EP4115344A1 - Computerimplementiertes verfahren zur berechnung von faltungsnetzwerken, computersystem, computerprogramm und automatisiert betreibbares fahrzeug - Google Patents
Computerimplementiertes verfahren zur berechnung von faltungsnetzwerken, computersystem, computerprogramm und automatisiert betreibbares fahrzeugInfo
- Publication number
- EP4115344A1 EP4115344A1 EP21709929.0A EP21709929A EP4115344A1 EP 4115344 A1 EP4115344 A1 EP 4115344A1 EP 21709929 A EP21709929 A EP 21709929A EP 4115344 A1 EP4115344 A1 EP 4115344A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- image
- convolution
- filter
- size
- convolution network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
Definitions
- the invention relates to a computer-implemented method for calculating convolution networks, to a computer system, to a computer program and to a vehicle that can be operated in an automated manner.
- Convolutional networks are known from the prior art, see for example “Gradient-Based Learning Applied to Document Recognition, ILA. Convolutional Networks, "Y. LeCun et al., Proc. of the IEEE, November 1998.
- Computer systems are also known from the prior art, for example systems-on-a-chip or microcontrollers.
- a microcontroller has a relatively limited computing and storage capacity. This makes the use of artificial neural networks difficult because many arithmetic operations are required to calculate the output of the artificial neural network. This depends on the architecture of the artificial neural network. In general, the size of the network has a positive effect on the ability to classify and / or localize.
- the object of the invention was to find out how an artificial neural network that is as large as possible, which requires a large number of arithmetic operations, can be run on a microcontroller.
- the invention solves this problem by saving computing steps. This means that much larger artificial neural networks can be executed on computer systems with limited computing and storage capacity and, at the same time, better classification and / or localization results can be obtained. According to the invention, intermediate results are noted in the correct places in hidden layers of the artificial neural network during the execution of an artificial neural network, that is to say cached to these results Reuse the past and thus determine a recognition result with fewer calculation steps.
- the invention provides a computer-implemented method for computing convolution networks.
- the process comprises the steps:
- a first dimension of the first convolutional image is reduced by the update size and the reduced first convolutional image is stored in a first memory area of a computer system
- the first filter is applied in the first dimension of the second image, restricted to the sum of the update size and a corresponding dimension of the first filter reduced by a font size of the first filter, and values of the reduced first convolution image from the first Storage area can be obtained.
- the invention provides a computer system.
- the computer system comprises at least one processor unit and a memory area.
- the processor unit and the memory area are adapted so that the computer system can execute a convolution network and the steps of the method according to the invention.
- the invention provides a computer program.
- the computer program comprises commands which, when the computer program is executed by a computer system according to the invention, cause the Steps of the method according to the invention and thereby lead trainees a convolution network.
- the invention provides a vehicle that can be operated in an automated manner.
- the vehicle includes
- At least four acoustic sensors the acoustic sensors arranged in a front and rear respective right and left area of the vehicle and
- a computer system designed as a microcontroller that receives signals from the acoustic sensors
- a multi-core processor of the computer system comprises at least four cores and each core carries out a convolution network and the steps of the method according to the invention based on signals from one of the acoustic sensors and the results are merged.
- Computer-implemented means that the steps of the method are carried out by a data processing device, for example a computer system, or parts thereof. According to one aspect of the invention, the method is carried out on a microcontroller.
- the commands of the computer program include software code sections, machine code or binary code.
- the invention provides a computer-readable data carrier on which the computer program according to the invention is stored.
- the data carrier is, for example, an SSD, volatile, RAM, non-volatile, ROM, or flash disk.
- the computer system is a one-chip system, also called a system-on-a-chip.
- a one-chip system for example a One-chip computer system, all or at least a large part of the functions of a programmable electronic system, for example a computer, are integrated on a chip, that is to say a die.
- the convolution network comprises the following structure:
- an image is entered into an input layer.
- the image is stored, for example, in a memory unit of the microcontroller.
- This storage unit is also called an image buffer.
- the image buffer stores and provides the inputs for the first layer of the convolution network.
- the microcontroller comprises two image buffers, an active one, from which is currently being read, and a passive one for intermediate storage.
- the image is folded with the filter kernel and its dimensions are reduced.
- filter kernels are used to filter different features or to weight different areas of the input image differently.
- 10, 20, 30, 40, 50, 60, 70, 80, 90 or 100 filter kernels are used.
- 64 filter kernels are used.
- the filter kernels include, for example, Sobel or Scharr operators in order to classify so-called low-level features in the input image, for example edges.
- 64 folding images are obtained from the input image.
- the parameters of the filter kernel are stored, for example, in a second memory unit of the microcontroller and are read out from this.
- neuron connections of the convolution network to each image point, that is, pixels, of the convolution images obtained are activated with activation functions.
- the activation function is, for example, a so-called rectifier function known to the person skilled in the art.
- the output of a rectifier function is the maximum of input and zero.
- the activation functions for each pixel are rectifier functions, that is to say the convolution network comprises so-called rectifier linear units.
- different activation functions are activated individually or in groups on different pixels applied. The activation results are saved and are input for further calculations in the following layers of the convolution network.
- the activation results i.e. images again, are read out and the dimensions of the output of the input layer are reduced in each case using maxpooling.
- the results after maxpooling are temporarily stored, for example, in a memory area of a fourth memory unit of the microcontroller.
- the results in the most up-to-date area which corresponds to the corresponding dimension of the filter kernel reduced by the step size of the filter kernel, cached.
- the outputs of the first layer are read from the buffer and convoluted with the filter kernel. Since the dimensions are reduced again. Subsequently, neuron connections to each image point, i.e. pixels, of the convolution images now obtained are activated by rectifier linear units and the activation results are temporarily stored.
- Maxpooling and intermediate storage, repeated folding and intermediate storage are carried out alternately in the subsequent third to sixth layers, where the dimensions of the images obtained in the layers decrease from layer to layer, until finally the sixth layer images of the dimension 1x1, i.e. 1x1 pixels in x- and y-direction, outputs.
- the seventh layer there is an intermediate storage in order to record associations between all pixels of all images in a completely connected layer. This is also called dense.
- the outputs of the seventh layer are linearized using so-called flattening. Flattening is carried out, for example, with the softmax function known to the person skilled in the art. Each node of the convolution network of this layer is connected to each pixel of the input images and activated with the Softmax function.
- the eighth layer is a so-called dense layer.
- the output of the eighth layer is, for example, 64 numerical values.
- the layer immediately before the output layer of the convolution network comprises several dense layers, for example eight.
- a subsequent ninth layer is again a dense layer.
- an output layer which, for example, supplies a prediction about a classified and / or localized siren signal.
- the first to eighth layers are hidden layers between the input and output and are called hidden layers.
- the convolution network comprises recurrent layers, for example long-short-term memories, in order to remember intermediate results.
- the series of measurements includes initial data and is expanded with further data in each update.
- the convolution network is output for each update.
- the interval from which the data is obtained in order to process the first and second images has a fixed width. This means that the second, updated interval still partially includes the first data, depending on the size of the update.
- the series of measurements includes audio data.
- the convolution network is trained to classify and / or localize sirens of emergency vehicles in the audio data. In Germany, the siren sounds of emergency vehicles are specified in the DIN 14610 standard.
- the siren sound comprises two main frequencies in the range between 360 Hz and 630 Hz with a ratio of 1: 1,333, which alternate in a standardized sequence.
- the siren signal is a time-discrete signal.
- the interval is 1 s wide, which is sufficient to detect the siren tones.
- the update size is 100 ms. This means that for each update the first 90% of an image is the same as the following. Only 10% new data is added. In other words, in this example the second image matches the first image by 90%. This overlap enables intermediate results to be reused.
- Environment detection sensors are sensors with which an autonomous system, for example a robot, a robotic vehicle, or even partially automated vehicles, records its environment in order to perceive the environment and move around in it.
- an acoustic sensor is an environment detection sensor that emits sound waves the environment.
- an ultrasonic sensor is an environment detection sensor.
- Environment detection sensors also include, for example, radiation sensors that scan or scan an environment with rays.
- a radar, a lidar or a time-of-flight camera is a radiation sensor.
- the first and second images are, for example, two-dimensional pixel arrays.
- a first dimension is spanned, for example, in an x-direction of a Cartesian coordinate system.
- a second dimension is spanned, for example, in a y-direction of a Cartesian coordinate system.
- the font size of the filter kernel indicates how far the filter kernel moves from application to application.
- the font size is 1. That is, each time the filter kernel is used, a pixel column / row in the input image changes. If the first and the second image, also called the input image, each have the dimension AxB and the filter kernel the dimension UxV, then the resulting convolution image with a font size of 1 has the dimension [A- (U- 1)] x [B - (V-1)].
- A, U, and A- (U-1) indicate the extent, that is to say the number of pixels, in the y-direction, which corresponds to the lines of a pixel array.
- B, V, and B- (V-1) indicate the extent, that is, the number of pixels, in the x-direction, which corresponds to the columns of a pixel array.
- the first and second images are each obtained, for example, as follows:
- the siren signals detected with acoustic sensors are broken down into their frequency components by means of an algorithm, for example by means of fast Fourier transformation, abbreviated FFT.
- Mel-frequency-cepstrum coefficients are calculated from the frequency spectrum obtained in this way.
- the first and second images are a representation of the frequency spectrum in mel-frequency-cepstrum coefficients.
- the x-direction then indicates the time.
- the y-direction indicates the features.
- the second picture is the same size as the first picture. If the first image has a number of B-pixels in the x-direction, for example, then after updating the second image in the x-direction also has a number of B-pixels. If the update size in the x direction corresponds, for example, to a number of US pixels, then (B-US) pixels in the x direction of the second image are identical to the corresponding pixels of the first image. Only US pixels are added to the known (B-US) pixels in the positive x direction.
- a first dimension of the convolution image to reduce the update size means that the reduced convolution image in the x direction comprises a number of [B- (V-1) - US] pixels, where B- (V-1) is the number of Pixels of the convolution image in the x-direction and US correspond to the update size in the x-direction.
- the filter kernel when the second image is folded, the filter kernel only needs to be used in the x direction over the area [US + (V-1) j.
- the values for the remaining area [B- (V-1) -US] are known from the reduced convolution image and are read out from the memory area when the second image is folded without having to be recalculated again. This saves computing power.
- the first convolution image and the second convolution image are each discretized in a second layer of the convolution network with at least one second filter. The discretized first convolutional image is reduced in the first dimension and the reduced, discretized first convolutional image is stored in a second memory area of the computer system.
- the second filter in the first dimension of the second convolutional image is restricted to the corresponding dimension of the first convolutional image reduced by the update value, if the value of the corresponding dimension of the first convolutional image reduced by the updated value is divisible by is the value of the corresponding dimension of the second filter. Otherwise, the second filter in the first dimension of the second convolutional image is restricted to the corresponding dimension of the first convolutional image reduced by the update size and is applied modulo the value of the dimension of the second filter. Values of the reduced, discretized first convolutional image are obtained from the second memory area.
- the discretization reduces the dimensions of the convolution images. This corresponds to reducing the dimensions of the input images by means of convolution.
- maxpooling is used to discretize.
- the size of the maxpooling kernel is different in the individual layers of the convolution network. With the maxpooling filter, the size of the maxpooling kernel is shifted and the maximum is output in the corresponding area of the respective input image, here the convolution images.
- zero padding is used in maxpooling. This means that zeros are entered in the rows and columns of the maxpooling kernel until the dimension of the maxpooling kernel corresponds to the dimension of the input image. This advantageously ensures that each pixel of the input image, here the convolution image, lies once in the center of the maxpooling kernel in order to evaluate the pixels of the input image as homogeneously as possible.
- the first and second images are convoluted with a plurality of first filters in the first layer.
- first filters are applied. That is, 64 convolution images are obtained from each of the first and second images.
- first filters several low-level features are advantageously recorded, which leads to an improved understanding of the first and second images.
- the at least first and / or second filters are used in a plurality of layers in order to fold images that are input into the respective layers and / or to discretize respective convolution images.
- the series of measurements is updated several times and the method according to the invention is repeated for each update variable in each layer.
- the method according to the invention is thus carried out on larger convolution networks, for example in the convolution network described by way of example at the beginning, comprising eight hidden layers.
- the method is therefore applied to the subsequent convolution and maxpooling layers of the convolution network.
- the sizes of the first and / or second filters differ in the respective layers of the convolution network.
- results of the individual layers are temporarily stored.
- the computer system comprises a microcontroller.
- the series of measurements includes audio data from at least one acoustic sensor that detects traffic noise.
- the first and second images are each a Mel Frequency Cepstrum Coefficient representation of the audio data.
- a microcontroller is a one-chip computer system which comprises a processor, a memory area comprising read-only memories and read / write memories, an input / output capability and peripheral functions.
- the individual elements are connected to each other via a bus system to transmit signals.
- the periphery includes input and output devices, external memories and the corresponding interfaces.
- Peripheral functions are, for example, Controller Area Network, abbreviated CAN, Local Interconnect Network, abbreviated to LIN, Universal Serial Bus, abbreviated to USB, serial or Ethernet interfaces.
- the computer system is a microcontroller of a control device, also called an electronic control unit, which has been developed for use in the automotive sector.
- the control unit prepares input signals, processes them by means of the microcontroller and provides logic and / or power levels as rule and / or control signals, in particular for the automated operation of a vehicle.
- the subject matter of the invention can advantageously be used in Fahrzeu conditions.
- the convolution network is trained to classify and / or localize operational noises from emergency vehicles, where the interval of the measurement series from which the inputs for the convolution network are obtained comprises a period of 1 s and the update size in the range from 10 ms to 100 ms, preferably in the range from 50 ms to 100 ms, and particularly preferably 100 ms.
- a further advantage of the invention is that in the event of a siren detection it can be output as often as desired whether a siren can currently be heard. If the update size and thus the output interval are reduced to 50 ms, for example, only 5% of the data is new. Accordingly, fewer operations need to be calculated.
- the computer system comprises a multi-core processor.
- the storage area comprises at least four storage units, wherein
- a first storage unit stores classification and / or localization results of the convolution network
- a second storage unit stores parameters of the convolution network
- a third storage unit stores at least one look-up table for transformations with which a series of measurements is preprocessed
- the multi-core processor reads and / or writes to the storage units during the execution of the convolution network.
- Multi-core processors achieve higher computing power and are cheaper to implement in a chip compared to multi-processor systems in which each individual core is arranged in a processor socket and the individual processor sockets are arranged on a motherboard.
- the individual processors are, for example, central processors, so-called central processing units, abbreviated CPU.
- the central processors have a 64-bit architecture, for example.
- the first storage unit stores, for example, the predictions of the convolution network from its output layer.
- the first memory unit is, for example, a local memory unit, abbreviated LMU.
- the local memory unit comprises hardware interfaces to a CPU core, for example via so-called shared resource interconnection, abbreviated SRI, connections.
- the multi-core processor comprises four processor cores, each of which calculates the convolution network and thereby executes the method according to the invention.
- the individual predictions of the processor cores, obtained after the respective execution of the convolution network, are stored in the first memory unit and, for example, on a first of the four processor cores and, for this purpose, redundantly on a second of the four processor cores plausibility checked. This increases operational reliability in accordance with the ISO 26262 standard and the final prediction can be validated, in particular through the detection of a siren signal.
- Parameters of the convolution network include, for example, weights for connections from nodes of the convolution network.
- the second memory unit is also a local memory unit, for example.
- the first and second memory units each have a size of 512 kB, for example.
- the third memory unit is, for example, a local memory unit distributed over the individual processor cores, also called distributed local memory unit, abbreviated DLMU.
- the third memory unit has a size of 64 kB, for example. This increases the performance of read and write operations on the memory area.
- the fourth memory unit is, for example, a so-called data scratchpad ram main memory, abbreviated DSPR.
- the fourth memory unit has a size of 96 kB, for example.
- each processor core of the multi-core processor comprises its own DSPR memory unit.
- a scratchpad memory is a processor core of internal high-speed memory. This accelerates the temporary storage of the convolution images and the discretized convolution images.
- the fourth memory unit comprises a first and a second buffer memory.
- the first buffer stores inputs for layers of the convolution network.
- the second buffer stores outputs from the layers.
- the first buffer has a size of 26.25 kB and stores the entries for the first, fourth, sixth and eighth layers of the convolution network.
- the second buffer has, for example, a size of 16.25 kB and stores the outputs of the second, fourth, sixth and eighth shift. This improves the storage process, the reading and writing of the storage area in the individual layers of the convolution network.
- the microcontroller is, for example, a microcontroller from the Infineon Automotive Realtime Integrated NeXT Generation Architecture family, AURIX for short.
- Automated operable vehicle for example with internal combustion engine, electrical Shem drive, hybrid electric vehicles or fuel cell vehicles, preferably road vehicles with one of these drive technologies, include technical equipment to control the vehicle to cope with driver tasks.
- the invention for driving functions of Levels SAE J3016 Level 2 to Level 5 used.
- the multi-core processor includes, for example, four processor cores, one for each acoustic sensor.
- the results are merged as follows, for example: Using beam finder and beam tracker algorithms, each processor core detects the stem frequencies of the siren signal starting from the ceptrum by executing the convolution network and the method according to the invention. A detection occurs when a jump in frequency is detected between successive detected frequencies. For example, a first processor core evaluates the audio data of an acoustic sensor arranged on the front left of the vehicle and a second processor core correspondingly evaluates the audio data of an acoustic sensor arranged on the front right of the vehicle.
- a third processor core evaluates the audio data of an acoustic sensor arranged on the rear left of the vehicle and a fourth processor core correspondingly evaluates the audio data of an acoustic sensor arranged on the rear right of the vehicle.
- the detections of the first, second, third and fourth processor core are merged in a first fusion result on the first processor core.
- a redundancy is achieved by the first and second merger result, which is advantageous in terms of operational safety in accordance with the ISO 26262 standard.
- the first and second fusion results are checked for plausibility on the first processor core.
- a front direction angle on the first processor core determines a front direction angle on the first processor core.
- a rearward direction angle is determined on the third processor core. The results of the front and rear direction angle determination are merged on the first processor core and a direction determination is obtained from which the siren signal arrives.
- cepstrum it is determined on the basis of the siren signal whether an emergency vehicle is approaching or moving away relative to the host vehicle, for example according to the disclosure of the application DE 10 2019 213 697.4, the content of which is explicitly stated here with regard to the detection of approaching or removal of an audio signal source by means of cepstrums Reference is included.
- FIG. 3 shows a schematic representation of the buffering according to the invention for maxpooling
- FIG. 6 shows an embodiment of an output image of a first layer of the folding network
- FIG. 7 shows an exemplary embodiment of a memory area of a computer system according to the invention and 8 to 18 exemplary embodiments of a method according to the invention in respective layers of an exemplary embodiment of a convolution network.
- Fig. 1 shows, for example, an input image for a convolution network F.
- the input image comprises X Y pixels.
- a first filter 31, also convolutional, for short conv., Filter kernel of the dimension or size V ⁇ U pixels is applied to this input image.
- the convolution image has the size (X- (V-1)) x (Y- (U-1)) pixels.
- This convolution image is now discretized with a second filter 32, for example maxpooling kernel of size T x S.
- the resulting image has the size [(X- (V-1)) / T] x [(Y- (U-1) ) / S] pixels.
- any number of kernel filters is used in each layer of the convolution network F.
- U is the update rate in the x direction. After the update rate U, new data is added to the measurement series.
- the new input image for the convolution network is shown. If the first filter 31 of size V x U runs over the new input image, it only has to run over [U + (V-1)] x Y, since the results of the remaining area have already been calculated in the previous step.
- the hatched area of the right sketch is the result of the first filter 31 over the hatched area of the left sketch. All values that are contained in the non-hatched area of the right-hand sketch have already been calculated and can therefore be called up from an intermediate memory, for example from a memory area of a fourth memory unit DSPR of a microcontroller.
- FIG. 3 shows which values of a following second filter, the maxpooling kernel, have to be recalculated and which values can be accessed via the buffer.
- FIGS. 2 and 4 the principle of FIGS. 2 and 4 is repeatedly applied to subsequent layers of the convolution network, see FIGS. 8 to 18.
- FIG. 4 An example of the method according to the invention is shown in FIG. 4.
- a first procedural step V1 a first image is used as input for the convolution network F from an interval R of a series of measurements of at least one surroundings detection sensor and a second image of the same size as the first image is used as input for the convolution network F obtained from the interval R of the measurement series updated by an update variable U.
- the first and second images are input into the convolution network F, folded with at least the first filter 31 and a first and second convolution image are obtained.
- a first dimension of the first convolution image is reduced by the update size U and the reduced first convolution image is stored in a first memory area of a computer system in a third method step V3.
- the first filter 31 is applied in the first dimension of the second image, restricted to the sum of the update size U and a corresponding dimension of the first filter 31 reduced by a step size of the first filter 31 and obtain values of the reduced first convolution image from the first memory area.
- the 5 shows a Mel Frequency Cepstrum Coefficient representation of audio data of at least one acoustic sensor that detects traffic noise.
- This representation is the first and second images that are entered into the convolution network.
- the x-axis is the time t axis.
- the features f of the frequency spectrum are plotted on the y-axis.
- the image size is 32 x 100 pixels. 10 pixels correspond to 100 ms.
- the update size U is 100 ms.
- the interval R is 1 s, which means that new data is added every 100 ms and the interval shifts to the left by 100 ms in the positive time direction.
- 6 shows the result image of the first layer of the convolution network F.
- the area W of the image that is to say reuse, is reused from a previous output. Only one area C, i.e. calculation, has to be recalculated.
- the memory unit of the microcontroller includes a first memory unit LMUO which, among other things, stores classification and / or localization results of the convolution network F.
- a second memory unit LMU1 stores parameters P of the convolution network F.
- the first and second memory units are designed, for example, as local memory units and have a size of 512 kB. About 485 kB are required to save the results and the parameters.
- a third storage unit DLMU stores at least one look-up table for transformations with which a series of measurements is preprocessed, for example Fourier transformations and coefficients.
- the third memory unit is designed as a distributed local memory unit. The method requires approx.
- a fourth storage unit DSPR stores the convolutional images and discretized convolutional images.
- the fourth memory unit is designed, for example, as a data scratchpad ram main memory. The method requires approx. 78 kB of the 96 kB provided by the third storage unit, for example.
- the fourth storage unit DSPR comprises a storage unit 21 with a size of 3 kB for storing the input images of the convolution network F.
- the fourth storage unit DSPR comprises a first buffer 23 of size 26.25 kB and a second buffer 24 of size 16.25 kB.
- the first buffer store 23 stores inputs for layers of the convolution network F and the second buffer store 24 stores outputs of the layers. This includes results from previous layers as described for FIGS. 2 and 3.
- the memory unit comprises a further memory unit 22.
- the memory unit 22 is, for example, an expanded memory, abbreviated EMEM.
- An exemplary layer structure of the convolution network is shown in FIGS. 8 to 18.
- the zeroth layer LO is the input layer of the convolution network F.
- the 32 x 12 pixel image is read from the storage unit 21, entered into the convolution network F and convoluted 3 x 3 with the first filter. For example, 64 first 3 x 3 filters are applied to the 32 x 12 pixel image. This means that 64 images of the size 30 ⁇ 10 are created. Each pixel of these images is activated with the activation function A and the images obtained in this way are stored in the memory unit 22.
- the output of the zeroth layer LO are 64 images with a size of 30 x 10 pixels.
- the input is the 64 images with a size of 30 x 10 pixels from the storage unit 22. These are discretized with the second filter 32, a maxpooling kernel with a size of 2 x 2 pixels. 64 images with a size of 15 x 5 pixels are obtained.
- the results in the most up-to-date area which corresponds to the dimension of the filter kernel reduced by the step size of the filter kernel, are displayed corresponds to, cached. That is 64 pictures of the size 15 x 2 pixels.
- 64 images of the size 15 ⁇ 7 pixels are then obtained and stored in the first buffer memory 23.
- the output of the first layer L1 is 64 images with a size of 15 x 7 pixels.
- Fig. 10 shows the second layer L2.
- the input is the 64 images of the size 15 x 7 pixels from the first buffer memory 23. These are convoluted with the first filter 31, the 3 x 3 kernel filter. 64 images with a size of 13 x 5 pixels are obtained. Each pixel of these images is activated with the activation function A and the images obtained in this way are stored in the second buffer memory 24.
- the output of the second layer L2 is 64 images with a size of 13 x 5 pixels.
- 11 shows the third layer L3.
- the input is the 64 images with a size of 13 x 5 pixels from the second buffer memory 24. These are discretized with the second filter 32, a maxpooling kernel with a size of 2 x 3 pixels. 64 images with a size of 7 x 2 pixels are obtained.
- the results in the most current area are displayed of the filter kernel, cached. That is 64 pictures of the size 7 x 2 pixels. 64 images of the size 7 ⁇ 4 pixels are then obtained and stored in the first buffer memory 23.
- the output of the third layer L3 is 64 images with a size of 7 x 4 pixels.
- Fig. 12 shows the fourth layer L4.
- the input is the 64 images of the size 7 x 4 pixels from the first buffer memory 23. These are convoluted with the first filter 31, the 3 x 3 kernel filter. 64 images with a size of 5 x 2 pixels are obtained. Each pixel of these images is activated with the activation function A and the images obtained in this way are stored in the second buffer memory 24.
- the output of the fourth layer L4 is 64 images with a size of 5 x 2 pixels.
- the input is the 64 images with a size of 5 x 2 pixels from the second buffer memory 24. These are discretized with the second filter 32, a maxpooling kernel with a size of 2 x 3 pixels. 64 images with a size of 3 x 1 pixels are obtained.
- the results in the most current area are displayed Filter kernels, cached. That is 64 pictures of the size 3 x 2 pixels. 64 images of the size 3 ⁇ 3 pixels are then obtained and stored in the first buffer memory 23.
- the output of the fifth layer L5 is 64 images with a size of 3 x 3 pixels.
- 14 shows the sixth layer L6.
- the input is the 64 images of the size 3 x 3 pixels from the first buffer memory 23. These are convoluted with the first filter 31, the 3 x 3 kernel filter. 64 images with a size of 1 x 1 pixel are obtained. Each pixel of these images is activated with the activation function A and the images obtained in this way are stored in the second buffer memory 24.
- the output of the sixth layer L6 is 64 images with a size of 1 x 1 pixel.
- Fig. 15 shows the seventh layer L7.
- the input is the 64 images of the size 1 x 1 pixel from the second buffer memory 24. These are discretized with the second filter 32, a maxpooling kernel of the size 2 x 2 pixels. 64 images with a size of 1 x 1 pixel are obtained.
- the results in the most current area which corresponds to the dimension of the reduced by the step size of the filter kernel, are displayed Filter kernels, cached. That is 64 pictures of the size 1 x 1 pixel.
- 64 images with a size of 1 ⁇ 2 pixels are then obtained and stored in the first buffer memory 23.
- the output of the seventh layer L7 are 64 images with a size of 1 ⁇ 2 pixels.
- the 16 shows the eighth layer L8.
- the input is the 64 images of the size 1 x 2 pixels from the first buffer memory 23. These are flattened. 64 rows or column vectors with 128 entries each are obtained. The vectors are transformed into numerical values with a corresponding matrix multiplication M. 64 numerical values are obtained and stored in the second buffer memory 24. The output of the eighth layer L8 is 64 numerical values.
- the eighth layer is a dense layer.
- FIG. 17 shows the ninth layer L9, which, analogously to the eighth layer L8, is a dense layer.
- the input of the ninth layer L9 is 64 numerical values.
- Output O of the ninth layer L9 are also 64 numerical values.
- 18 shows the tenth layer L10.
- the tenth layer L10 is the output layer of the convolution network F.
- the output O is a prediction of a detected siren signal.
- the prediction is stored in the first memory unit LMUO.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Image Processing (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| DE102020202871.0A DE102020202871B4 (de) | 2020-03-06 | 2020-03-06 | Computerimplementiertes Verfahren zur Berechnung von Faltungsnetzwerken, Computersystem, Computerprogramm und automatisiert betreibbares Fahrzeug |
| PCT/EP2021/055152 WO2021175834A1 (de) | 2020-03-06 | 2021-03-02 | Computerimplementiertes verfahren zur berechnung von faltungsnetzwerken, computersystem, computerprogramm und automatisiert betreibbares fahrzeug |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| EP4115344A1 true EP4115344A1 (de) | 2023-01-11 |
Family
ID=74858410
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP21709929.0A Withdrawn EP4115344A1 (de) | 2020-03-06 | 2021-03-02 | Computerimplementiertes verfahren zur berechnung von faltungsnetzwerken, computersystem, computerprogramm und automatisiert betreibbares fahrzeug |
Country Status (3)
| Country | Link |
|---|---|
| EP (1) | EP4115344A1 (de) |
| DE (1) | DE102020202871B4 (de) |
| WO (1) | WO2021175834A1 (de) |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102019213697B4 (de) | 2019-09-10 | 2021-09-16 | Zf Friedrichshafen Ag | Verfahren zum Erkennen einer Annäherung und/oder Entfernung eines Einsatzfahrzeugs relativ zu einem Fahrzeug |
-
2020
- 2020-03-06 DE DE102020202871.0A patent/DE102020202871B4/de active Active
-
2021
- 2021-03-02 WO PCT/EP2021/055152 patent/WO2021175834A1/de not_active Ceased
- 2021-03-02 EP EP21709929.0A patent/EP4115344A1/de not_active Withdrawn
Also Published As
| Publication number | Publication date |
|---|---|
| DE102020202871B4 (de) | 2022-12-29 |
| DE102020202871A1 (de) | 2021-09-09 |
| WO2021175834A1 (de) | 2021-09-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| DE102020128978A1 (de) | Trainieren von tiefen neuronalen netzwerken mit synthetischen bildern | |
| WO2019201400A1 (de) | Steuergerätetesteinrichtung zum testen, absichern und entwickeln von funktionen | |
| DE102021112616A1 (de) | Selbstüberwachte schätzung einer beobachteten fahrzeugpose | |
| EP3430423B1 (de) | Verfahren zur verbesserung einer erfassung zumindest eines objekts in einer umgebung eines kraftfahrzeugs durch eine indirekte messung mit sensoren, steuereinrichtung, fahrerassistenzsystem sowie kraftfahrzeug | |
| DE102018126511A1 (de) | Lidar-signal-komprimierung | |
| DE102021122407A1 (de) | Segmentierung und klassifizierung von punktwolkendaten | |
| DE102022121109A1 (de) | Visuelle Wahrnehmung mit einem Fahrzeug basierend auf einem Kamerabild und einer Ultraschallkarte | |
| DE102020127051A1 (de) | Verfahren zur Bestimmung von sicherheitskritischen Ausgabewerten mittels einer Datenanalyseeinrichtung für eine technische Entität | |
| EP4115344A1 (de) | Computerimplementiertes verfahren zur berechnung von faltungsnetzwerken, computersystem, computerprogramm und automatisiert betreibbares fahrzeug | |
| WO2021190922A1 (de) | Vorrichtung, system und verfahren zur erkennung von objekten in einem umfeld eines automatisierten fahrsystems | |
| DE102020132042A1 (de) | Verfahren, System, Testsystem und Computerprogrammprodukt zur prädiktiven Bestimmung von Testergebnissen für ein technisches Testobjekt | |
| DE112021005568T5 (de) | Datenverarbeitungsvorrichtung und Datenverarbeitungsverfahren | |
| DE102024114791A1 (de) | Vorrichtung und verfahren zur objekterkennung | |
| DE102023106974A1 (de) | Verfahren zum erkennen von objekten mit eigenbewegung | |
| DE102022119949A1 (de) | Dynamischer fahrzeugbetrieb | |
| DE102023206446A1 (de) | Verfahren zum Verarbeiten einer Punktwolke | |
| DE102018107838A1 (de) | Verfahren zum simulativen Bestimmen von zumindest einer Messeigenschaft eines virtuellen Sensors, sowie Rechensystem | |
| DE102015201868A1 (de) | Optimierte Speicherung einer Umfeldkarte im Speicher eines Fahrzeugs | |
| DE102020201544A1 (de) | Verarbeitung von Messdaten | |
| DE102020134331A1 (de) | Verfahren zum Bestimmen eines Freiraums in einer Fahrzeugumgebung | |
| DE102020213624A1 (de) | Vorrichtung zur Vorverarbeitung von Sensordaten mehrerer Sensoren in einem Fahrzeug | |
| DE102015103632B4 (de) | Verfahren zum Überprüfen einer Auswirkung einer Veränderung von Sensordaten auf ein Fahrerassistenzsystem und programmierbare Hardwarekomponente | |
| DE102020200875A1 (de) | Verfahren zum Bereitstellen von Sensordaten durch eine Sensorik eines Fahrzeugs | |
| DE102019111608A1 (de) | Verfahren zum Bestimmen einer Eigenbewegung eines Kraftfahrzeugs, elektronische Recheneinrichtung sowie elektronisches Fahrzeugführungssystem | |
| DE102023116653A1 (de) | Verfahren zum segmentieren von objekten mit eigenbewegung |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: UNKNOWN |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
| PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
| 17P | Request for examination filed |
Effective date: 20220707 |
|
| AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
| DAV | Request for validation of the european patent (deleted) | ||
| DAX | Request for extension of the european patent (deleted) | ||
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: EXAMINATION IS IN PROGRESS |
|
| 17Q | First examination report despatched |
Effective date: 20250325 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN |
|
| 18D | Application deemed to be withdrawn |
Effective date: 20250726 |