WO2022096130A1

WO2022096130A1 - Device and method for vehicular traffic signal optimization

Info

Publication number: WO2022096130A1
Application number: PCT/EP2020/081417
Authority: WO
Inventors: Cristian AXENIE; Stefano BORTOLI; Goetz BRASCHE
Original assignee: Huawei Cloud Computing Technologies Co., Ltd.
Priority date: 2020-11-09
Filing date: 2020-11-09
Publication date: 2022-05-12
Also published as: CN116508081A

Abstract

Described is a vehicular traffic control device (1801) configured to control traffic signals (303, 1001) by: storing (1701) a plurality of datasets, each dataset comprising, for a respective past period of time, representations of (i) conditions of the traffic signals over that past period of time, (ii) first traffic conditions prior to that past period of time and (iii) second traffic conditions subsequent to that past period of time; detecting (1702) current traffic conditions; comparing (1703) the current traffic conditions to the first traffic conditions; and in dependence on that comparison, selecting (1704) one of the datasets and causing the traffic signals to adopt the conditions represented in that dataset. This may allow for the optimization of traffic signals by exploiting decision making in previous scenarios given the current traffic situation.

Description

DEVICE AND METHOD FOR VEHICULAR TRAFFIC SIGNAL OPTIMIZATION

TECHNICAL FIELD

This disclosure relates to vehicular traffic control, in particular to optimizing traffic control decision making and signals by exploiting decision making in previous scenarios given the current traffic situation.

BACKGROUND

Traffic congestion poses serious challenges for city infrastructure facilities and also affects the socio-economic lives of residents due to time wasted whilst waiting in traffic.

A traffic responsive signal control system is a means of adjusting the setting of traffic signals (for example, adjusting cycles, green splits and offsets) located in a road traffic infrastructure. The settings optimize a given objective function, such as minimizing travel time or stops, in real-time based upon estimates of traffic conditions in order to improve the flow of vehicles, as depicted in Figure 1. The associations between the control signals (for example, the green time I red time of traffic lights) and the measured outcomes (for example, the number of vehicles passing a point in an arbitrary and fixed amount of time , referred to as the traffic flow) describe the temporal and spatial dynamics of traffic on a road, intersection, or region. To optimize flow and minimize delay time, the system needs to find the best traffic signal timing.

In Figure 2, a typical road traffic scenario is depicted. For a single intersection, the traffic dynamics (i.e. traffic light timing for each of the Red (R), Green (G), and Yellow (Y), determine flow control from transients of vehicles transiting (for example, start loss, effective flow) sequenced by their order and distance to the stop light (for example, queue discharge rate) up to continuous green time where the saturation flow can be reached. In Figure 2, is the yellow plus all-red interval that occurs between phases of a traffic signal to provide for clearance of the intersection before conflicting movements are released; I2 is the time between signal phases during which an intersection is not used by any traffic; dj is the component of delay that results when a control signal causes a lane group to reduce sped or to stop, which is measured by comparison with the uncontrolled condition; Ci is the total time for a signal to complete one cycle length; Rj is the period in the signal cycle during which, for a given phase or lane group, the signal is red; Gj is the duration of the green indication for a given movement at a signalized intersection; and h is the additional time consumed by the first few vehicles in a queue at a signalized intersection above and beyond the saturation headway, because of the need to react to the initiation of the green phase and to accelerate.

In order to minimize waiting time, methods to optimize the sequence of traffic light signals among adjacent intersections have been considered. These methods need to assess the number of vehicles passing and adjust the timing of the traffic lights accordingly in order to maximize the flow.

One reason that many previous models do not excel in practice in large-scale scenarios is that they fail to fully exploit the causal coupling or associations and intrinsic dynamics of the traffic flow when controlling traffic signals. Moreover, typical (static) systems do not scale as the computational effort increases, both in describing the system and its continuous changes (variance).

It is desirable to develop a device and method for traffic control that overcomes these problems.

SUMMARY

According to a first aspect there is provided a vehicular traffic control device configured to control traffic signals by: storing a plurality of datasets, each dataset comprising, for a respective past period of time, representations of (i) conditions of the traffic signals over that past period of time, (ii) first traffic conditions prior to that past period of time and (iii) second traffic conditions subsequent to that past period of time; detecting current traffic conditions; comparing the current traffic conditions to the first traffic conditions; and in dependence on that comparison, selecting one of the datasets and causing the traffic signals to adopt the conditions represented in that dataset.

This may allow for the optimization of traffic signals by exploiting decision making in previous scenarios given the current traffic situation.

The representations of the first and second traffic conditions may comprise data indicative of traffic flow. The representation of the conditions of the traffic signals may comprise data indicative of green time (or another flow permission mode) of the traffic signals. The device may be configured to select one of the datasets in further dependence on a predetermined objective. The device may be configured to, based on the current traffic conditions, select one of the datasets in order to optimize the objective. The objective may be maximising traffic flow. Therefore, the device may allow the flow of vehicles to be maximized, which may reduce traffic congestion and waiting time.

The detected current traffic conditions may comprise a timeseries of traffic data. This may allow the traffic conditions to be evaluated over a period of time.

The detected current traffic conditions may be derived from sensory data collected from one or more sensors, each sensor comprising one of a camera and an induction loop. Such sensors may be located in the environment and may allow the current conditions to be captured. For example, a camera may record a series of images which can be used to determine traffic flow (defined as the number of cars passing a point per arbitrary fixed unit of time).

The device may be configured to detect current traffic conditions from a partial timeseries of input data. This may allow the current traffic conditions to still be detected using a broken or intermittently operating sensor.

The device may be configured to detect current traffic conditions according to one of multiple traffic contexts. The multiple traffic contexts may comprise green time of a traffic lane and traffic flow of a traffic lane. The device may retrieve a complete context (for example, all flows and green time sequence) upon presentation of only partial information from that data (i.e. from broken traffic sensors or failing traffic lights). Alternatively, the device can learn memories such as to recall an associated data context from one category (i.e. green time of a lane) upon presentation of data context from another category (i.e. traffic flow on a lane).

The plurality of datasets may be learned using an artificial intelligence model. This may allow associations between data to be learned.

The artificial intelligence model may be a neural network. This may be a convenient implementation. For example, the neural network may be a recurrent neural network, such as a Hopfield network, that uses correlation learning to converge towards a solution in which the weights minimize an energy function of the constraints in traffic measurements (for example, constraints between the measured flow and green time allocated). The device may be configured to learn a mapping between the conditions of the traffic signals at that past period of time and the second traffic conditions subsequent to that past period of time. This may allow associations to be learned that can be subsequently used by the device.

The device may be configured to iteratively update parameters of the model over time in dependence on the detected current traffic conditions. The device may therefore allow for adaptation in real-time to changes, which is an important feature for traffic control. This may result in improved performance.

The device may be further configured to generate, in dependence on the selected dataset, a time plan for the traffic signals. The time plan may be for a future period of time. The future period of time may be subsequent to the period of time in which the current traffic conditions are detected. The output of the device can therefore be applied to a traffic control unit configured to update traffic signals in order to maximize the traffic flow based on the stored datasets and the current conditions.

The device may be implemented by one or more computer processors. This may allow the device to be deployed in dedicated hardware, or in the cloud.

According to a second aspect there is provided a method for controlling vehicular traffic signals, the method comprising: storing a plurality of datasets, each dataset comprising, for a respective past period of time, representations of (i) conditions of the traffic signals over that past period of time, (ii) first traffic conditions prior to that past period of time and (iii) second traffic conditions subsequent to that past period of time; detecting current traffic conditions; comparing the current traffic conditions to the first traffic conditions; and in dependence on that comparison, selecting one of the datasets and causing the traffic signals to adopt the conditions represented in that dataset.

The representations of the first and second traffic conditions may comprise data indicative of traffic flow. The representation of conditions of the traffic signals may comprise data indicative of green time (or another flow permission mode) of the traffic signals.

The method may comprise selecting one of the datasets in further dependence on a predetermined objective. The method may comprise, based on the current traffic conditions, selecting one of the datasets in order to optimize the objective. The objective may be maximising traffic flow. Therefore, the method may allow the flow of vehicles to be maximized, which may reduce traffic congestion and waiting time.

The detected current traffic conditions may be derived from sensory data collected from one or more sensors, each sensor comprising one of a camera and an induction loop. The data received from each of the sensors may comprise a timeseries of values. This may allow data generated continuously by the sensors to measure traffic metrics as time progresses.

The method may comprise detecting current traffic conditions from a partial timeseries of input data. This may allow the current traffic conditions to be detected using a broken or intermittently operating sensor.

The method may comprise detecting current traffic conditions according to one of multiple traffic contexts. The multiple traffic contexts may comprise green time of a traffic lane and traffic flow of a traffic lane. The method may comprise retrieving datasets having a complete context (for example, flows and green time sequences) upon presentation of only partial information from that data (i.e. from broken traffic sensors or failing traffic lights). Alternatively, memories may learned such as to recall an associated data context from one category (i.e. green time of a lane) upon presentation of data context from another category (i.e. traffic flow on a lane).

The artificial intelligence model may be a neural network. This may be a convenient implementation. For example, the neural network may be a recurrent neural network, such as a Hopfield network, that uses correlation learning to converge towards a solution in which the weights minimize an energy function of the constraints in traffic measurements (for example, constraints between the measured flow and green time allocated).

The method may comprise learning a mapping between the conditions of the traffic signals at that past period of time and the second traffic conditions subsequent to that past period of time.

The method may comprise iteratively updating parameters of the model over time in dependence on the detected current traffic conditions. The method may therefore allow for adaptation in real-time to changes, which is an important feature for traffic control. This may result in improved performance.

The method may comprise generating, in dependence on the selected dataset, a time plan for the traffic signals. The time plan may be for a future period of time. The future period of time may be subsequent to the period of time in which the current traffic conditions are detected. The output of the device can therefore be applied to a traffic control unit configured to update traffic signals in order to maximize the traffic flow based on the stored datasets and the current conditions.

The vehicular traffic signals are preferably traffic lights. However, other traffic signals may be used.

BRIEF DESCRIPTION OF THE FIGURES

The present disclosure will now be described by way of example with reference to the accompanying drawings. In the drawings:

Figure 1 shows a graphical depiction of a regional traffic control system, local traffic lights and flow control.

Figure 2 shows an overview of a generic traffic light signal control system and associated internal dynamics.

Figure 3 shows an overview of the architecture used by the device.

Figure 4 shows the modules comprising the device.

Figure 5 illustrates the internal neural network dynamics.

Figure 6 illustrates a sample neural network structure and attractor/memory dynamics explanation.

Figure 7 shows an exemplary neural network and mapping learning rule to traffic dynamics.

Figure 8 illustrates the memory dynamics.

Figure 9 shows an example of the runtime functionality of the device described herein. Figure 10(a) shows an exemplary instantiation showing the regional layout of a road traffic infrastructure having a plurality of traffic signals each located at a respective road traffic intersection. Figure 10(b) shows how the infrastructure is modelled as a 5x5 (N=5) urban road infrastructure described by a coupled system of oscillators used to control flow in the road network.

Figure 11 shows an exemplary instantiation of the device showing the network internals.

Figure 12 shows an exemplary instantiation of the device showing network learning of memory patterns from timeseries.

Figure 13 shows an exemplary instantiation of the device showing the memory recall capacity at scale.

Figures 14(a)-14(c) show an exemplary instantiation of the device showing the heteroassociative memory mode. There are three inputs: the keys (M = 2, control signals) as shown in Figure 14(a), the values (K = 3, traffic flow) as shown in Figure 14(b), and a learning signal, as shown in Figure 14(c).

Figure 15(a) shows an exemplary instantiation of the device showing the heteroassociative memory mode analysis visualization.

Figure 15(b) shows an exemplary instantiation of the device showing the heteroassociative memory mode analysis.

Figure 16 shows preliminary performance evaluation of the method described herein.

Figure 17 shows a flowchart illustrating the steps of a method for for controlling vehicular traffic signals.

Figure 18 shows an example of a device configured to implement the method described herein.

DETAILED DESCRIPTION

Embodiments of the present disclosure relate to the area of traffic control, in particular to optimizing traffic control decision making by exploiting decision making in previous scenarios given the current traffic situation. In some embodiments, the device described herein can also learn contextual memories (associations) for traffic signal optimization. A typical traffic control system senses various traffic variables (such as the traffic flow, or the queue of waiting cars) in order to compute the timing of the traffic signals that maximizes the flow of vehicles. In order to optimize traffic control, the device described herein can exploit previously acquired knowledge of the actions (for example, traffic signal timing configuration, or green time) that were taken and what the measured consequence was (for example, what was the measured traffic flow) as a result of this action and can apply this to the current traffic conditions to optimize an objective, such as maximizing the traffic flow. This can make the control loop faster by alleviating the need for re-computing a signal, through recalling the corresponding association, or memory.

Traffic data is generally received as a timeseries, which is a sequence of time indexed values (i.e. tuples containing various types of data, such as the number of vehicles, speed of vehicles etc.) that are collected from various sources (for example, sensors such as cameras, street induction loop sensors, pollution sensors, noise sensors, weather sensors) in a chronologically ordered fashion. These sensors are typically spatially distributed in urban infrastructures.

Despite being highly nonlinear, traffic dynamics is regular on certain timescales. Such regularities and the available sensory data can be used to judiciously extract traffic contexts that can be subsequently used in optimizing traffic situations.

The richness of spatial and temporal information describing road traffic, together with the control signals generated by the traffic signals, describe complex dynamics. Such dynamics can be captured through associations, which may be obtained by learning pairs of the action taken (such as traffic light timing) and the measured outcome (such as the flow of vehicles). Despite its highly nonlinear nature, traffic is described by a regular shape at a high-granularity (e.g. the daily profile) or medium granularity (e.g. mornings and afternoons). These regularities can support learning associations that can speed-up the capability of the traffic control device to adapt to changes and improve control objectives.

The device described herein may aggregate various road traffic variables or timeseries describing the current traffic context (such as vehicular traffic flows determined from sensor data) and the generated control signals; learn associations from such aggregated actions and consequences of the traffic situation; and, finally, given information sensed from the current traffic (for example, the current traffic flow), predict the best control signals for the traffic signal timing based on the learned associations from previous experience. As mentioned above, this may be done in order to optimize an objective, such as maximising traffic flow. In summary, the device is configured to store a plurality of datasets, also referred to herein as memories or associations. Each dataset comprises, for a respective past period of time, representations of (i) conditions of the traffic signals over that past period of time, (ii) first traffic conditions prior to that past period of time and (iii) second traffic conditions subsequent to that past period of time. The memories may be stored locally within the device or may be stored remotely, for example in the cloud, and retrieved by the device as required.

The device then detects the current traffic conditions. The current traffic conditions may be evaluated over a period of time, for example 5 minutes or 30 minutes. The detected current traffic conditions may comprise a timeseries of traffic data. For example, the device may detect the current traffic conditions by receiving sensory data from one or more sensors, such as a camera or induction loop, which may give an indication of the current traffic flow. As discussed in more detail later, the device may detect current traffic conditions from a partial timeseries of input data. For example, from data collected by a broken or intermittently operating sensor.

The device then compares the current traffic conditions to the first traffic conditions, prior to the respective past period of time for multiple datasets. In dependence on that comparison, the device selects one of the datasets and causes the traffic signals to adopt the conditions represented in that dataset.

The device may be configured to select one of the datasets in further dependence on a predetermined objective, and may select one of the datasets in order to optimize the objective. For example, the predetermined objective may be optimizing traffic flow.

The device may be configured to detect current traffic conditions according to one of multiple traffic contexts. The multiple traffic contexts may comprise green time of a traffic lane and traffic flow of a traffic lane. As will be described in more detail later, the device can use an autoassociative or a heteroassociative mode.

The plurality of datasets may be learned using an artificial intelligence model, in particular a neural network, as will be described in more detail below. The device may be configured to learn a mapping between the conditions of the traffic signals at that past period of time and the second traffic conditions subsequent to that past period of time. To improve the accuracy of the model, the device may be configured to iteratively update parameters of the model over time in dependence on the detected current traffic conditions.

The device may be further configured to generate, in dependence on the selected dataset, a time plan for the traffic signals. This may allow a time plan for the signals to be generated that maximizes the flow of vehicles. Such a method can therefore be used to control traffic signal sequences to subsequently maximize the flow of vehicles and alleviate congestion.

The optimization of road traffic flow is preferably performed continuously, accommodating the incoming stream of current traffic data (such as detected traffic flow, green time etc.) and updating the learned model accordingly. The considered scenario is well-suited for such a device that is capable of continuously computing optimal timing describing the traffic situation, learning to associate the variables describing the dynamics of the traffic control loop (i.e. measured flow, traffic signal timing) and adapting to changes for improved road traffic control.

As the data enters the device, the computations have a limited time span to be handled, bounded resource allocation and execution time, yielding for simple operations to be executed. Embodiments of the present disclosure may overcome the resource greedy, computationally expensive and complex state-of-the-art approaches, such as complex analytical flow models based on differential equations and numerical methods, empirical methods, nonlinear control systems,

The overall system 300 has a modular structure, depicted in Figure 3. The modules operate in concert on the incoming sensory timeseries data. The elements of the modules will be described in more detail later with reference to Figure 4. Generally, the system has a flexible infrastructure that acquires measured sensory data describing the traffic situation (for example, flow of cars) or other sensors, and traffic signal control timeseries. Such contextual information contains intrinsic association patterns that the context memory learning module 301 extracts and represents as memories in an efficient computation substrate. The system has an interpretable output/prediction applicable to any road traffic layout, size and architecture (i.e. number of lanes per direction etc.), functionality supported by the context memory prediction module 302. The output from module 302 can be used to control traffic signals 303, which may be located at road traffic intersections 304 (an intersection of two or more roads). The current traffic data may be determined by at least one sensor located at one or more of the intersections.

In order to enhance its general applicability, the device does not need to explicitly encode traffic constraints, but exploits and learns the temporal correlation and causal association between data, such as the green time control signal (or other flow permission mode of the traffic signals) and the flow of vehicles passing given the allocated green time. The device may minimize execution costs by exploiting the dynamics of causal association intrinsic in traffic signals in a learning system. More precisely, traffic data pairs, i.e. action (for example, green time of traffic lights) and outcome or consequence (e.g. flow/number of vehicles passing, or other quantities) are translated into a contextual representation, depicting the memory of that specific traffic situation at particular past period of time.

When deployed and used for prediction, the device may be input only with the currently measured traffic variables (such as traffic flow) and can recall, given the learnt memories depicted in the multiple datasets, what would be a suitable control signal configuration or green time for the current situation. The device has a flexible infrastructure to learn and store associations among measured traffic data and the corresponding traffic control signal generating it and recall them for speeding up computation in similar traffic situations. The device therefore provides a lightweight learning system capable of solving local or global traffic dynamics by exploiting the causal coupling in traffic context, among measured traffic data and corresponding control signals.

An overall aim may be to control the timing of the traffic signals (for example, the green time of traffic lights, or the timing of another flow permission mode of traffic signals) in order to achieve maximum throughput by exploiting the causality among the action and the outcome temporal correlations in a dynamic and adaptive system.

In terms of implementation details, starting from the generic architecture of the device in Figure 4, the device may comprise the following sub-modules.

An example of the context memory learning module is shown at 401 in Figure 4. The learning model is capable of capturing the dynamics and constraints of the local and global interactions in the road traffic infrastructure between measured traffic data and traffic control signals, shown at 403.

The system receives measured data 403 from a road traffic infrastructure, such as measured flow and the allocated green time (or other flow permission mode) of a lane, direction or total (in all lanes and directions) through an intersection. The full traffic context data is fed to a recurrent neural network (for example, a Hopfield Network, as described below) that uses correlation learning to converge towards a solution in which the weights minimize an energy function of the constraints in traffic measurements (i.e. constraints between the measured flow and green time allocated). The neural network implements temporal dynamics given by a differential equation that uses the recurrent (temporal co-activation) character of the system to converge to a solution of minimum energy depicting a certain traffic context. The system is trained on historical data. Basically, these are different traffic data context patterns (for example, tuples of traffic flow and corresponding allocated green time). The system learns to store independent points in state space (memories) that represent the specific correlation patterns in the input data. The system can also practically embed spatial information about the road infrastructure (adjacency matrices) in the temporal dynamics, making the dynamics rich and more specific, which is helpful to build more distinct memories.

The device may therefore advantageously learn data patterns as associations. The device can learn associations between signals and the observed outcome compared to the traffic conditions beforehand (i.e., conditions of the traffic signals over a past period of time, first traffic conditions prior to that past period of time and second traffic conditions subsequent to that past period of time) to speed up operation when facing new traffic situations. The device may recall these associations as memories. The idea of associative memory is that when the current data is presented (for example, a partial tuple, such as just the observed outcome), the actual memory that is most like the data can still be recapitulated.

As mentioned above, one way that this may be done is using a Hopfield Network. A Hopfield Network is a recurrent neural network with as many neurons as input size and dynamics given as a dynamical system, as shown in Figure 5. Attractors, in Hopfield Networks terminology, correspond to memories. In this case a memory is a tuple (for example, measured traffic data (flow), and control signal (green time) of a traffic signal). There is an underlying energy function describing the activity dynamics. Starting in any initial state, the state of the system evolves to a final state that is a minimum of the energy function. The phenomenon of associative memory matches the idea of dynamics controlled by an energy function or potential, as described in Figure 6.

Mapping the dynamics of road traffic control to the network dynamics can be performed using an intuitive equation describing how each quantity (i.e. flow, green time) are included in the dynamics of the network, as shown in Figure 7, upper panel. As previously mentioned, the mechanism behind learning associations is based on energy minimization, a process intrinsic to Hopfield Networks, as illustrated in Figure 7, lower panel.

The device can therefore map sensory data (from which traffic condition information, such as flow, may be derived) and control signal values to a learning mechanism capable of extracting the causal association and temporal correlations that can subsequently support prediction.

The context memory prediction module is shown at 402 in Figure 4. This module deploys the trained learning module 401 and can also be used separately. This module of the device employs a trained recurrent neural network that stores the various learnt memories, in order to infer or recall full information given the current traffic conditions, shown at 404, including partial or incomplete input patterns. For instance, to infer green time given current traffic flow or infer traffic flow given allocated green time.

The system can also advantageously recall data patterns (memories) given noisy or incomplete input patterns. Recall that attractors correspond to memories. In this case a memory is a tuple (for example, measured traffic data, control signal (for example, green time) of traffic signal). The phenomenon of associative memory matches the idea of dynamics controlled by an energy function or potential. The energy function terms describe how the system dynamics penalizes traffic constraint violation (e.g. larger green times as a preset value, negative green time, or impossible flow value per sample time) and the network enforces the minimization of the objective function/ energy, alleviating the need for costly optimization (see Figure 7, lower panel).

In one example, consider a set of states M that are desired as memories. Suppose network weights are chosen so that each memory vector M is a local minimum of the energy function E. Starting with partial information about some memory k means starting relatively nearer to Mk than to other memories (basin), as shown in Figure 8. This starting state is then likely to be within the ‘valley’ of the E that has Mk as its lowest point. If so, the dynamics result in the final state Mk, the correct memory reconstruction.

The device can therefore advantageously learn in two modes. The device can learn memories such as to retrieve a complete context (i.e. all stored flows and green time sequences) upon presentation of only partial information from that data (i.e. from broken traffic sensors or failing traffic lights), referred to as autoassociation. Alternatively, the device can learn memories such as to recall an associated data context from one category (i.e. green time of a lane) upon presentation of data context from another category (i.e. traffic flow on a lane), referred to as heteroassociation.

The aforementioned modules 401 and 402 can act in concert at system runtime, as depicted in Figure 9.

The runtime sequence and corresponding functional modules may therefore exploit learned memories from historical data associations among measured data and control signals to infer new control signals given measured traffic data, and can use both autoassociative and heteroassociative modes. The device can also learn from historical data associations among patterns of measured data and control signals by training a neural network that exploits the intrinsic traffic constraints and temporal dynamics among traffic context sequences. The computation cost is minimal, accounting to essentially a dot product between the input traffic pattern and the weight matrix of the trained recurrent neural network. The capacity of memories to store is given by the number of neurons, which essentially share the same dynamic equations that describe temporal evolution and temporal correlation extraction.

The concept is applicable to a core business intelligence solution vertical (i.e. traffic optimization) which requires fast and efficient computation of traffic light signal to optimize for flow of vehicles and the device can be deployed without reconfiguration to arbitrary size road traffic situations and at every granularity (e.g. lane, intersection level).

An exemplary implementation is described below with reference to Figures 10 to 15.

The operation of the device is instantiated in a sample traffic optimization scenario and the functionality and performance will now be described step by step in the computation pipeline of the disclosure. Both autoassociative and heteroassociative operation modes are described.

As shown in Figure 10(a), the regional layout of the road infrastructure comprises a plurality of traffic signals 1001 , each traffic signal being located at a respective road traffic intersection 1002. In this example, as shown in Figure 10(b), a portion of this infrastructure is modelled as a 5x5 (N=5) urban road infrastructure described by a coupled system of oscillators used to control flow in the road network. The 5x5 (N=5) urban road infrastructure used in this example is used to control flow in the entire road network. In this exemplary network, traffic signals 1001 are located at intersections 1002. Sample data for the autoassociative mode of operation of the system along with the mapping of a number of patterns of timeseries to train the system are depicted in Figure 11. This process replaces the computationally expensive optimization and ensures that the system converges to a stable state, which describes the best traffic light signals.

The configuration in Figure 11 is instantiated for M = 3 patterns of traffic light control signal and flow timeseries. The module analyzes how well the network is able to translate noisy vectors into clean versions (i.e. recover the memory). One of the patterns (P3) is chosen and is randomly distorted by a small amount (standard deviation noise = 0.6), and the network is initialized in a state corresponding to this distorted pattern. The similarity of the network's state to each of the patterns corresponding to a desired stable point/memory can then be plotted, as shown in Figure 12. The network's state preferably converges towards the pattern that it is closest to (similarity).

Figure 13 demonstrates how well the network is able to generalize to M = 10 patterns when the architecture is extended, with the goal of avoiding placing any restriction on the kinds of vectors of traffic data that can serve as stable points in the network's state space. This may allow the network to handle interfering inputs (i.e. mixed patterns or weighted combinations). In this example, the patterns P1 and P10 are chosen, and are randomly distorted by a small amount (standard deviation noise=0.65). They are then combined linearly (here using input = 0.6*P1 + 0.4*P10). The network then converges to the most prominent pattern and also, due to interference, partially to other similar patterns, as depicted in Figure 13.

For the heteroassociative traffic memory instantiation, a 5x5 (N=5) urban road infrastructure is used to control flow in the entire road network. Being able to learn an input-output mapping (or a heteroassociative memory) is useful for storing and recalling associations. There are three inputs: the keys (M = 2, control signals) as shown in Figure 14(a), the values (K = 3, traffic flow) as shown in Figure 14(b), and a learning signal, as shown in Figure 14(c). The model is run continuously in two phases: the first phase learns the set of associations, and the second phase tests recall of the memory by just switching the learning signal from 0 to -1 (i.e. inhibit learning).

The overall learning process is depicted in Figure 15(a), where the training errors decrease during learning as the network extracts the temporal associations among the keys and the values, corresponding to traffic signals and flow of cars. During recall, or inference, the spikes describe the moment when the network is presented with a new pattern (i.e. a new dataset).

To visually illustrate how the global network dynamics work, Figure 15(b) shows the dynamics of the heteroassociative operation along with the localization of the memories before and after training. In this example, each memory comprises a pair of traffic signal and traffic flow values; in other words, the causal associations describing road traffic.

As shown in Figure 16, a preliminary performance evaluation was performed against state-of- the art methods used in typical systems and extrapolated, looking at the relationship between run-time and demand factor of a certain sequence of intersections. As shown by the black arrow, the cumulative vehicle delay versus demand factor of the method described herein is superior to state-of-the-art methods at the highest demand factor (i.e. 10s vs 31.8s). The table depicts the histogram values in the middle panel. Moreover, in some embodiments, the runtime of the method described herein is one order of magnitude smaller than the Dynamic Programming (MILP) method, a de-facto standard.

In summary, the traffic context learning and memory system learns from historical data, mainly sensory data (i.e. traffic flow timeseries on a certain lane I direction I in an intersection) and control data (i.e. allocated green time for each lane/direction) and builds associations (i.e. memories) - essential, which action (i.e. control signal) generated a respective outcome (i.e. flow of cars).

Figure 17 shows a flowchart illustrating a method for controlling vehicular traffic signals. At step 1701 , the method comprises the method comprises storing a plurality of datasets, each dataset comprising, for a respective past period of time, representations of (i) conditions of the traffic signals over that past period of time, (ii) first traffic conditions prior to that past period of time and (iii) second traffic conditions subsequent to that past period of time. At step 1702, the method comprises detecting current traffic conditions. At step 1703, the method comprises comparing the current traffic conditions to the first traffic conditions. At step 17-4, the method comprises in dependence on that comparison, selecting one of the datasets and causing the traffic signals to adopt the conditions represented in that dataset.

Figure 18 shows an example of a device 1801 configured to use the method describe herein for controlling vehicular traffic. Such a device 1801 typically includes some onboard processing capability. This could be provided by the processor 1802. The device may also comprise a memory 1803 for storing the datasets and/or for storing the models used by the device.

The device may also comprise a transceiver 1804 capable of communicating over a network with other entities 1805, 1806. Those entities may be physically remote from the device 1801. The network may be a publicly accessible network such as the internet. The entities 1805, 1806 may be based in the cloud. Entity 1805 is a computing entity. Entity 1806 is a command and control entity. These entities are logical entities. In practice they may each be provided by one or more physical devices such as servers and data stores, and the functions of two or more of the entities may be provided by a single physical device. Each physical device implementing an entity comprises a processor and a memory. The devices may also comprise a transceiver for transmitting and receiving data to and from the transceiver 1804 of device 1801 . The memory stores in a non-transient way code that is executable by the processor to implement the respective entity in the manner described herein.

The command and control entity 1806 may in some implementations be used to train the artificial intelligence models used in the device. This is typically a computationally intensive task, even though the resulting model may be efficiently described, so it may be efficient for the development of the algorithm to be performed in the cloud, where it can be anticipated that significant energy and computing resource is available. It can be anticipated that this is more efficient than forming such a model at a typical imaging device. In one implementation, once the algorithms have been developed in the cloud, the command and control entity can automatically form a corresponding model and cause it to be transmitted to the device. In this example, the model is implemented at the device 1801 by processor 1802.

In another possible implementation, traffic data may be received by the device and may be sent by the transceiver 1804 to the cloud for processing. The resulting models could then be sent back to the device 1801 , as shown at 1807 in Figure 18.

Therefore, the method may be deployed in multiple ways, for example in the cloud, on the device, or alternatively in dedicated hardware. As indicated above, the cloud facility could perform training to develop new models or refine existing ones. Depending on the compute capability near to the data corpus, the training could either be undertaken close to the source data, or could be undertaken in the cloud, e.g. using an inference engine. The method may also be implemented at the device, in a dedicated piece of hardware, or in the cloud.

The device and method described herein may in some implementations solve the constrained optimization problem intrinsic in the area of traffic optimization (i.e. maximize traffic flow minimizing the waiting time at a stop light). The device can map the constrained optimization to a generic temporal dynamics (i.e. loss function), avoiding explicit problem formulation and optimization. The device uses efficient recurrent neural networks whose dynamics enforce the intrinsic constraints in traffic optimization when learning the memories.

Embodiments of the present disclosure may therefore provide a new approach for controlling traffic signal offset in an urban region exploiting the temporal correlation between contextual traffic information, sequences of control signals (such as green time allocation) and measured outcomes (for example, the resulting traffic flow). This may enable exploitation of temporal correlation between road traffic context in a system capable of learning associations between the traffic elements (for example, the control signals and the measured outcome) and use the learned associations to recall or reconstruct partial patterns of traffic context or infer the corresponding missing part of a context (for example, infer green time given traffic flow or infer traffic flow given allocated green time).

This may also enable the automatic learning of traffic constraints from measured data to extract the underlying associations from historical data in order to speed up computation for new data and adapt to changing situations in traffic, and enable the efficient processing of local and global constraints of traffic using a neural network (i.e. the same processing paradigm can be applied at lane, direction or intersection level in an urban region). Further, the system may enable the use of learnt memories (associations) as means to control the offset signal in the region intersections and enable inexpensive computation, which avoids constrained optimization, by using the principles of temporal dynamics learning in recurrent neural networks.

The system can be deployed for multiple scenarios (independent of road geometry, size and configuration and available data) for which flexibility and scalability are required and can minimize costs associated with expensive computational systems in all road crosses by exploiting the intrinsic character of traffic flow context and capable of embedding spatial information to drive the system dynamics towards optimal control.

The device and method described herein may help to efficiently represent traffic context using the measured data (for example, traffic flow) and control signal (for example, allocated green time) in a system capable of learning multiple associations between such causal data. This may help to avoid costly optimization methods and control signal re-computation by exploiting previously learnt patterns of traffic data and infer, given partial information (i.e. either traffic flow or allocated green time ), what would be the best corresponding full context corresponding to the partial information. Scalability is supported by storing multiple memories (i.e. multiple full traffic contexts) and deployable at different granularities (for example, per lane, per direction, per intersection). The device and method may generally allow for efficient computation of novel traffic control signals (i.e. green time) that embed and exploit the intrinsic traffic constraints and physics without an explicit need to model the constraints. It may also provide a solution for road traffic infrastructure and geometry agnostic control, allowing for the modelling of traffic dynamics at different levels and using the same mechanisms for learning, independent of the number of lanes, or the intersection layout.

Where Hopfield networks are used, the peculiarities of Hopfield networks are exploited, but extended to allow for representation of various contextual traffic timeseries data in associations (associative memory) or in the form of heteroassociative memory (i.e. key - value pairs of traffic timing signal - flow of cars). The idea of associative memory is that when a memory clue is presented (e.g. partial tuple - just the observed outcome), the actual memory that is most like the clue will be recapitulated. The heteroassociative behavior assumes that the system is able to learn an input-output mapping (or a heteroassociative memory) useful for storing and recalling associations, as the ones between traffic light green time and flow of cars passing. The device may operate in both modes without modifications to the system, by employing an efficient implementation in (spiking) neural networks and avoids introducing the burden of constrained optimization typical for integer programming by exploiting the temporal dynamics of the Hopfield network. Regardless of the deployment scenario, embodiments of the disclosure may provide a learning mechanism for building memories of traffic contexts, a device capable of learning temporal correlations among control signals and measured outcome in associations that can be recalled from noisy or incomplete input data, and a device and method capable of describing traffic dynamics using recurrent neural networks implementing differential equations modelling and capturing traffic specificities. This may allow for fast and efficient modelling, representation, learning and prediction with fixed resource budget. Such a resource-light system (memory, size, computation) can be advantageously deployed on various edge or cloud devices.

The device may act as a cache mechanism in computing systems, capable of recognizing ongoing traffic patterns and matching them with traffic signal configurations performing well for that specific current situation. The proposed apparatus can be used with any traffic sensory data available. The device is supported by a flexible instrumentation which can help to ensure updates with low-latency, high incoming event rates and a fixed resource budget. Further, the device can be deployed in any type of intersection without pre-training, agnostic of the intersection layout, size, and available traffic data metrics. This offers major advantages in terms of deployment cost reduction because the underlying contextual learning mechanisms are efficient and can be transferred to new road infrastructures with different geometries but similar traffic dynamics (i.e. typical urban traffic dynamics).

The device and method described herein have the capability to optimize decision making time and “short-circuit” re-computation of control signals (such as green time allocation) by exploiting the previously learned patterns. Such memories of associations have the power to speed up operation when facing new traffic situations. The device can therefore advantageously exploit the learned context associations among patterns of traffic to infer what the traffic flow would be when a control signal is available, and what would be the green time for measured flow values.

The applicant hereby discloses in isolation each individual feature described herein and any combination of two or more such features, to the extent that such features or combinations are capable of being carried out based on the present specification as a whole in the light of the common general knowledge of a person skilled in the art, irrespective of whether such features or combinations of features solve any problems disclosed herein, and without limitation to the scope of the claims. The applicant indicates that aspects of the present invention may consist of any such individual feature or combination of features. In view of the foregoing description it will be evident to a person skilled in the art that various modifications may be made within the scope of the disclosure.

Claims

1. A vehicular traffic control device (1801) configured to control traffic signals (303, 1001) by: storing (1701) a plurality of datasets, each dataset comprising, for a respective past period of time, representations of (i) conditions of the traffic signals over that past period of time, (ii) first traffic conditions prior to that past period of time and (iii) second traffic conditions subsequent to that past period of time; detecting (1702) current traffic conditions (404); comparing (1703) the current traffic conditions to the first traffic conditions; and in dependence on that comparison, selecting (1704) one of the datasets and causing the traffic signals to adopt the conditions represented in that dataset.

2. The device (1801) as claimed in claim 1 , wherein the device is configured to select one of the datasets in further dependence on a predetermined objective.

3. The device (1801) as claimed in claim 2, wherein the device is configured to, based on the current traffic conditions (404), select one of the datasets in order to optimize the objective.

4. The device (1801) as claimed in any preceding claim, wherein the detected current traffic conditions (404) comprises a timeseries of traffic data.

5. The device (1801) as claimed in any preceding claim, wherein the detected current traffic conditions (404) are derived from sensory data collected from one or more sensors, each sensor comprising one of a camera and an induction loop.

6. The device (1801) as claimed in any preceding claim, wherein the device is configured to detect current traffic conditions from a partial timeseries of input data.

7. The device (1801) as claimed in any preceding claim, wherein the device is configured to detect current traffic conditions (404) according to one of multiple traffic contexts.

8. The device (1801) as claimed in claim 7, wherein the multiple traffic contexts comprise green time of a traffic lane and traffic flow of a traffic lane.

9. The device (1801) as claimed in any preceding claim, wherein the plurality of datasets are learned using an artificial intelligence model.

10. The device (1801) of claim 9, wherein the artificial intelligence model is a neural network.

11. The device (1801) as claimed in claim 9 or claim 10, wherein the device is configured to learn a mapping between the conditions of the traffic signals at that past period of time and the second traffic conditions subsequent to that past period of time.

12. The device (1801) as claimed in any of claims 9 to 11 , wherein the device is configured to iteratively update parameters of the model over time in dependence on the detected current traffic conditions.

13. The device (1801) as claimed in any preceding claim, wherein the device is further configured to generate, in dependence on the selected dataset, a time plan for the traffic signals.

14. The device (1801) as claimed in any preceding claim, wherein the device is implemented by one or more computer processors.

15. A method (1700) for controlling vehicular traffic signals (303, 1001), the method comprising: storing (1701) a plurality of datasets, each dataset comprising, for a respective past period of time, representations of (i) conditions of the traffic signals over that past period of time, (ii) first traffic conditions prior to that past period of time and (iii) second traffic conditions subsequent to that past period of time; detecting (1702) current traffic conditions; comparing (1703) the current traffic conditions to the first traffic conditions; and in dependence on that comparison, selecting (1704) one of the datasets and causing the traffic signals to adopt the conditions represented in that dataset.

16. The method (1700) as claimed in claim 15, comprising selecting one of the datasets in further dependence on a predetermined objective.

17. The method (1700) as claimed in claim 15 or claim 16, wherein the method comprises, based on the current traffic conditions, selecting one of the datasets in order to optimize the objective.

18. The method (1700) as claimed in any of claims 15 to 17, wherein the detected current traffic conditions (404) comprises a timeseries of traffic data.

19. The method (1700) as claimed in any of claims 15 to 18, wherein the detected current traffic conditions (404) are derived from sensory data collected from one or more sensors, each sensor comprising one of a camera and an induction loop.

20. The method (1700) as claimed in any of claims 15 to 19, wherein the method comprises detecting current traffic conditions from a partial timeseries of input data.

21. The method (1700) as claimed in any of claims 15 to 20, wherein the method comprises detecting current traffic conditions according to one of multiple traffic contexts.

22. The method (1700) as claimed in any of claims 15 to 21 , wherein the multiple traffic contexts comprise green time of a traffic lane and traffic flow of a traffic lane.

23. The method (1700) as claimed in any of claims 15 to 22, wherein the plurality of datasets are learned using an artificial intelligence model.

24. The method (1700) of claim 23, wherein the artificial intelligence model is a neural network.

25. The method (1700) as claimed in claim 23 or 24, wherein the method comprises learning a mapping between the conditions of the traffic signals at that past period of time and the second traffic conditions subsequent to that past period of time.

26. The method (1700) as claimed in any of claims 23 to 25, wherein the method comprises iteratively updating parameters of the model over time in dependence on the detected current traffic conditions.

27. The method (1700) as claimed in any of claims 15 to 26, wherein the method comprises generating, in dependence on the selected dataset, a time plan for the traffic signals. Road traffic infrastructure

Flow / control signal dynamics (associations and dependency)

Traffic flow control dynamics

300 303

Figure 4 404 405

I*** Tift

Time 402

Time

Context Memory Prediction Module atractor basin

^,nfcfWce

Consider a set of states M that are desired as memories (attractors) . Suppose network weights are chosen so that each memory vector M is a local minimum of the energy function E .

Starting with partial information about some memory k means starting relatively nearerto Mk than toother memories (basin). If so, the dynamics must result in the final Abate Mk, the correct memory reconstruction

Temporaldynamics with neurons

• |is>sWe9ffcw«w

• ^Mfedipttotenami

• Ijis thewesmtw neareoi ®w»J,

' -rn - I, - yu ,

The phenomenon of memory matches the idea of dynamics controlled by a energy function or potential.

403

SUBSTITUTE SHEET (RULE 26) (Q

C

CD

CH

Hopfield Network - Temporal dynamics with neurons

A; ^xi

— _ 4. X* I + 1 j

• \’_} denotes output of the /-th neuron

• ij is the state of the neuron

• is direct input to the neuron

• is the weight from neuron k to neuron j

• ,v is an activation function i.e _; V, = x(.i; i.

Ono steady state

AHraclinn doniam of one attractor

< State space (projected)

/ i Domain esjitrol (transition

L,

■ Transition from one domain to another)

— A-.- —. *ra ‘4 -. v.ra >

Ara

S Sttatyseme

State space (projected)

Internal Dynamics

Neural Network

he update rule and energy function are redefined here to incorporate a differential equationat describes the rate of change of the input to a given neuron with respect to time. here:

• Xj is the TLS control signal for lane j

• T_jk coupling of j with neighboring (incoming) lane k

• g(x_k ) cars flow (outgoing) from k

• I; cars from other directions his is a core modelling assumption that sets up in the description of traffic dynamics he energy function has the following form:

£ ₌ -Isiy v, v + -z [ g-¹ (vwv

2 « Jo

Figure 8

Sttae

SUBSTITUTE SHEET (RULE 26) Figure 9

403

SUBSTITUTE SHEET (RULE 26)

1002

Figure 11

SUBSTITUTE SHEET (RULE 26) Figure 12

SUBSTITUTE SHEET (RULE 26) Sample training data (M=10 patterns)

Traffic Flow Paterns across intersections

Similarity of Network State to Desired Stable Points

-

P2

- P3

: s ™ _P4

P6

P7

P8

04 -

PIO

0 20 • a MO

Figure 14

SUBSTITUTE SHEET (RULE 26)

SUBSTITUTE SHEET (RULE 26) Figure 15 (b)

Traffic Flow

Control Signals

I I i I i

SUBSTITUTE SHEET (RULE 26) State-of-the-art benchmark*

Dynamic Programming 6.9 H.5 1 1.6 2(1 29.3 bu 67.1 y- Learning 5.2 6.7 9,2 10.4 13.7 22.2 44.1

1 iio 23.6 25 26,2 27,2 90 ft .51.8

III OpiBlfc f

’Extrapolated (no one-to-one experiment) w,r.t Dynamic Programming timing evaluation,

Run-time is execution time of the effective calculation from measurements to new timing for traffic light,

*http;//rep05itoryAudelft,nl/

Figure 17

1701

1702

1703

1704

SUBSTITUTE SHEET (RULE 26) Figure 18

1805 1806

1802

1803

SUBSTITUTE SHEET (RULE 26)