Method and device for predicting the remaining time of a signal phase

Description

[0001] Increasing traffic volumes pose considerable challenges, especially for large cities. Accordingly, many places are striving to reduce fuel consumption, pollutant emissions, noise emissions, and waiting times through various traffic control measures.

[0002] Publication EP3438946A2 discloses a method in which the switching times of a traffic signal or the remaining time until the signal is switched are predicted using artificial intelligence. Using this information, road users approaching a traffic signal can be prompted to adjust their speed, initiate braking early, or avoid unnecessary acceleration. In this way, fuel consumption, pollutant emissions, noise emissions, and waiting times can often be effectively reduced. However, to achieve sufficient forecasting accuracy, the described artificial intelligence typically requires highly computationally intensive training and a large amount of training data.

[0003] EP 3 333 823 A1 discloses a system for predicting a future signal pattern of a traffic signal system, comprising: a first communication interface for receiving traffic data for a traffic signal system, based on which the traffic signal system can determine a future signal pattern in order to display the determined signal pattern; a second communication interface for receiving signal pattern data of the traffic signal system corresponding to the signal pattern determined by the traffic signal system; and a computing unit comprising artificial intelligence trainable on the basis of the traffic data and the signal pattern data of the traffic signal system for predicting a future signal pattern based on the traffic data. The artificial intelligence comprises one or more neural networks.

[0004] From Sirat JA ET AL: "Neural trees: a new tool for classification: Network: Computation in Neural Systems: Vol 1, No 4", July 9, 2009 (2009-07-09), XP055797581, a neural network is known that is organized as a binary tree, with each neuron of each neural tree receiving input data from an input layer. Output data from the neural tree is used in a decision tree. The neural network has a set of perceptrons functionally organized into a binary tree.

[0005] It is an object of the present invention to provide a method for traffic control, a traffic control device and a method for training the same, which allow a better remaining time forecast or require less training effort.

[0006] This object is achieved by a method having the features of patent claim 1, by a method having the features of patent claim 5, by a traffic control device having the features of patent claim 7, by a computer program product having the features of patent claim 8 and by a computer-readable, preferably non-volatile storage medium having the features of patent claim 9.

[0007] According to the invention, traffic data from the surroundings of a traffic signal head and a signal phase indication distinguishing different signal phases of the traffic signal head are recorded for traffic control purposes. The signal phase indications for a traffic signal head indicate the signal phase the traffic signal head is currently in. Two complementary signal phases are distinguished for the traffic signal head, with a first signal phase "green" for clear traffic and a complementary second signal phase "non-green" for blocked traffic. The traffic data are fed as input data to an artificial neural network, which comprises a first subnetwork and a different second subnetwork, as well as a combination network for combining output data from both subnetworks.The artificial neural network is trained to use traffic data to reproduce the remaining time until a phase change of the traffic signal transmitter, with the first subnetwork being specifically optimized for remaining time forecasts for the green phase and the second subnetwork being specifically optimized for remaining time forecasts for the non-green phase. According to the invention, the output data of the first subnetwork and the output data of the second subnetwork are controlled in a complementary manner depending on the signal phase information. Furthermore, output data of the combined network or forecast data derived therefrom are transmitted to a means of transport or a road user as a remaining time forecast for traffic control.

[0008] To train a traffic control system with an artificial neural network, traffic data from the environment of a traffic signal head and a signal phase indication distinguishing the various signal phases of the traffic signal head are acquired. The signal phase indications for a traffic signal head indicate the current signal phase of the traffic signal head. Two complementary signal phases are distinguished for the traffic signal head, with a first signal phase, "green," indicating clear traffic and a complementary second signal phase, "off-green," indicating no traffic. The traffic data are fed as input data to the artificial neural network, which comprises a first subnetwork and a different second subnetwork, as well as a combination network for combining output data from both subnetworks.The artificial neural network is trained to use the traffic data to reproduce the remaining time until a phase change of the traffic signal transmitter, with the first subnetwork being specifically optimized for remaining time predictions for the green phase and the second subnetwork being specifically optimized for remaining time predictions for the non-green phase. According to the invention, the output data of the first subnetwork and the output data of the second subnetwork are controlled in a complementary manner depending on the signal phase information.

[0009] To carry out a method according to the invention, a traffic control device, a computer program product and a computer-readable, in particular non-volatile, storage medium are provided.

[0010] The methods according to the invention and the traffic control device according to the invention can be carried out or implemented, for example, by means of one or more processors, computers, application-specific integrated circuits (ASICs), digital signal processors (DSPs) and / or so-called "field programmable gate arrays" (FPGAs).

[0011] By using different subnetworks controlled by specific signal phases, the quality of a remaining time forecast can generally be improved and / or the required training effort reduced. An improved remaining time forecast can, in turn, reduce noise emissions, waiting times, pollutant emissions, and / or traffic delays. Furthermore, in many cases, only minor or no modifications to existing traffic systems are required to use the invention.

[0012] Advantageous embodiments and further developments of the invention are specified in the dependent claims.

[0013] According to the invention, the output of the output data of the two subnetworks is controlled depending on the signal phase specification such that, in a first signal phase, the first subnetwork outputs a forecast value and the second subnetwork outputs a neutral value to the combination network. In a second signal phase, the second subnetwork outputs a forecast value and the first subnetwork outputs a neutral value to the combination network. A numerical or logical zero is output as the neutral value, which, when combined with a quantified forecast value, changes it only slightly or not at all.

[0014] According to the invention, the combination network combines output data from the first subnetwork with output data from the second subnetwork by means of numerical addition. When a neutral value is added to a quantified forecast value, the latter changes little or not at all. In this way, a signal-phase-specific forecast value of the first subnetwork in a first signal phase and a signal-phase-specific forecast value of the second subnetwork in a second signal phase can each be output or selected as the result of the addition.

[0015] According to an advantageous embodiment of the invention, a signal-specific signal phase information can be recorded for each of several traffic signal heads. Depending on the recorded signal phase information, the first subnetwork can output signal-specific forecast values ​​for traffic signal heads in a first signal phase, and the second subnetwork can output signal-specific forecast values ​​for traffic signal heads in a second signal phase to the combined network. In this way, signal-specific forecast values ​​for several or all participating traffic signal heads can be determined and output in parallel.

[0016] Furthermore, the signal phase information and / or other data influencing the switching behavior of the traffic signal head can be fed to the artificial neural network as input data, wherein the artificial neural network is trained to additionally reproduce remaining times until a respective phase change based on signal phase information and / or other data influencing the switching behavior of the traffic signal head. Such data influencing the switching behavior of the traffic signal head can include, in particular, time information, date information, environmental data, e.g., regarding weather conditions, lighting conditions, slipperiness, or pollution levels, information about peak times, and / or event data about events, the approach of prioritized vehicles, accidents, or other traffic events.

[0017] According to a further advantageous embodiment of the invention, depending on the remaining time forecast transmitted to the vehicle or road user, an automatic start-stop system, a brake, a recuperation device, an autonomous vehicle, a navigation device, and / or a route planner can be controlled, or a notification can be issued to the road user. In this way, the vehicle or road user can be prompted to adjust speed, initiate braking early, avoid unnecessary acceleration, or otherwise react to the expected phase change.

[0018] Furthermore, the artificial neural network can be trained to reproduce a mean, a median, a quantile, a probable value, a minimum, a maximum, and / or a statistical fluctuation range of the remaining time or a probability of a phase change in a given time interval as a remaining time forecast. For example, the neural network can be trained to output a probability of a change to green in the next second as a remaining time forecast.

[0019] An embodiment of the invention is explained in more detail below with reference to the drawings, each of which shows a schematic representation: Fig. 1 a road intersection with a traffic signal system, Fig. 2 a neural network according to the invention and Fig. 3 a traffic control device according to the invention.

[0020] Fig. 1 shows a schematic representation of an example of a road intersection KR with a traffic signal system that includes several traffic lights S1 ,..., S4 as traffic signal heads. Traffic lights controlled in phase each form a signal group. In the present embodiment, S1 and S3 form a first signal group, and S2 and S4 form a second signal group. The traffic signal heads S1 ,..., S4 are each used to control or regulate local traffic depending on current traffic data or other influencing factors.

[0021] In the present embodiment, at a respective traffic signal generator S1,.. . or S4, two complementary signal phases are distinguished. This means that if a first signal phase is not currently present, then the second is, and vice versa. Thus, a first signal phase "green" can be used to indicate clear passage, and a complementary second signal phase "off-green" can be used to indicate no passage.

[0022] The traffic signal heads S1 ,..., S4 and its signal phases and phase changes are each controlled by a connected traffic light controller (CTL). The latter can be implemented locally or as part of a traffic control system spanning multiple signal systems. The control of the traffic signal heads S1,..., S4 by the traffic light control CTL is carried out depending on locally recorded traffic data or other influencing factors.

[0023] In the present embodiment, two vehicles F1 and F2 approaching the intersection KR are considered. According to standard traffic regulations, the traffic lights at the intersection KR determine which vehicle has free passage and which vehicle must wait. In the present embodiment, traffic light S1 is relevant for vehicle F1, and traffic light S2 is relevant for vehicle F2.

[0024] Furthermore, a traffic control device (RSU) according to the invention is arranged in the vicinity of the intersection KR. The latter serves to influence vehicles, in particular vehicles F1 and F2, or other road users. In this case, an internal control system of the vehicles or a smartphone can be influenced by cyclists or pedestrians.

[0025] According to the invention, the traffic control unit RSU is intended to determine a remaining time forecast, i.e., predict the time remaining until a subsequent phase change of a respective traffic signal. Alternatively or additionally, the remaining time forecast can also determine a probability of a phase change to green or non-green in the next second or in another predetermined time interval. Based on the determined remaining time forecasts, vehicles, in this case F1 and F2, or other road users can be prompted to adjust their speed, initiate braking early, or avoid unnecessary acceleration when approaching the traffic signal.For this purpose, depending on the remaining time forecast, an automatic start / stop system, a brake, a recuperation system, a navigation device, or an autonomous vehicle control system in a vehicle can be controlled or a message can be issued to a road user.

[0026] In the present exemplary embodiment, the traffic control unit RSU determines signal head-specific or signal group-specific remaining time forecasts RP1 and RP2 and transmits them to vehicles F1 and F2 for influencing them. RP1 quantifies a predicted remaining time until the next phase change of the first signal group S1, S3, and RP2 quantifies a predicted remaining time until the next phase change of the second signal group S2, S4.

[0027] The RSU traffic control unit can preferably be implemented as a so-called roadside unit. Alternatively or additionally, the RSU traffic control unit can be fully or partially integrated into or coupled to a so-called SPaT box (SPaT: Signal Phase and Timing). The remaining time forecasts RP1 and RP2 of the RSU traffic control unit can be transmitted to an onboard unit or another control device in a vehicle or to a road user's smartphone. Remaining time forecasts can also be transmitted via a smartphone, in particular to pedestrians or cyclists.

[0028] The remaining time forecasts RP1 and RP2 are calculated based on traffic data VD currently recorded in the vicinity of the traffic signal heads S1,..., S4, signal phase information SPA and, if applicable, other switching behavior of the traffic signal heads S1 ,...,S4 influencing data is determined. Traffic data VD can include, in particular, the number and speed of vehicles, their waiting times, approaches of prioritized vehicles, traffic events such as accidents or other traffic disruptions, or other information about current traffic load. To record or measure traffic data, the traffic control unit RSU has a sensor system S, which can include, for example, vehicle sensors, cameras, speed sensors, or other sensors. Alternatively or additionally, the traffic control unit RSU can have a receiving device for data recorded elsewhere that influences switching behavior.

[0029] The signal phase information SPA indicates the current signal phase for a particular traffic signal head or signal group. Alternatively or additionally, the time of the last phase change can be quantified for each signal head or signal group using a respective signal phase information. The current signal phase information SPA is stored in the traffic light control unit (CTL) and is transmitted continuously or upon phase changes from the traffic light control unit (CTL) to the traffic control unit (RSU).

[0030] The other data influencing the switching behavior of the traffic signal system can include, in particular, time information, date information, environmental data, e.g. on weather conditions, lighting conditions, slipperiness or pollution levels, information on rush hours and / or event data on events, on the approach of priority vehicles, on accidents or on other traffic events.

[0031] To determine the remaining time forecasts RP1 and RP2, the traffic control system RSU has an artificial neural network NN, which is coupled to the sensor system S. The neural network NN is trained to predict the switching behavior of the traffic signal heads S1 based on the traffic data VD, the signal phase information SPA and, if applicable, the other ,..., S4 influencing data remaining time until a respective phase change of the traffic signal head S1 ,..., S4 to predict.

[0032] The traffic control unit (RSU) also has a transmitter (TX) coupled to the neural network (NN). The transmitter (TX) transmits the remaining time forecasts (RP1 and RP2), preferably recalculated every second, to the vehicles (F1 and F2), preferably via radio.

[0033] In addition to influencing or controlling road traffic, the invention can also be used advantageously to influence or control rail traffic, robots, shipping or air traffic in order to determine, in an analogous manner, remaining time forecasts for traffic signal transmitters used there, e.g. radio-based ones, and to transmit them to the affected means of transport or road users.

[0034] Fig. 2 illustrates a neural network NN according to the invention. Insofar as Fig. 2 the same or corresponding reference numerals as in Fig. 1 are used, these reference symbols denote the same or corresponding entities, which may be implemented or configured in particular as described above.

[0035] The neural network is formed by several neural layers A to M. Data propagation between the neural layers A to M is carried out in Fig. 2 illustrated by arrows.

[0036] The neural network NN comprises, in particular, two neural subnetworks TNA and TNB, with the first neural subnetwork TNA being formed by layers A, E, G, I, and K, and the second neural subnetwork TNB being formed by layers C, F, H, J, and L. In the present embodiment, the two subnetworks TNA and TNB are separated from each other and have separate, independent data propagation paths. In particular, no neurons are shared. Alternatively, the subnetworks TNA and TNB can also be only partially separated, so that the two subnetworks TNA and TNB share individual neurons.

[0037] The neural network NN also has an input network EN. This serves to receive input data from the neural network NN, to preprocess it, and to forward the preprocessed input data to both subnetworks TNA and TNB. In the present embodiment, the input network EN is formed by the input layer B and the hidden layer D. Traffic data VD, signal phase information SPA, and other data (not shown) influencing the switching behavior of the traffic signal transmitters are fed to the input layer B as input data. The supplied input data can be Fig. 1 described. The preprocessed input data, i.e., the output data of the input network EN, are fed from layer D to both layer E of the first subnetwork TNA and layer F of the second subnetwork TNB. Alternatively, such an input network can be dispensed with. In this case, the input data can be fed directly to both subnetworks TNA and TNB.

[0038] The neural network NN further comprises a combination network, which in the present embodiment is formed by the single layer M, which simultaneously functions as the output layer of the neural network NN. The combination network M serves to combine output data from the subnetworks TNA and TNB. In the present embodiment, one output neuron of layer M is provided for each respective signal generator or signal group. Each of these signal generator-specific or signal group-specific output neurons outputs a quantified remaining time value, here RP1 or RP2, for each respective traffic signal generator or signal group.

[0039] Layers A to M can each contain one or more neurons. A connection between any two layers is Fig. 2 indicated by arrows. With such a connection between two layers, every neuron in an output layer can potentially be connected to every neuron in a target layer.

[0040] In the present embodiment, the subnetwork TNA with the hidden layers E, G, I, and K is used for remaining time forecasts of signal groups currently in the "green" signal phase, while the subnetwork TNB with the hidden layers F, H, J, and L is used for remaining time forecasts of signal groups currently in the "non-green" phase. As long as both subnetworks TNA and TNB are supplied with the same data from the input network EN, both layer I and layer J generate remaining time forecasts for all signal groups. However, for layer I, only those remaining time forecasts assigned to the signal groups currently in the "green" signal phase influence the final result. Similarly, for layer J, only those remaining time forecasts assigned to the signal groups currently in the "non-green" signal phase influence the final result.

[0041] This different, or more precisely complementary, consideration of the various signal phases by the subnetworks TNA and TNB is performed or enforced by layers K and L in conjunction with layers A and C. Thus, layer K adopts the remaining time forecasts of layer I only for those signal groups that are currently in the "green" signal phase, while a neutral value, e.g., a numerical zero, is entered for those signal groups that are currently in the "non-green" signal phase. The information required for this differentiation is provided by layer A.Depending on the supplied signal phase information SPA, the latter outputs a numeric or logical one to layer K for all signal groups currently in the "green" signal phase and a logical or numeric zero for those signal groups currently in the "non-green" signal phase. Similarly, layer L adopts the remaining time forecasts of layer J only for those signal groups currently in the "non-green" phase, while a neutral value is entered for those signal groups currently in the "green" phase. The information required for this differentiation is provided analogously by layer C depending on the supplied signal phase information SPA.

[0042] The combination network M summarizes the forecast results of layers K and L, preferably by simple addition, so that the combination network M contains the remaining time forecasts of all signal groups, regardless of whether a respective signal group is currently in the "green" or "non-green" signal phase. In the present embodiment, the combination network M outputs a remaining time forecast RP1 for the signal group consisting of traffic lights S1 and S3 and a remaining time forecast RP2 for the signal group consisting of traffic lights S2 and S4 as output data of the neural network NN.

[0043] Insofar as during training of the entire neural network NN only predictions of the subnetwork TNA for "green" and the subnetwork TNB for "non-green" are taken into account due to the complementary control of the layers K and L, the normal learning process of the neural network NN has the effect that the subnetwork TNA is specifically optimized for remaining time predictions for the green phase and the subnetwork TNB is specifically optimized for remaining time predictions for the non-green phase.

[0044] Fig. 3 shows a learning-based traffic control system RSU according to the invention in a schematic representation. Fig. 3 the same or corresponding reference symbols are used as in the preceding figures, these reference symbols denote the same or corresponding entities which can be implemented or designed in particular as described above.

[0045] The traffic control device RSU has a processor PROC for executing the method according to the invention and a memory MEM coupled to the processor PROC for storing data generated during execution.

[0046] Furthermore, the traffic control unit RSU has a sensor system S. In the present embodiment, the sensor system S serves to measure or otherwise record traffic data VD or other data that influence the switching behavior of a traffic signal system. For reasons of clarity, the latter are referred to below as traffic data and are assigned to the traffic data VD.

[0047] Furthermore, the traffic control unit RSU has an artificial neural network NN coupled to the sensor system S. The neural network NN receives the traffic data VD from the sensor system S and signal phase information SPA as input data. In the present exemplary embodiment, the neural network NN is to be trained to determine signal-group-specific and signal-phase-specific remaining time forecasts, here RP1 and RP2, for all signal groups based on the supplied traffic data VD and signal phase information SPA and to output them as output data.

[0048] Furthermore, the traffic control unit RSU comprises a transmitter TX coupled to the neural network NN for the preferably radio-based transmission of determined remaining time forecasts, here RP1 and RP2, to means of transport, here vehicles F1 and F2 and / or to other road users.

[0049] As explained above, the neural network NN comprises an input network EN, neural subnetworks TNA and TNB, and a combination network M. The traffic data VD and the signal phase information SPA are fed to the input network EN as input data. The input network EN preprocesses the input data VD and SPA and transmits the preprocessed input data to both neural subnetworks TNA and TNB.

[0050] By training the neural network NN as a whole, the sub-networks TNA and TNB are to be implicitly trained to determine signal-phase-specific forecast values, each quantifying a remaining time until the next phase change. In the present exemplary embodiment, the sub-network TNA outputs forecast values ​​P1A and P2A, and the sub-network TNB outputs forecast values ​​P1B and P2B as output data. The forecast values ​​P1A, P2A, P1B, and P2B are selected and output by the sub-networks TNA and TNB, each signal-phase-specific. To select the forecast values, the sub-network TNA has a selection network SELA, and the sub-network TNB has a selection network SELB. The selection network SELA can Fig. 2 described layers K and A and the selection network SELB comprises layers L and C.

[0051] The signal phase information SPA is fed to the selection networks SELA and SELB in order to select and output the forecast values ​​P1A, P2A, P1B, and P2B in a signal-phase-specific manner depending on the signal phase. In the present exemplary embodiment, the selection network SELA outputs the green-phase-specific forecast value P1A of the subnetwork TNA, which quantifies a remaining time, for all signal groups in the green phase according to the signal phase information SPA. In contrast, for all signal groups in the non-green phase, the selection network SELA outputs a numerical zero as the non-green-phase-specific forecast value P2A. Analogously, the selection network SELB outputs the non-green-phase-specific forecast value P2B of the subnetwork TNB, which quantifies a remaining time, for all signal groups in the non-green phase according to the signal phase information SPA.Accordingly, for all signal groups in the green phase, the selection network SELB outputs a numerical zero as the green phase-specific forecast value P1B.

[0052] The output data of subnetworks TNA and TNB are thus set to zero by the selection networks SELA and SELB, depending on the signal phase information SPA, in a signal-phase-specific, complementary manner. This means that in the green phase of a respective signal group, a quantified forecast value is output by subnetwork TNA and, complementarily, a forecast value from subnetwork TNB is suppressed to a certain extent. Similarly, in the non-green phase, a quantified forecast value from subnetwork TNB is output and, complementarily, a forecast value from subnetwork TNA is suppressed to a certain extent.

[0053] The forecast values ​​P1A, P2A, P1B, and P2B are output from the subnetworks TNA and TNB to the combination network M. In each signal phase, a quantified forecast value for all signal groups is passed on, either from the subnetwork TNA or from the subnetwork TNB to the combination network M. A zero is output from the other subnetwork in each case. As already mentioned above, the combination network M is used to combine output data from the subnetworks TNA and TNB and to output the remaining time forecasts, here RP1 and RP2.

[0054] The combination network M adds the forecast values ​​P1A and P2A of the TNA subnetwork to the forecast values ​​P1B and P2B of the TNB subnetwork, specific to the signal phase and signal group. This means that for the green phase, the forecast values ​​P1A and P1B are added, and for the non-green phase, the forecast values ​​P2A and P2B are added. The respective result is then output by the combination network M as a remaining time forecast RP1=P1A+P1B or as a remaining time forecast RP2=P2A+P2B.

[0055] To the extent that adding zero to a quantified forecast value does not change it, the complementary combination of the output data from subnetworks TNA and TNB by the combination network M has the effect that, in the green phase, quantified forecast values ​​from subnetwork TNA and, in the non-green phase, quantified forecast values ​​from subnetwork TNB are output as the respective quantified remaining time forecasts RP1 and RP2, respectively. In this way, a quantified forecast value is always available for all signal groups.

[0056] As mentioned above, the neural network (NN) is initially trained in a training phase to reproduce the remaining times until the next phase change of the traffic signal heads based on the input data VD and SPA supplied as training data, i.e., to output the most accurate remaining time forecasts RP1 and RP2. In this case, "reproduction" means that a respective remaining time, predicted based on input data available up to a particular point in time, should correspond as closely as possible to a phase change that actually occurs later.

[0057] Training is generally understood as the optimization of a mapping from input data, here VD and SPA of a parameterized system model, here the neural network NN, to output data, here the remaining time forecasts RP1 and RP2. This mapping is optimized during a training phase according to predetermined, learned or to-be-learned criteria. For prediction models, the minimization of a prediction error can be used as a criterion. Through training, for example, a network structure of neurons in the neural network NN and / or weights of connections between the neurons can be adjusted or optimized so that the predetermined criteria are met as closely as possible. Training can therefore be viewed as an optimization problem for which a variety of efficient optimization methods are available.The parameters optimized in this way, in particular the optimized connection weights between neurons, can be saved as a training structure, output or transferred to another neural network in order to configure it in an optimized manner.

[0058] In the present embodiment, to train the neural network NN, its output data RP1 and RP2 are compared with the times of phase changes that actually occur later, and time deviations Δt between predicted phase changes and actually occurring phase changes are determined. The time deviations Δt can be formed by an absolute value or a square of a respective time difference. The times of the actually occurring phase changes can be determined using the signal phase information SPA. The time deviations Δt represent a prediction error of the neural network NN and are fed back to it. Based on the fed-back time deviation Δt, the neural network NN - as in Fig. 3The neural network (NN) is trained to minimize the time deviations Δt on average, indicated by a dotted arrow. This training enables the neural network (NN) to determine relatively accurate predictions of the expected remaining times until the next respective phase change.

[0059] Alternatively or additionally, the training can also be performed entirely or partially in a cloud or by another external computer. In this case, the training data, here VD and SPA, are transmitted there in order to train an inventive external neural network as described above. The training structure of the trained external neural network thus obtained can then be transferred from the external neural network to the traffic control unit (RSU) in order to load this training structure into its neural network (NN). In this way, the external neural network is, in a sense, copied from the external computer to this traffic control unit (RSU).

[0060] The signal-phase-specific selection of the forecast values ​​of the subnetworks TNA and TNB generally ensures that changes in the parameterization of the respective subnetwork TNA or TNB for the forecast of non-green phases or green phases do not affect the output of the neural network NN and are therefore not optimized during training. In particular, the training of the subnetwork TNA is not disrupted by data from the non-green phase, and the training of the subnetwork TNB is not disrupted by data from the green phase. This separation of the training paths is advantageous because the learning tasks in the different signal phases usually differ significantly from one another. For example, the red phase of many traffic signal systems is essentially fixed-time controlled and therefore varies only relatively slightly, whereas the green phase is often traffic-dependent and thus variably controlled.

[0061] The training paths in the neural network (NN) are essentially switched on and off based on the signal phase, or switched between the subnetworks (TNA and TNB). The normal learning process of the neural network (NN) then ensures the signal-phase-specific learning of the subnetworks (TNA and TNB). It turns out that the neural network (NN) described in this way often delivers significantly better remaining time forecasts or requires significantly less training effort than conventional neural networks used for remaining time forecasts.

[0062] After completing a training phase, the trained neural network NN can be used for traffic control. For this purpose, as described above, current traffic data VD is recorded by the sensor system S and fed to the trained neural network NN along with current signal phase information SPA from the traffic light controller CTL. From this, the trained neural network NN determines, as described above, remaining time forecasts RP1 and RP2, which are transmitted to the transmitter TX and from there to the vehicles F1 and F2 or to other road users. Preferably, the neural network NN can be further trained during operation based on the currently recorded traffic data VD and the signal phase information SPA.

Claims

1. Computer-implemented method for influencing traffic, wherein a) traffic data (VD) relating to an environment of a traffic signal generator (S1-S4) are captured, b) a signal phase specification (SPA) distinguishing various signal phases of the traffic signal generator (S1-S4) is captured, wherein the signal phase specifications (SPA) indicate, for a traffic signal generator (S1-S4), the signal phase the traffic signal generator (S1-S4) is currently in, and wherein two signal phases that are complementary to one another are distinguished in the traffic signal generator (S1-S4), and a first signal phase "green" is provided for unimpeded travel and a second signal phase "non-green" that is complementary thereto is provided for blocked travel, c) the traffic data (VD) are supplied, as input data, to an artificial neural network (NN) comprising a first subnetwork (TNA) and a second subnetwork (TNB) that differs from the latter as well as a combination network (M) for combining output data (PIA, P2A, P1B, P2B) from both subnetworks (TNA, TNB), wherein the combination network (M) combines the output data (PIA, P2A) from the first subnetwork (TNA) with the output data (P1B, P2B) from the second subnetwork (TNB) by way of an addition, wherein the artificial neural network (NN) is trained to reproduce a time still remaining to a phase change of the traffic signal generator (S1-S4) on the basis of traffic data (VD), wherein the first subnetwork (TNA) is specifically optimized for predictions of the remaining time for the green phase and the second subnetwork (TNB) is specifically optimized for predictions of the remaining time for the non-green phase, d) output of the output data (PIA, P2A) from the first subnetwork (TNA) and output of the output data (P1B, P2B) from the second subnetwork (TNB) are controlled in a manner complementary to one another on the basis of the signal phase specification (SPA) in such a manner that - in the first signal phase, the first subnetwork (TNA) outputs a prediction value (PIA) and the second subnetwork (TNB) outputs a neutral value (P1B) to the combination network (M), and - in the second signal phase, the second subnetwork (TNB) outputs a prediction value (P2B) and the first subnetwork (TNA) outputs a neutral value (P2A) to the combination network (M), - wherein a numeric or logical zero is output as the neutral value (P1B, P2A), and e) output data from the combination network (M) or prediction data derived therefrom are transmitted to a means of transport (F1, F2) or to a road user as a prediction of the remaining time (RP1, RP2) for influencing traffic.

2. Method according to Claim 1, characterized in that a signal-generator-specific signal phase specification (SPA) is respectively captured for a plurality of traffic signal generators (S1-S4), and in that, on the basis of the captured signal phase specifications (SPA), - the first subnetwork (TNA) outputs signal-generator-specific prediction values (PIA) for traffic signal generators in a first signal phase and - the second subnetwork (TNB) outputs signal-generator-specific prediction values (P2B) for traffic signal generators in a second signal phase to the combination network (M).

3. Method according to Claim 1 or 2, characterized in that the signal phase specification (SPA) and / or other data influencing a switching behaviour of the traffic signal generator is / are supplied to the artificial neural network (NN) as input data, wherein the artificial neural network (NN) is trained to additionally reproduce times remaining to a respective phase change on the basis of signal phase specifications (SPA) and / or other data influencing a switching behaviour of the traffic signal generator.

4. Method according to one of the preceding claims, characterized in that, on the basis of the prediction of the remaining time (RP1, RP2) transmitted to the means of transport (F1, F2) or to the road user, an automatic start / stop system, a brake, a recuperation device, an autonomous vehicle, a navigation device and / or a route planner is / are controlled or a notification is output to the road user.

5. Method according to one of the preceding claims, characterized in that the input data are supplied to an input network (EN) of the artificial neural network (NN) and output data from the input network (EN) are supplied to both subnetworks (TNA, TNB).

6. Computer-implemented method for training a traffic influencing device (RSU) using an artificial neural network (NN), wherein a) traffic data (VD) relating to an environment of a traffic signal generator (S1-S4) are captured, b) a signal phase specification (SPA) distinguishing various signal phases of the traffic signal generator (S1-S4) is captured, wherein the signal phase specifications (SPA) indicate, for a traffic signal generator (S1-S4), the signal phase the traffic signal generator (S1-S4) is currently in, and wherein two signal phases that are complementary to one another are distinguished in the traffic signal generator (S1-S4), and a first signal phase "green" is provided for unimpeded travel and a second signal phase "non-green" that is complementary thereto is provided for blocked travel, c) the traffic data (VD) are supplied, as input data, to the artificial neural network (NN) comprising a first subnetwork (TNA) and a second subnetwork (TNB) that differs from the latter as well as a combination network (M) for combining output data (PIA, P2A, P1B, P2B) from both subnetworks (TNA, TNB), wherein the combination network (M) combines the output data (PIA, P2A) from the first subnetwork (TNA) with the output data (P1B, P2B) from the second subnetwork (TNB) by way of an addition, and d) the artificial neural network (NN) is trained to reproduce a time still remaining to a phase change of the traffic signal generator (S1-S4) on the basis of the traffic data (VD), wherein the first subnetwork (TNA) is specifically optimized for predictions of the remaining time for the green phase and the second subnetwork (TNB) is specifically optimized for predictions of the remaining time for the non-green phase, wherein output of the output data (PIA, P2A) from the first subnetwork (TNA) and output of the output data (P1B, P2B) from the second subnetwork (TNB) are controlled in a manner complementary to one another on the basis of the signal phase specification (SPA) in such a manner that - in the first signal phase, the first subnetwork (TNA) outputs a prediction value (PIA) and the second subnetwork (TNB) outputs a neutral value (P1B) to the combination network (M), and - in the second signal phase, the second subnetwork (TNB) outputs a prediction value (P2B) and the first subnetwork (TNA) outputs a neutral value (P2A) to the combination network (M), - wherein a numeric or logical zero is output as the neutral value (P1B, P2A).

7. Method according to Claim 6, characterized in that the artificial neural network (NN) is trained to reproduce - a mean value, a median, a quantile, a probable value, a minimum, a maximum and / or a statistical fluctuation range of the remaining time or - a probability of a phase change in a predefined interval of time as a prediction of the remaining time (RP1, RP2).

8. Traffic influencing device (RSU) configured to carry out a method according to one of the preceding claims.

9. Computer program product configured to carry out a method according to one of Claims 1 to 7.

10. Computer-readable storage medium having a computer program product according to Claim 9.