System and method for processing information signals

Abstract

Some examples relate to a system 10 for processing information signals 11, 12, comprising a signal input 13 for receiving the information signals 11, 12 and a common unit 14 of the neural network of the system 10. The common unit 14 of the neural network is configured for a first signal processing step of each of the information signals 11, 12. Furthermore, at least two splitting units 15a, 15b of the neural network are provided downstream of the common unit 14 in the signal flow. A first unit 15a of the at least two splitting units 15a, 15b is configured for a second signal processing step of the first information signal 11, and a second unit 15b of the at least two splitting units 15a, 15b is configured for a second signal processing step of the second information signal 12. Furthermore, a sensor system, a vehicle 40, and a method 30 for signal processing are also proposed.

Description

[Technical Field]

[0001] The present invention relates to a system for processing at least one first and at least one second information signal, and to a system for processing an information signal. Further embodiments relate to a sensor system and to a motor vehicle. Furthermore, a method for signal processing of an information signal is proposed. [Background technology]

[0002] Modern automobiles have to process many information signals, for example, signals from different types of sensors, which can be used for various vehicle functions.

[0003] Signal processing using artificial intelligence is known, for example. In particular, the concepts of machine learning (e.g., supervised and unsupervised for clustering) and convolutional neural networks (CNNs) are known in the field of computer vision, for example, for object recognition, object classification and segmentation. Neural networks can be implemented, for example, in hardware.

[0004] CNN hardware accelerators are known, for example, which can have a positive effect on processing speed, in contrast to software-based CNN systems. One goal may be to achieve as much computing power as possible while minimizing the load on hardware resources. In this way, for example, a desired performance (e.g., a desired number of processed images per second) can be achieved. Furthermore, for example, the processor area and therefore the cost can be reduced.

[0005] However, in known systems, especially in the field of vehicles, the energy demands for high functional requirements may still be too high. Hardware-implemented neural networks, for example, cannot be used very flexibly, since they cannot be trained further. Summary of the Invention [Problem to be solved by the invention]

[0006] The objective of this disclosure is to provide a better understanding of systems with neural networks. [Means for solving the problem]

[0007] This problem is solved by the subject matter of the independent claims. Further advantageous embodiments are described in the dependent patent claims, in the description and in the drawings.

[0008] A system for processing at least one first information signal and at least one second information signal is therefore proposed. The system comprises a signal input for receiving the information signals and a common unit of the neural network of the system, where the common unit of the neural network is configured for a first signal processing step for each of a plurality of information signals. The system further comprises at least two splitting units of the neural network arranged downstream of the common unit in the signal flow, where a first of the at least two splitting units is configured for a second signal processing step for the first information signal and a second of the at least two splitting units is configured for a second signal processing step for the second information signal. A signal output of the system is configured for outputting the processed information signals.

[0009] The system thus comprises a plurality of separate units of the neural network. The proposed system can make signal processing more efficient. A common unit of the neural network can thus be used to process both the first and second information signals. The first signal processing step of the commonly used unit can, for example, be suitable for both the first and second information signals.

[0010] After the first signal processing step, the system provides the opportunity to select whether the signal should be further processed by the first or second of the two division units in the second signal processing step, thereby making it possible to meet, for example, different requirements for processing different information signals, for example, the first division unit may be configured to process signals differently from the second division unit.

[0011] In this way, it is possible to jointly use (share) portions of the neural network (e.g., share layers of the neural network), thereby eliminating the need to provide duplicated portions as in other systems. For example, general signal processing can be performed by a common unit, while more specialized signal processing steps can be performed by the division units (e.g., two or more division units, e.g., at least three or at least four division units, to increase the flexibility of the system). For example, the common unit associated with the first unit of both division units can be considered a first neural network of the system, and the common unit associated with the second unit of both division units can be considered a second neural network of the system.

[0012] The two information signals can be, for example, two different signals from, for example, at least two different signaling devices. Such signaling devices can also be, for example, sensors such as cameras. The system can, for example, perform general signal processing steps of camera images in a first common part of the neural network, and allow more specific steps (for example, specific functions, e.g., specialized processing suited to the type of camera used) to be performed later in separate parts of the neural network.

[0013] For example, the at least two splitting units of the neural network may include at least one software-implemented neural network unit and at least one hardware-implemented neural network unit. This allows the system to simultaneously utilize the advantages of both software and hardware implementations. While a hardware implementation may reduce flexibility in signal processing, it may advantageously enable faster processing speeds and / or lower energy consumption during processing. In contrast, a software implementation may allow for greater flexibility, even later, during use of the system, by reprogramming processing parameters (e.g., neural network weights). For example, the first splitting unit can be used for a fixed, pre-defined function, while the second splitting unit can be used when new functions need to be adapted.

[0014] For example, at least two splitting units of the neural network may be arranged in parallel in the signal flow, so that before two information signals are processed by the at least two splitting units in the second signal processing, it is possible to select which of the splitting units to use. This selection can be made, for example, depending on the type of signal information and / or the function performed by the system. The selection of the processing path can be controlled, for example, depending on which signal source (e.g., which sensor, e.g., which camera) the first or second information signal comes from.

[0015] For example, the common unit of a neural network may be implemented in hardware. For example, a hardware implementation of the common unit can be efficiently used for more general signal processing steps (e.g., pre-processing of signals by the neural network, e.g., extraction of general features from information signals). This type of signal processing step may be required, for example, for signals from different sensors before more specific, sensor-specific signal processing steps can be performed.

[0016] For example, a common unit of a neural network may include a Convolutional Neural Network (abbreviated as CNN; e.g., a deep convolutional neural network). A Convolutional Neural Network may have one or more convolutional layers, which may be followed by, for example, a pooling layer. A CNN may be used, for example, to perform a classification function (e.g., extract features from an image), e.g., for image or speech recognition.

[0017] For example, the common unit of the neural network may comprise an autoencoder or part of an autoencoder. The autoencoder can be used to enable more efficient coding. The purpose of an autoencoder may also be to reduce data, for example, an autoencoder can be used for dimensionality reduction. The autoencoder may have an input layer and at least one other layer that is significantly smaller than the input layer (the layer that forms the coding; e.g., the encoding layer). The encoding layer can be used to output the signal processed in the first signal processing step from the common unit. In this way, the system may be used for data compression (e.g., compressed sensor data).

[0018] According to one embodiment, the system may further comprise a pre-processing unit (e.g., a pre-processing unit). In this case, the pre-processing unit may be arranged in the signal flow between the signal input and the common unit of the neural network. The pre-processing unit is configured, in particular, to allocate the information signals with respective processing times for the first signal processing step. For example, since the common unit is used for signal processing of the first and second information signals (and further signals, e.g., from two separate signal sources), scheduling (e.g., when the first signal can be processed and when, e.g., subsequently, the second signal can be processed) must be provided. For example, the information signals may be processed sequentially (e.g., in the order in which they arrive at the signal input of the system, e.g., by processing signals from different signal sources alternately). Alternatively or additionally, prioritization of the information signals may be taken into account. For example, it may be sensible to process more important signal types first before the common unit of the neural network, even if there are still other lower-priority signals to be processed in the queue. For example, two cameras on a vehicle may capture images into the system (e.g., a first information signal from a first camera and a second information signal from a second camera). In this case, for example, one of the cameras may be used for a more safety-critical task than the other of the cameras. Priority may be given to processing the information signal of this camera for the more safety-critical task.

[0019] The pre-processing unit may, for example, be configured to convert the information signals into a signal standard that is compatible with a common unit of the neural network, thus enabling the adaptation of signals from different signal sources (e.g., different cameras with different resolutions and / or different frame rates) so that all signals from the different signal sources (e.g., sensors) can be uniformly processed using the common unit.

[0020] For example, the first and second information signals may suitably both be image signals, and the pre-processing unit may be configured to convert the image signals into respective standard image signals of a predetermined frame rate and / or a predetermined resolution.

[0021] For example, the first and second information signals may both be image signals, and the pre-processing unit may be configured to select only a predetermined selection of information from the image signals for signal processing by the neural network, based on the function for which the respective image signals are used. For example, instead of processing all image frames of the information signals, only a portion of them (e.g., every second or every fifth frame) may be processed. For example, for more safety-critical functions, image frames may be analyzed more frequently (e.g., at a higher frequency, e.g., more image frames) than for less safety-critical functions. For example, for less critical functions, analysis may be performed only every fifth (or ten or twenty) frame of the image signal. For example, for a function to detect driver drowsiness, a relatively low image recognition rate may be sufficient, whereas for functions such as pedestrian recognition, image frames must be analyzed more frequently (e.g., every two frames). The frequency of the analyzed frames may be selected depending on the image frequency of the camera.

[0022] For example, the pre-processing unit may be configured to select, depending on the information signal present at its signal input, whether a first or second sub-unit of the neural network is used for the second signal processing. For example, the information signal may be provided with an identifier indicating the source from which the information signal originates. This identifier can then be used to select, in the second signal processing, whether the information signal is processed using the first or second sub-unit of the neural network.

[0023] An aspect of the present disclosure further relates to a system for processing an information signal, the system having a signal input for receiving the information signal and a common unit of a neural network of the system, the common unit of the neural network being configured for a first signal processing step of the information signal.

[0024] The system further comprises at least two splitting units of a neural network arranged in parallel downstream of the common unit in the signal flow, wherein the system is configured to use a first of the at least two splitting units for the second signal processing of the information signal in a first operating mode and to use a second of the at least two splitting units for the second signal processing of the information signal in a second operating mode, and a signal output of the system is configured to output the processed information signal.

[0025] The proposed system advantageously allows for the use of common units for signal processing steps required for different functions (e.g. for the first and second operating modes), which, in contrast to other systems, allows for the omission of additional neural network layers, since these layers do not need to be duplicated.

[0026] In contrast, at least two of the division units of the neural network (e.g., the final layer of the neural network) that enable subsequent signal processing can be specialized for each function. For example, at least one of the division units can be implemented in software (e.g., trainable after implementation, as opposed to being implemented in hardware). For example, a trainable final layer can be used when a first function is to be realized using signals from a certain sensor, and another hardware final layer (e.g., a second division unit of the neural network) can be used when a second function (e.g., various types of detection) is to be realized using signals from the same sensor.

[0027] For example, a system may be provided in which both the first dividing unit and the second dividing unit are used in parallel for the second signal processing in order to implement the first and second operating modes in parallel.

[0028] One aspect relates to a sensor system, comprising any of the systems described above or below. The sensor system further comprises at least two sensors connected to a signal input of the system, wherein the at least two sensors are configured to provide different information signals to the signal input. Advantageously, for example, a common unit of a neural network of the system and a first unit of the dividing units can be used to process the information signal of a first sensor, and a common unit of a neural network of the system and a second unit of the dividing units can be used to process the information signal of a second sensor. Sharing common parts of the neural network allows for an efficient provision of the sensor system (e.g., lower cost; e.g., smaller size).

[0029] For example, the at least two sensors may include at least one of an optical sensor, a camera, a current sensor, a temperature sensor, a state-of-charge sensor, or a sensor device configured to predict a regeneration event. Separating different elements of the system's neural network advantageously allows it to be used efficiently for many applications (e.g., classifying features from images; e.g., reducing sensor data for efficient subsequent processing).

[0030] One aspect relates to a vehicle having a system according to any of the systems described above or below, the system being deployed in a control unit of the vehicle, the vehicle further comprising at least two sensors connected to signal inputs of the system via an on-board network of the vehicle.

[0031] In particular, an ever-increasing number of sensors, such as cameras, are used in automobiles. The proposed system can enable neural networks or layers that are at least partially shared among multiple sensors, rather than requiring each sensor to have its own complete neural network. This can enable more efficient signal processing within the vehicle.

[0032] One aspect relates to a method for processing an information signal, wherein a first step involves processing the information signal by an upstream unit of a neural network in a first signal processing step, a second step involves identifying the type of signal information and / or the function for which the information signal is used, and a third step involves further processing the information signal by a first downstream unit of the neural network or a second downstream unit of the neural network in a second signal processing step, wherein the first downstream unit or the second downstream unit of the neural network is selected depending on the type of information signal and / or the type of function used for signal processing.

[0033] For example, one of the proposed systems can be used in an efficient manner to implement the proposed method. The use of neural network splitting units allows for flexible selection of the units of signal processing that are best suited to the signal to be processed and / or the function assigned to the signal.

[0034] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings. [Brief explanation of the drawings]

[0035] [Figure 1] FIG. 1 shows a schematic example of a system for processing first and second information signals. [Figure 2] FIG. 1 shows a schematic example of a system for processing an information signal using a first unit or a second unit downstream of a neural network. [Figure 3]FIG. 1 shows a schematic example of a method for processing an information signal. [Figure 4] FIG. 1 is a diagram showing an example of a sensor system equipped with two cameras in a vehicle. DETAILED DESCRIPTION OF THE INVENTION

[0036] Various embodiments will now be described in more detail with reference to the accompanying drawings, in which some embodiments are shown. In the figures, the thickness of lines, layers, and / or regions may be exaggerated for clarity. In the following description of the accompanying drawings, which show only some illustrative embodiments, the same reference numerals may indicate the same or equivalent components.

[0037] An element that is described as being "connected" or "coupled" to another element may be directly connected or coupled to the other element, or there may be elements between them. Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as understood by one of ordinary skill in the art to which the embodiment belongs.

[0038] The increasing use of artificial intelligence (AI) or machine learning applications in vehicles may also lead to increased energy demands. One example of this would be autonomous driving, where a large number of sensors must be analyzed, for example, in camera image analysis using machine learning, which plays a key role in object recognition, object classification and / or segmentation.

[0039] The various camera systems within a vehicle run algorithms, for example using convolutional neural networks (CNNs), which, if combined with other approaches such as distributing the workload across multiple controllers and / or simply implementing it in software, for example, would increase energy demands and costs. Providing the energy to run AI algorithms can be an obstacle to the introduction of energy-efficient autonomous driving. Furthermore, other approaches may require a large number of additional controllers, which can increase hardware costs.

[0040] Below we propose ideas that may make it possible to reduce energy demand and / or hardware costs. The illustrated example of autonomous driving should be considered here as exemplary only. The proposed ideas may be beneficial for many applications. Worth mentioning here are applications related to implementations in the field of energy management using artificial intelligence, for example.

[0041] 1 shows a schematic example of a system 10 for processing first and second information signals 11, 12. The system 10 has a signal input 13 for receiving the information signals 11, 12.

[0042] The system 10 further comprises a neural network for data processing. A common unit 14 of the neural network of the system 10 is configured for a first signal processing step for each of a plurality of information signals 11, 12. For example, the information signals 11, 12 are received in sequence at a signal input 13. The common unit 14 can then perform a first signal processing step for the first signal 11, followed by a first signal processing step for the second signal 12. Alternatively, an optional pre-processing unit (see also FIG. 4 ) may provide scheduling for the processing of the input signals by the common unit 14.

[0043] The system 10 comprises at least two splitting units 15a, 15b of a neural network arranged downstream of the common unit 14 in the signal flow, a first unit 15a of the at least two splitting units 15a, 15b being configured for a second signal processing of a first information signal 11 and a second unit 15b of the at least two splitting units 15a, 15b being configured for a second signal processing of a second information signal 12. The information signals 11', 12' processed by the system 10 can be output at a signal output 16 of the system 10.

[0044] This example shows how several layers can be efficiently shared to process different signals in a system 10 with a neural network. For example, a first signal 11 can come from a first sensor, and a second signal 12 can come from a second sensor. A common unit 14 can be provided to perform processing steps for both signal types. This makes it possible to avoid duplicating layers of the neural network formed in the common unit 14, as opposed to separate systems for each of the sensors.

[0045] For example, the optional control unit can determine which sensor generated the input signal being processed at each time and can consistently control which of the two splitting units 15a, 15b of the neural network is used to process the respective signal in the second processing step.

[0046] Further details and aspects are set forth in the context of the above or below described embodiments. The embodiment shown in Figure 1 may include one or more optional additional features corresponding to one or more aspects set forth in the context of the proposed concepts or in the context of one or more above or below described (e.g., Figures 2-4) embodiments.

[0047] FIG. 2 shows a schematic example of a system 20 for processing an information signal using a first unit 25a or a second unit 25b of a neural network of the system 20.

[0048] The system 20 here has a signal input 23 for receiving information signals 21, e.g. signals from a sensor such as a camera, etc. In particular, a plurality of sensor signals can be received and processed, for example, for use in various functions (e.g. pedestrian detection, traffic sign detection, vehicle surroundings and / or vehicle state detection, occupant detection).

[0049] The common unit 24 of the neural network of the system 20 is formed for a first signal processing step of the information signal 21 .

[0050] The system 20 further comprises at least two splitting units 25a, 25b of a neural network arranged in parallel downstream of the common unit 24 in the signal flow, wherein the system 20 is configured to use a first one 25a of the at least two splitting units 25a, 25b for the second signal processing of the information signal 21 in a first operating mode and to use a second one 25b of the at least two splitting units 25a, 25b for the second signal processing of the information signal 21 in a second operating mode. The processed information signal 21' can be output via a signal output 26.

[0051] For example, a sensor signal may need to be processed in the same way (e.g., a basic image processing step) for both the first and second functions of system 20 (or even for other functions). This signal processing step can be efficiently performed for both functions by common unit 24 of the neural network. In contrast, the functions may have different specific signal processing requirements. Thus, a first unit 25a of the splitting unit may, for example, be provided to initiate or complete signal processing for the first function, and a first unit 25b of splitting units 25a, 25b may, for example, be provided to initiate or complete signal processing for the second function. Again, layers of the neural network in common unit 24 (e.g., layers shared for both functions) may be used efficiently.

[0052] Further details and aspects are set forth in the context of the above or below described embodiments. The embodiment shown in Figure 2 may include one or more optional additional features corresponding to one or more aspects set forth in the context of the proposed concepts or in the context of one or more above (e.g., Figure 1) or below (e.g., Figures 3-4) described embodiments.

[0053] 3 shows a schematic example of a method 30 for processing an information signal. The method 30 comprises processing the information signal by a unit upstream of the neural network in a first signal processing step 31. The method 30 further comprises identifying 32 the type of signal information (e.g. the signal source that generated the information signal) and / or the function for which the information signal is to be used.

[0054] The method 30 further comprises further processing 33 of the information signal by a first unit 15a downstream of the neural network or by a second unit 15b downstream of the neural network in a second signal processing step. The choice of whether to use both units 15a, 15b downstream of the neural network for the second signal processing step depends on the type of signal information and / or the type of function used to process the signal.

[0055] Further details and aspects are set forth in the context of the above or below described embodiments. The embodiment shown in Figure 3 may include one or more optional additional features corresponding to one or more aspects set forth in the context of the proposed concepts or in the context of one or more above (e.g., Figures 1-2) or below (e.g., Figure 4) described embodiments.

[0056] FIG. 4 shows an example of a sensor system with two cameras 42a, 42b (for example, conventional sensors) in a vehicle 40, in particular in an automobile, that is made for (partially) automated driving.

[0057] The vehicle 40 has a system such as those described above or below installed in a control device 40a of the vehicle 40. The vehicle 40 further has at least two sensors 42a, 42b connected to signal inputs of the system via an on-board network 41 of the vehicle 40.

[0058] The cameras 42a, 42b can include, for example, a front camera, a rearview camera, a side camera, and / or an interior camera of the vehicle 40. The in-vehicle network 41 can include a communication network within the vehicle 40, for example, an Ethernet network, such as a LIN bus or a CAN bus. The cameras 42a, 42b can, for example, attach an identifier to the information signal they transmit, so that information about which camera 42a, 42b the currently received information signal came from is available to the system. Alternatively, the signal input unit of the system can be configured to recognize which camera 42a, 42b sent the information signal and attach the appropriate identifier to it.

[0059] The control device 40a is provided with a pre-processing unit 43 between the signal input and the common unit 44 of the neural network of the control device. This unit can, for example, comprise a hardware-based signal processing unit. The pre-processing unit 43 can be used to normalize the image signals of the different cameras to enable signal processing in a common system. For example, the images processed by the pre-processing unit 43 can be normalized and of an appropriate size (e.g., resolution) and sent to the common unit 44 of the neural network as messages with identifiers for each camera 42a, 42b.

[0060] The neural network common unit 44 can have neural network layers that are shared by (e.g., shared between) the different cameras 42 a, 42 b. These shared (e.g., convolutional neural network) layers can be implemented in hardware. The output signal of the common unit 44 can be a feature map from a CNN or a flattened vector with an identifier for each camera 42 a, 42 b.

[0061] The identifier allows for control (eg by a control unit not shown) which of the at least two division units 45a, 45b of the neural network is used for further signal processing (eg second signal processing).

[0062] For example, the first unit 45a of the two (or multiple) segmentation units 45a, 45b may then be implemented in software so that it can be trained. The first unit 45a may, for example, contain models of downstream layers of neural networks (e.g., CNNs and fully connected neural networks) for different tasks. The signals thus processed can be used for segmentation, detection and / or classification.

[0063] For example, the second unit 45b of the two (or multiple) segmentation units 45a, 45b may be implemented in software so that it cannot be trained. The second unit 45b may also include a model of a downstream layer of a neural network (e.g., CNN and fully connected) for a different task. The signals processed in this way can be used for segmentation, detection, and / or classification.

[0064] The splitting units 45a, 45b can each form a complete neural network in combination with the common unit 44, for example. Sharing layers of a neural network can improve efficiency. This may be particularly true when multiple (e.g., three, four, or at least five) splitting units (e.g., including the final layer of the neural network) each share a layer of the common unit 44. For example, the common unit can have more layers than at least one of the multiple splitting units. For example, two splitting units can have different numbers of layers. In this way, the system can be used, for example, to process multiple different sensors (e.g., cameras) and / or use different functions. For example, using four splitting units for a second signal processing for one sensor can realize four different functions.

[0065] A typical use case of the system is described below. Different camera systems record images, and algorithms such as object recognition, object classification, and segmentation can be executed on these images. Because some tasks require redundant or similar algorithms, for example, a dedicated controller with a neural network for person detection (front camera, interior camera) does not exist. The convolutional layers of a CNN can be considered a feature extraction method. In the proposed concept, the convolutional layers of a trained CNN are implemented in hardware (e.g., a common unit of a neural network) within the controller. All tasks, such as object recognition, object classification, and segmentation, can utilize these shared convolutional layers (e.g., a common unit of a neural network) and connect task-specific convolutional and fully connected neural networks (e.g., a segmentation unit of a neural network) to them. These can be implemented in software (e.g., trainable) or hardware (e.g., non-trainable). When people need to be detected and this needs to be done by both the indoor camera and the front camera, the advantage of a lean approach with a flexible performance trade-off becomes apparent (e.g. common units of the neural network can be used in signal processing for a common task, e.g. a pre-processing unit can enable time sharing of signal processing by common units of the neural network).

[0066] The described idea of shared neural network layers is not limited to image processing and is used here for illustrative purposes only. For example, the proposed idea can also be used in energy management, for example, by sharing autoencoder layers, and for dimensionality reduction (e.g., data reduction).

[0067] Further details and aspects are set forth in the context of the embodiments described above or below. The embodiment shown in Figure 4 may include one or more optional additional features corresponding to one or more aspects described in the context of the proposed concepts or in the context of one or more of the embodiments described above (e.g., Figures 1-3).

[0068] Examples relate to methods and systems for implementing machine learning in an energy-efficient manner, e.g., for image processing in autonomous driving, while simultaneously reducing costs. By using both shared layers and separate neural network layers (e.g., for different signal types or applications) in a single system, efficiency gains can be achieved compared to conventional systems.

[0069] The proposed aspects enable high energy efficiency due to the reduced number of controllers and hardware implementation of convolutional layers, and also reduce costs due to the reduced number of controllers and shared convolutional layers. This application relates to the invention described in the claims, but the disclosure of this application also includes the following: 1. A system (10) for processing at least one first information signal (11) and at least one second information signal (12), comprising: a signal input unit (13) for receiving the information signals (11, 12); a common unit (14) of the neural network of the system (10), the common unit (14) of the neural network being configured for a first signal processing step of each of the information signals (11, 12); at least two division units (15a, 15b) of a neural network arranged downstream of the common unit (14) in the signal flow, wherein a first unit (15a) of the at least two division units (15a, 15b) is configured for a second signal processing of a first information signal (11) and a second unit (15b) of the at least two division units (15a, 15b) is configured for a second signal processing of a second information signal (12); a signal output unit (16) for outputting the processed information signals (11', 12'); A system (10) having: 2. The at least two splitting units (15a, 15b) of the neural network include at least one unit of the neural network implemented by software and at least one unit of the neural network implemented by hardware. 1. The system (10) according to claim 1. 3. At least two division units (15a, 15b) of the neural network are arranged in parallel in the signal flow. 3. The system (10) according to claim 1 or 2. 4. The common unit (14) of the neural network is formed by hardware. A system (10) according to any one of 1 to 3 above. 5. The common unit (14) of the neural network has a convolutional neural network. 5. A system (10) according to any one of 1 to 4 above. 6. The common unit (14) of the neural network has an autoencoder 6. A system (10) according to any one of 1 to 5 above. 7. 7. A system (10) according to any one of 1 to 6 above, further comprising a pre-treatment unit (43); The pre-processing unit (43) is arranged in the signal flow between the signal input section (13) and the common unit (14) of the neural network, The pre-processing unit (43) is configured to allocate to said information signals (11, 12) respective processing times for a first signal processing step by a common unit (14) of the neural network. System (10). 8. The pre-processing unit (43) is configured to convert the information signals (11, 12) into a signal standard compatible with the common unit (14) of the neural network. 7. The system (10) according to claim 7. 9. The first information signal (11) and the second information signal (12) are both image signals, and the pre-processing unit (43) is configured to convert the image signals into respective standard image signals of a predetermined frame rate and / or a predetermined resolution. 10. The system (10) according to claim 8. 10. The first information signal (11) and the second information signal (12) are both image signals, and the pre-processing unit (43) is configured to select only information from a predetermined selection of the image signals for signal processing by the neural network, based on the function of the object for which the respective image signal is used. The system (10) according to claim 8 or 9. 11. The pre-processing unit (43) is configured to select, depending on the information signals (11, 12) present at the signal input (13), whether to use the first unit (15a) or the second unit (15b) of the two division units (15a, 15b) of the neural network for the second signal processing. A system (10) according to any one of 7 to 10 above. 12. A system (20) for processing an information signal (21), comprising: a signal input unit (23) for receiving an information signal (21); a common unit (24) of the neural network of the system (20), the common unit (24) of the neural network being configured for a first signal processing step of the information signal (21); and at least two neural network splitting units (25a, 25b) arranged in parallel downstream of the common unit (24) in the signal flow, A system (20) configured to use a first unit (25a) of at least two division units (25a, 25b) for second signal processing of an information signal (21) in a first operating mode and to use a second unit (25b) of the at least two division units (25a, 25b) for second signal processing of the information signal (21), Furthermore, a signal output unit (26) for outputting the processed information signal (21') is provided. A system (20) having: 13. 13. The system (20) according to claim 12, A system (20) using both the first dividing unit (25a) and the second dividing unit (25b) in parallel for second signal processing to implement the first and second operating modes in parallel. 14. A system (10, 20) according to any one of 1 to 13 above; and at least two sensors (42a, 42b) connected to the signal inputs (13, 23) of the system (10, 20), the at least two sensors (42a, 42b) being configured to send different information signals to the signal inputs (13, 23). Sensor system. 15. The at least two sensors (42a, 42b) include at least one of an optical sensor, a camera (42a, 42b), a current sensor, a temperature sensor, a state-of-charge sensor, or a sensor device configured to predict a regeneration event. 15. The sensor system according to claim 14. 16. a system (10, 20) according to any one of 1 to 13 above, which is installed in a control device (40a) of a vehicle (40); and at least two sensors (42a, 42b) connected to the signal inputs (13, 23) of the system (10, 20) via an in-vehicle network (41) of the vehicle (40). Automobiles (40). 17. A method (30) for signal processing at least one information signal (11, 12), comprising: In a first signal processing step, the information signal (11, 12) is processed (31) by an upstream unit (14) of the neural network; Identifying (32) the type of information signal (11, 12) and / or the function for which the information signal (11, 12) is used; Further processing (33) the information signals (11, 12) by a first unit (15a) downstream of the neural network or by a second unit (15b) downstream of the neural network in a second signal processing step; The selection of the first downstream unit (15a) or the second downstream unit (15b) of the neural network is then made depending on the type of information signal (11, 12) and / or the type of function used to process the signal. A method (30) comprising:

Claims

1. A system (10) for processing at least one first information signal (11) and at least one second information signal (12), comprising: a signal input unit (13) for receiving the information signals (11, 12) in sequence; a common unit (14) of the neural network of the system (10), which is formed for a first signal processing step of the information signals (11, 12), respectively; at least two division units (15a, 15b) of a neural network arranged downstream of the common unit (14) in the signal flow, wherein a first unit (15a) of the at least two division units (15a, 15b) is configured for a second signal processing of a first information signal (11) and a second unit (15b) of the at least two division units (15a, 15b) is configured for a second signal processing of a second information signal (12); a signal output unit (16) for outputting the processed information signals (11', 12'); and a common unit (14) performing a first signal processing step for the first information signal (11) and subsequently performing a first signal processing step for the second information signal (12); A system (10) in which either the first unit (15a) or the second unit (15b) of the dividing units (15a, 15b) is selected depending on the type of the information signal (11, 12) processed by the common unit (14) and / or the function assigned to the information signal (11, 12).

2. The at least two splitting units (15a, 15b) of the neural network include at least one unit of the neural network implemented by software and at least one unit of the neural network implemented by hardware. The system (10) of claim 1.

3. At least two division units (15a, 15b) of the neural network are arranged in parallel in the signal flow. A system (10) according to claim 1 or 2.

4. The common unit (14) of the neural network is formed by hardware. A system (10) according to any one of claims 1 to 3.

5. The common unit (14) of the neural network comprises a convolutional neural network. A system (10) according to any one of claims 1 to 4.

6. The common unit (14) of the neural network comprises an autoencoder A system (10) according to any one of claims 1 to 5.

7. A system (10) according to any one of claims 1 to 6, further comprising a pre-treatment unit (43), The pre-processing unit (43) is arranged in the signal flow between the signal input section (13) and the common unit (14) of the neural network, The pre-processing unit (43) is configured to allocate to said information signals (11, 12) respective processing times for a first signal processing step by a common unit (14) of the neural network. System (10).

8. The pre-processing unit (43) is configured to convert the information signals (11, 12) into a signal standard compatible with the common unit (14) of the neural network. The system (10) of claim 7.

9. The first information signal (11) and the second information signal (12) are both image signals, and the pre-processing unit (43) is configured to convert said image signals into respective standard image signals of a predetermined frame rate and / or a predetermined resolution. The system (10) of claim 8.

10. The first information signal (11) and the second information signal (12) are both image signals, and the pre-processing unit (43) is configured to select only information from a predetermined selection of the image signals for signal processing by the neural network, based on the function of the object for which the respective image signal is used. A system (10) according to claim 8 or 9.

11. The pre-processing unit (43) is configured to select, depending on the information signal (11, 12) present at the signal input (13), whether to use the first unit (15a) or the second unit (15b) of the two division units (15a, 15b) of the neural network for the second signal processing. A system (10) according to any one of claims 7 to 10.

12. A system (20) for processing an information signal (21), comprising: a signal input unit (23) for receiving an information signal (21); a common unit (24) of the neural network of the system (20), the common unit (24) of the neural network being configured for a first signal processing step of the information signal (21); and at least two neural network splitting units (25a, 25b) arranged in parallel downstream of the common unit (24) in the signal flow, A system (20) configured to use a first unit (25a) of at least two division units (25a, 25b) for a second signal processing of an information signal (21) in a first operating mode and to use a second unit (25b) of the at least two division units (25a, 25b) for a second signal processing of the information signal (21), Furthermore, a signal output unit (26) for outputting the processed information signal (21') is provided. and A system (20) in which either the first unit (25a) or the second unit (25b) of the division units (25a, 25b) is selected depending on the type of the information signal (21) processed by the common unit (24) and / or the function assigned to the information signal (21).

13. 13. A system (20) according to claim 12, comprising: A system (20) using both the first dividing unit (25a) and the second dividing unit (25b) in parallel for second signal processing to implement the first and second operating modes in parallel.

14. A system (10, 20) according to any one of claims 1 to 13, and at least two sensors (42a, 42b) connected to the signal inputs (13, 23) of the system (10, 20), the at least two sensors (42a, 42b) being configured to send different information signals to the signal inputs (13, 23). Sensor system.

15. The at least two sensors (42a, 42b) include at least one of an optical sensor, a camera (42a, 42b), a current sensor, a temperature sensor, a state-of-charge sensor, or a sensor device configured to predict a regeneration event. The sensor system of claim 14.

16. A system (10, 20) according to any one of claims 1 to 13, arranged in a control device (40a) of a motor vehicle (40); and at least two sensors (42a, 42b) connected to the signal inputs (13, 23) of the system (10, 20) via an in-vehicle network (41) of the vehicle (40). Automobiles (40).

17. A method (30) for signal processing at least one information signal (11, 12), respectively, comprising: In a first signal processing step, the information signal (11, 12) is processed (31) by an upstream unit (14) of the neural network; Identifying (32) the type of information signal (11, 12) to be processed in the first signal processing step and / or the function for which said information signal (11, 12) is used; In a second signal processing step, further processing (33) of the information signals (11, 12) by a first unit (15a) downstream of the neural network or by a second unit (15b) downstream of the neural network; The selection of the first downstream unit (15a) or the second downstream unit (15b) of the neural network is then made depending on the type of the information signal (11, 12) and / or the type of function used to process the signal. A method (30) comprising:

Images