# US 12,600,379 B2 · Computer system and method for determining reliable vehicle control instructions using traffic signal information

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,600,379 B2
- Document ID: US12600379B2
- Status: Granted
- Kind: B2
- Application number: 18/300,765
- Filed: 2023-04-14
- Published: 2023-08-10
- Granted: 2026-04-14
- Earliest claimed: 2020-10-16
- Assignees: Urban Software Institute GmbH
- Applicants: Urban Software Institute GmbH
- Inventors: Jürgen Mück; Aristotelis Hadjakos; Stefan Radomski; Michele Piccolini
- Web page: https://patentgrep.com/patent/US12600379B2

## CPC

- B60W 60/0015: {specially adapted for safety}
- G08G 1/0141: {for traffic information dissemination}
- B60W 40/04: Traffic conditions
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/08: Learning methods
- G08G 1/07: Controlling traffic signals
- G08G 1/096725: {where the received information generates an automatic action on the vehicle control}
- G08G 1/096783: {where the origin of the information is a roadside individual element}
- B60W 2556/45: External transmission of data to or from the vehicle
- B60W: CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT
- G08G: TRAFFIC CONTROL SYSTEMS (guiding railway traffic, ensuring the safety of railway traffic B61L; arrangement of road signs or traffic signals E01F9/00; radar or analogous systems, sonar systems, lidar systems specially adapted for traffic control G01S13/91, G01S15/88, G01S17/88; {radar or analogous systems, sonar systems, lidar systems specially adapted for anti-collision purposes G01S13/93, G01S15/93, G01S17/93})
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Described techniques receive signal state data from a controller controlling one or more signal groups of a traffic light system, with a current signal state of respective signal groups, and information about the time interval elapsed since the last state switch of respective signal groups. A feature vector is generated with a first value indicating the elapsed time since the latest switch of the respective signal group to a pass-state, and a second value indicating the elapsed time since the latest switch of the respective signal group to a stop-state. Based on the current feature vector, a quantile regression neural network predicts a low conditional quantile representing the minimum-end-time, a medium conditional quantile representing the likely-time and a high conditional quantile representing the maximum-end-time for respective signal groups. A message which includes the predicted minimum-end-time, likely-time, and maximum-end-time is provided to a receiving device associated with the vehicle.

## Claims

### Claim 1

A computer implemented method for providing Signal Phase and Timing (SPaT) messages for influencing operation of a vehicle, wherein the SPaT messages include timing parameters for a minimum-end-time, likely-time, and maximum-end-time for a next signal switch of at least one signal group of a traffic light system affecting the vehicle, the method comprising:
receiving signal state data from a controller controlling one or more signal groups (S1 to S4) of a traffic light system, wherein the signal state data of a particular signal group (S*) comprises a current signal state (SD*) of the respective signal group (S*), and
information about a time interval (Te*) which has elapsed since a last state switch of the respective signal group (S*);
generating a feature vector (FV) for the one more signal groups (S1 to S4), the feature vector comprising for each of the one or more signal groups at least: a first value indicating the elapsed time (T*p) since a latest switch of the respective signal group (S*) to a pass-state (1), and a second value indicating the elapsed time (T*s) since a latest switch of the respective signal group (S*) to a stop-state (0);
predicting, by at least one quantile regression neural network with a particular current feature vector (T*p, T*s) as test input, for at least one of the one or more signal groups (S1 to S4), as one or more outputs (O1), a low conditional quantile (Ql) representing the minimum-end-time (ETmin), a medium conditional quantile (Qm) representing the likely-time (Tl) and a high conditional quantile (Qh) representing the maximum-end-time (ETmax) for the at least one of the one or more signal groups (S1 to S4), wherein the at least one quantile regression neural network has been trained with stochastic gradient descent, on past feature vectors of at least a subset of the one or more signal groups of the traffic light system and respective low, medium and high conditional quantile outputs, using a loss function being a sum of individual loss functions over the at least one of the one or more signal groups (S1 to S4), and a corresponding one or more conditional quantiles, wherein the individual loss functions are hinge loss functions (HL);
composing a SPaT message including the predicted minimum-end-time (ETmin), likely-time (Tl), and maximum-end-time (ETmax) of the at least one of the one or more signal groups (S1 to S4);
providing the SPaT message to a receiving device associated with the vehicle, and
if the time interval between an actual point in time (t−1) and the current predicted likely-time (Tl-1) is below a predefined filter time interval (FTI)
filtering out a current maximum-end-time prediction (Etmax-1) which predicts a later point in time than a previous filtered maximum-end-time prediction (Etmax-2),
filtering out a current minimum-end-time prediction (Etmin-1) which predicts an earlier point in time than the previous filtered minimum-end-time prediction (Etmin-3),
else,
only filtering out the current minimum-end-time prediction (Etmin-2) which predicts an earlier point in time than the previous filtered minimum-end-time prediction (Etmin-3).

### Claim 2 (depends on 1)

The method of claim 1, further comprising:
before applying filtering,
adding a first predetermined safety margin (SM-max) to a latest filtered maximum-end-time prediction, and
subtracting a second predetermined safety margin (SM-min) from the latest filtered maximum-end-time prediction,
wherein the first and second predetermined safety margins are determined based on a comparison between historic predictions and corresponding actual switching time points and are selected to limit a percentage of violating predictions to a percentage below a given threshold value, wherein violating predictions are predictions where an actual state switch of the corresponding signal group occurred before the predicted minimum-end-time or after the predicted maximum-end-time in a given time interval before the actual switching time point.

### Claim 3 (depends on 1)

The method of claim 1, wherein the at least one quantile regression neural network is implemented using any of the following neural network architectures: a recurrent neural network, a multi-layer perceptron, and a convolutional neural network.

### Claim 4 (depends on 1)

The method of claim 1, wherein a grid search or a random search of a parameter space is used to provide a plurality of layers of the at least one quantile regression neural network and a number of hidden units per layer.

### Claim 5 (depends on 1)

The method of claim 1, wherein the at least one quantile regression neural network comprises:
a first neural network model trained for predicting state switching from a pass-state to a stop-state; and
a second neural network model trained for predicting state switching from the stop-state to the pass-state.

### Claim 6 (depends on 1)

The method of claim 1, wherein the predicted maximum-end-time (ETmax) is only provided with the SPAT message to the receiving device if a future signal switch is predicted to occur within a predefined future interval.

### Claim 7 (depends on 6)

The method of claim 6, wherein the predictions by the at least one quantile regression neural network are used to distinguish between first and second conditions based on two value regions of the prediction, with the first condition being that the signal is expected to switch within a predicted time, and the second condition being that no signal switch is expected within the predicted time.

### Claim 8

A computer program product for providing Signal Phase and Timing (SPaT) messages for influencing operation of a vehicle, wherein the SPAT messages include timing parameters for a minimum-end-time, likely-time, and maximum-end-time for a next signal switch of at least one signal group of a traffic light system affecting the vehicle, and wherein the computer product is tangibly embodied in a non-transitory memory and includes instructions, which, when executed by at least one processor of a computer cause the computer to:
receive signal state data from a controller controlling one or more signal groups (S1 to S4) of a traffic light system, wherein the signal state data of a particular signal group (S*) comprises a current signal state (SD*) of the respective signal group (S*), and
information about a time interval (Te*) which has elapsed since a last state switch of the respective signal group (S*);
generate a feature vector (FV) for the one more signal groups (S1 to S4), the feature vector comprising for each of the one or more signal groups at least: a first value indicating the elapsed time (T*p) since a latest switch of the respective signal group (S*) to a pass-state (1), and a second value indicating the elapsed time (T*s) since a latest switch of the respective signal group (S*) to a stop-state (0);
predict, by at least one quantile regression neural network with a particular current feature vector (T*p, T*s) as test input, for at least one of the one or more signal groups (S1 to S4), as one or more outputs (O1), a low conditional quantile (Ql) representing the minimum-end-time (ETmin), a medium conditional quantile (Qm) representing the likely-time (Tl) and a high conditional quantile (Qh) representing the maximum-end-time (ETmax) for the at least one of the one or more signal groups (S1 to S4), wherein the at least one quantile regression neural network has been trained with stochastic gradient descent, on past feature vectors of at least a subset of the one or more signal groups of the traffic light system and respective low, medium and high conditional quantile outputs, using a loss function being a sum of individual loss functions over the at least one of the one or more signal groups (S1 to S4), and a corresponding one or more conditional quantiles, wherein the individual loss functions are hinge loss functions (HL);
composing a SPAT message including the predicted minimum-end-time (ETmin), likely-time (Tl), and maximum-end-time (ETmax) of the at least one of the one or more signal groups (S1 to S4);
provide the SPaT message to a receiving device associated with the vehicle, and
if the time interval between an actual point in time (t−1) and the current predicted likely-time (Tl-1) is below a predefined filter time interval (FTI)
a current maximum-end-time prediction (Etmax-1) which predicts a later point in time than a previous filtered maximum-end-time prediction (Etmax-2) is filtered out,
a current minimum-end-time prediction (Etmin-1) which predicts an earlier point in time than the previous filtered minimum-end-time prediction (Etmin-3) is filtered out,
else,
only filtering out the current minimum-end-time prediction (Etmin-2) which predicts an earlier point in time than the previous filtered minimum-end-time prediction (Etmin-3).

### Claim 9 (depends on 8)

The computer program product of claim 8, wherein:
before applying filtering,
a first predetermined safety margin (SM-max) is added to a latest filtered maximum-end-time prediction, and
a second predetermined safety margin (SM-min) is subtracted from the latest filtered maximum-end-time prediction,
wherein the first and second predetermined safety margins are determined based on a comparison between historic predictions and corresponding actual switching time points and are selected to limit a percentage of violating predictions to a percentage below a given threshold value, wherein violating predictions are predictions where an actual state switch of the corresponding signal group occurred before the predicted minimum-end-time or after the predicted maximum-end-time in a given time interval before the actual switching time point.

### Claim 10 (depends on 8)

The computer program product of claim 8, wherein the at least one quantile regression neural network is implemented using any of the following neural network architectures: a recurrent neural network, a multi-layer perceptron, and a convolutional neural network.

### Claim 11 (depends on 8)

The computer program product of claim 8, wherein a grid search or a random search of a parameter space is used to provide a plurality of layers of the at least one quantile regression neural network and a number of hidden units per layer.

### Claim 12 (depends on 8)

The computer program product of claim 8, wherein the at least one quantile regression neural network comprises:
a first neural network model trained for predicting state switching from a pass-state to a stop-state; and
a second neural network model trained for predicting state switching from the stop-state to the pass-state.

### Claim 13

A computer system for providing Signal Phase and Timing (SPaT) messages for influencing operation of a vehicle, wherein the SPAT messages include timing parameters for a minimum-end-time, likely-time, and maximum-end-time for a next signal switch of at least one signal group of a traffic light system affecting the vehicle, the computer system comprising:
an interface component configured to receive signal state data from a controller controlling one or more signal groups (S1 to S4) of a traffic light system, wherein the signal state data of a particular signal group (S*) comprises a current signal state (SD*) of the respective signal group (S*), and information about a time interval (Te*) which has elapsed since last state switch of the respective signal group (S*), and further configured to provide a SPAT message to a receiving device associated with the vehicle;
a data pre-processor configured to generate a feature vector (FV) for the one or more signal groups (S1 to S4) comprising for each of the one or more signal groups at least: a first value indicating the elapsed time (T*p) since a latest switch of the respective signal group (S*) to a pass-state (1), and a second value indicating the elapsed time (T*s) since a latest switch of the respective signal group (S*) to a stop-state (0);
at least one quantile regression neural network configured to receive a particular current feature vector (T*p, T*s) as test input, and to predict for at least one of the one or more signal groups (S1 to S4), as one or more outputs (O1), a low conditional quantile (Ql) representing the minimum-end-time (ETmin), a medium conditional quantile (Qm) representing the likely-time (Tl) and a high conditional quantile (Qh) representing the maximum-end-time (ETmax) for the at least one of the one or more signal groups (S1 to S4), wherein the at least one quantile regression neural network has been trained with stochastic gradient descent, on past feature vectors of at least a subset of the one or more signal groups of the traffic light system and respective low, medium and high conditional quantile outputs, using a loss function being a sum of individual loss functions over the at least one of the one or more signal groups (S1 to S4), and corresponding one or more conditional quantiles, wherein the individual loss functions are hinge loss functions (HL);
a message composer configured to compose the SPAT message including the predicted minimum-end-time (ETmin), likely-time (Tl), and maximum-end-time (ETmax) of the at least one of the one or more signal groups (S1 to S4); and
a filter module configured to filter out a current minimum-end-time prediction (Etmin-1 which predicts an earlier point in time than a previous filtered minimum-end-time prediction (Etmin-3), and to filter out a current maximum-end-time prediction (Etmax-1) which predicts a later point in time than the previous filtered maximum-end-time prediction (Etmax-2), if the time interval between an actual point in time (t−1) and the current predicted likely-time (Tl-1) is below a predefined filter time interval (FTI); else, only to filter out the current minimum-end-time prediction (Etmin-2) which predicts an earlier point in time than the previous filtered minimum-end-time prediction (Etmin-3).

### Claim 14 (depends on 13)

The computer system of claim 13, further comprising a safety margin module configured:
to add a first predetermined safety margin SM-max to a latest filtered maximum-end-time prediction (Etmax-3) before applying the filter module, resulting in a new latest filtered maximum-end-time prediction (Etmax-3s), and
to subtract a second predetermined safety margin (SM-min) from a latest filtered low minimum-end-time prediction (Etmin-3) before applying the filter module resulting in a new latest filtered minimum-end-time prediction (Etmax-3s),
wherein the first and second predetermined safety margins are determined based on a comparison between historic predictions and corresponding actual switching time points and are selected to limit a percentage of violating predictions to a percentage below a given threshold value, wherein violating predictions are predictions where an actual state switch of the corresponding signal group occurred before the predicted minimum-end-time or after the predicted maximum-end-time in a given time interval before the actual switching time point.

### Claim 15 (depends on 13)

The computer system of claim 13, wherein the at least one quantile regression neural network is implemented using any of the following neural network architectures: a recurrent neural network, a multi-layer perceptron, and a convolutional neural network.

### Claim 16 (depends on 13)

The computer system of claim 13, wherein a grid search or a random search of the parameter is used to provide a plurality of layers of the at least one quantile regression neural network and a number of hidden units per layer.

### Claim 17 (depends on 13)

The system of claim 13, wherein the at least one quantile regression neural network comprises:
a first neural network model trained for predicting state switching from the pass-state to the stop-state; and
a second neural network model trained for predicting state switching from the stop-state to the pass-state.