over time. Such a measure of certainty can be calculated using statistical calibration or uncertainty estimates. Suitable for: a) A statistical calibration method, where the output of the auto-encoder is weighted to represent the probabilistic uncertainty. This method is suitable when computational resources are very scarce. b) If sufficient computational resources are available, the model uncertainty can be estimated: for example, by augmenting each individual network layer using Monte Carlo dropout. Then, at runtime, this and the following network layers are iterated m times with randomized activations of each neural layer, resulting in a set of outputs y_i, i=1…m. c) Additionally, the measurement uncertainty is also estimated: for example by adding a regularization to the error function (Loss; Loss2) that measures the measurement uncertainty at run-time. The extension to unknown or uncertain scenes with uncertainty estimation allows the evaluation of scenes where the classifier makes a correct decision but there is a large uncertainty in that decision. Adding uncertainty estimation extends the robustness of unknown scene search by adding uncertainty scene search. Such an architecture allows the network to select not only unknown scenes but also scenes that are uncertain to the network.
[0062] Extension with temporal uniqueness
[0063] As mentioned above, autoencoders provide a way to identify data that was under-represented in the training data. Additional value can be gained by combining this principle with the idea of temporal consistency. For example, filtering the identified data over time to distinguish between continuously occurring samples and infrequent outliers can provide valuable additional information. Short-term outliers can suggest sensor-related factors, for example. In other words, when entering a tunnel, the white balance of the camera may fluctuate significantly at a certain time t. These data will be useful for the development of the respective sensors. If the identified data appears continuous, this is likely an unknown scenario and is useful for algorithm development.
[0064] The above-described embodiments and aspects provide the following advantages: -Special advantage: Optimized for runtime -Can be applied to any sensor data and their combinations -Easy to implement in development process, no separate data collection required -Comprehensive coverage compared to event-based approaches that only monitor the output of vehicle functions, e.g. lost object tracks -Automatically selects useful data specific to the algorithm being observed -Identify realistic scenarios that are difficult to reproduce (accidents, unlikely combinations of events...)
Claims
1. A vehicle data system (10) including a surroundings detection system (16) and a monitoring system (15) within a vehicle (2), comprising: a perimeter detection system (16) configured to receive input data X of the perimeter acquisition sensor (1), evaluate said input data X using a trained first artificial neural network K, and output perimeter detection data Y' as a result of said evaluation; The monitor system (15) transmits the same input data X of the peripheral acquisition sensor (1) to a second trained artificial neural network K R As a result, the reproduced data X' can be output. If the monitoring system detects a difference in the reproduced data X' relative to the input data X that exceeds a threshold, the monitoring system (15) prompts the transfer of the input data X to a separate data unit (20). A vehicle data system comprising:
2. Second artificial neural network K R 2. The vehicle data system (10) of claim 1, wherein the is an auto-encoder.
3. 3. The vehicle data system (10) of claim 1 or 2, wherein the monitor system (15) calculates a score estimating the usefulness of the input data X based on the difference.
4. 10. The vehicle data system (10) of claim 1, wherein a first artificial neural network K is trained using predetermined training data, and reference output data Y_1, Y_2, ..., Y_n are used for input data X_1, X_2, ..., X_n, respectively, and a first error function (Loss) indicating the difference between the output of the first neural network K for the input data X_1, X_2, ..., X_n and the corresponding reference output data Y_1, Y_2, ..., Y_n is minimized by adjusting weights of the first neural network K.
5. Second artificial neural network K R But the second artificial neural network K R A second error function (Loss 2 2. A vehicle data system (10) according to any one of the preceding claims, characterized in that:
6. 10. A vehicle data system (10) according to any one of the preceding claims, characterized in that in addition to the input data X of the perimeter capture sensor (1), additional meta-information is also transmitted, said meta-information corresponding to one or more of the following group of information: latest software version, calculated score of the monitoring system, GPS data, date, time, vehicle identification number (FIN = Fahrzeug-Identifications-Number) and cloud data making it possible to grasp the scene and / or vehicle situation.
7. A vehicle data system (10) according to any one of the preceding claims, characterized in that the periphery detection system (16) is configured to receive input data X from a plurality of periphery acquisition sensors (1) and to evaluate them collectively.
8. 10. The vehicle data system (10) of claim 1, wherein the monitor system (15) is configured to process input data X of the perimeter acquisition sensor (1) in parallel with the perimeter detection system (16).
9. 10. The vehicle data system (10) of claim 1, wherein the monitor system (15) is integrated into the periphery detection system (16) as an additional detection head, and the monitor system (15) and the periphery detection system (16) use a common encoder.
10. A vehicle data system (10) as claimed in any one of the preceding claims, characterized in that the input data of the surroundings capture sensor (1) is image data and the monitor system (15) is configured to reproduce the entire image or an excerpt of a part of the image.
11. 11. A vehicle data system (10) as claimed in any one of the preceding claims, wherein the monitoring system (15) is configured to determine and output an uncertainty measure, the uncertainty measure being a value giving a degree of certainty that the monitoring system (15) is in outputting the reconstruction data X'.
12. 10. The vehicle data system (10) of any one of the preceding claims, wherein the monitoring system (15) is configured to take into account the consistency of the reconstruction data X' over time.
13. 13. The vehicle data system of claim 12, further comprising a distinguishing feature as to whether the difference between the reconstruction data X' of the monitor system (15) and the input data X is continuous or only appears for a limited period of time.
14. 10. A vehicle data system according to any one of the preceding claims, characterised in that both the surrounding detection system (16) and the monitoring system (15) are adapted for over the air updates.
15. A method for determining vehicle data worth transmitting from a perimeter capture sensor (1), comprising the steps of: evaluating the input data X of the perimeter acquisition sensor (1) by a perimeter detection system (16) implemented using a first trained artificial neural network; outputting the perimeter detection data Y′ determined during the evaluation; A second trained artificial neural network K R evaluating the same input data X of the perimeter acquisition sensor (1) by a monitoring system (15) implemented using outputting the reconstruction data X' determined during the evaluation; and A step of instructing transmission of the input data X to the independent data unit (20) when a difference between the reproduced data X' and the input data X that exceeds a threshold is identified.
Citation Information
Patent Citations
Learning server, and assist system
JP2019021201A
Lane line detection method and apparatus, driving control method and apparatus, and electronic device
WO2020098708A1
System and method for obtaining training data
WO2020056331A1
