CN114287022A

CN114287022A - Multi-step traffic prediction

Info

Publication number: CN114287022A
Application number: CN201980099755.2A
Authority: CN
Inventors: 克里斯蒂安·阿克塞尼; 拉杜·都铎兰; 斯蒂法诺·波托利; 穆罕默德·啊·哈吉·哈桑; 戈茨·布拉舍
Original assignee: Huawei Cloud Computing Technologies Co Ltd
Current assignee: Huawei Cloud Computing Technologies Co Ltd
Priority date: 2019-09-25
Filing date: 2019-09-25
Publication date: 2022-04-05
Also published as: WO2021058099A1

Abstract

A road traffic flow prediction system for predicting road traffic flow from data received from a plurality of sensors, each sensor of the plurality of sensors for acquiring road traffic data at a respective traffic location, wherein the system is for: receiving data from the sensor; implementing a machine learning algorithm to learn a model of a relationship between traffic flow at a first location and traffic flow at other locations where the sensor is located according to the time series of the received data; implementing the model to predict a time series of traffic flows at the first location from other data received from the sensors.

Description

Multi-step traffic prediction

Technical Field

The present invention relates to traffic control systems, and more particularly to a system for optimizing traffic predictions by applying statistics and machine learning to big data processing.

Background

Traffic congestion presents a serious challenge to urban infrastructure and also affects people's socioeconomic life because time is wasted while waiting for traffic.

Although there are many characteristics that describe traffic flow models, we can follow a general assumption that traffic flow data is a time series, which is a time indexed sequence of values (i.e., tuples containing various types of data: number of vehicles, speed of vehicles, etc.). Traffic data may be collected chronologically from various sources, such as cameras or street induction loop sensors. These data are continuously generated over time by sensors that measure traffic indicators. These sensors are typically spatially distributed in the urban infrastructure.

Various traffic modeling and control methods have been developed, such as macro/micro models, filtering models, and other combination models. Such studies lack practical deployment for reasons including low prediction accuracy, inability to handle correlations, strict data requirements, lack of expressiveness or understanding of internal mechanisms, and intolerable prediction time costs.

A key reason for the failure of existing traffic prediction models in practical applications is that they fail to take full advantage of spatiotemporal correlations in available traffic flow sensor sources, such as dynamic spatial factors, i.e., the topological (spatial) structure and inherent time flow of vehicles and their correlations, which can be sensed by the number of inexpensive sensors available in the urban mass. At the same time, cameras are expensive, intrusive, difficult to maintain and require algorithms to process data, which increases the cost of such schemes, making them impractical to implement on a large scale.

Typical statistical and machine learning models for time series prediction, such as autoregressive moving average series (AR, MA, ARMA, ARIMA), bayesian inference, and regression trees, can only model and predict one-dimensional time series. Such predictive models assume that the correlation in the data can be adequately described by a global time-fixed parameter. Furthermore, they cannot be extended to multivariate spatial correlation predictions per se, which makes them unsuitable for situations where the correlation between data is dynamic and heterogeneous, especially in spatially driven situations, such as road traffic data.

In the field of machine learning-based time series modeling and prediction, there are various approaches that attempt to remedy the limitations of typical statistical methods and supplement new techniques to improve performance and flexibility. The prior art in this area focuses on offline procedures that employ bayesian integration, support vector regression, nonlinear least squares, combinatorial methods and expert systems.

The methods described in the following documents assume that the diversity and accuracy of the involved models are the most important factors to consider when selecting models, and a new method of time series prediction based on neural networks and meta-features was studied: fonseca et al, "automatic model selection in time series prediction combinatorial approach", IEEE Collection, Vol.14, No. 8, pages 3811 to 3819, 2016.8 months. The method automatically adjusts the desired balance between diversity and accuracy in selecting predictors and provides good results over highly non-linear time series. However, offline unsupervised training and complex model updates make the method difficult to adapt to real-time scenarios.

Another approach described in the following document assumes that knowledge of the complexity of the time series enables the design of an adaptive predictive decision support system to actively support the accuracy of the predicted behavior and outcome: adya et al, "develop and validate rule-based time series complexity scoring techniques to support the design of adaptive predictive Decision Support Systems (DSS)", volume 83, 2016. The system is based on a rule-based complexity scoring technique that generates a time series of complexity scores using twelve rules that depend on fourteen features of the sequence. Although embedding expert system rules in modeling is an interesting approach, the decision-making system requires a large number of rules and features in selecting a model, making it difficult to meet the low resource and low time budget requirements required for real-time processing. Furthermore, the predictive Decision Support System (FDSS) focuses on complex features such as discontinuity, cardinal trend direction, horizontal discontinuity and domain knowledge, which may indeed increase the adaptability but also increase the complexity and computation time. These indices are not calculated incrementally, but are calculated off-line, so the calculation time depends on the length of the sequence, but there is no real-time prediction requirement.

In the field of traffic prediction, in order to meet the requirement of traffic jam early warning, a real-time traffic flow state identification and prediction method based on a big data driving theory is provided in many researches. The traffic big data has the characteristics of time correlation, space correlation, historical correlation, polymorphism and the like.

The method described in the following documents quantifies traffic flow conditions by using a traffic cluster model based on fuzzy c-means (SAGA-FCM) of simulated annealing genetic algorithm, which is the basis of traffic flow condition identification: huapu Lu, Zhiyuan Sun and Wencong Qu, "real-time traffic flow state identification and prediction based on big data drive", natural and social discrete dynamics, volume 2015, article number 284906, page 11, 2015. Considering simple calculation and high prediction precision, a regional traffic flow correlation analysis double-layer optimization model is established to predict traffic flow parameters based on time-space-historical correlation. Although this model has flexibility, it makes many assumptions about the spatio-temporal parameterization of the model, and only achieves a 10% gain in traffic congestion mitigation.

Another approach described in the following documents uses a Stochastic Cell Transmission Model (SCTM) for probabilistic traffic state estimation, which incorporates covariance structures calibrated from spatial correlation analysis: t.l.pan, a.sumalee, r.x.zhong and n.indrapayonong, "short-term traffic status prediction based on spatio-temporal correlation", IEEE intelligent traffic systems collection, volume 14, phase 3, pages 1242 to 1254, month 9 in 2013. Despite the large computational load behind the model, the overall Maximum Absolute Percentage Error (MAPE) for all predictions for this system is about 16.2%, which number, unfortunately, is only obtained in static empirical studies (not in real scenes).

The following studies in the literature have employed a more real-time oriented approach to investigate the potential of Probabilistic Hypothesis Density (PHD) filters in real-time traffic state estimation: m. canaud, l.michaylova, j.sau et al, (2013), "probabilistic hypothesis density filtering for real-time traffic state estimation and prediction", (NHM, network and heterogeneous media), 8(3), pages 825 to 842. The study employed a Cell Transmission Model (CTM) coupled with a PHD filter, taking into account the uncertainty of the measurement source, and showed that this can provide high accuracy in traffic setup and real-time computation cost. Although this model is very attractive, it is only used in highway environments without a large number of lanes and intersections, and the complexity is often increased in real traffic environments.

US 2002/0067292 a1 describes the use of a sensor system for environmental sensing for intelligent scene interpretation. To determine the position of the motor vehicle relative to the roadway, data from a digital road map coupled with a navigation system is fused with data provided by a distance-resolving sensor. In this context, the characteristics of the received signals of the distance-resolving sensors and the distance-related changes are evaluated to determine the distance to the road edge.

There is a need to develop a traffic prediction method that can solve these problems.

Disclosure of Invention

According to a first aspect, there is provided a road traffic flow prediction system for predicting road traffic flow from data received from a plurality of sensors, each sensor of the plurality of sensors being for acquiring road traffic data at a respective location, wherein the system is for: receiving data from the sensor; implementing a machine learning algorithm to learn a model of a relationship between traffic flow at a first location and traffic flow at other locations where the sensor is located according to the time series of the received data; implementing the model to predict a time series of traffic flows at the first location from other data received from the sensors.

At least one of the sensors may be a camera for capturing one or more images of the vehicle at the respective location. This may allow for determining the number of vehicles passing a certain location within a certain time period.

The model may comprise weights determined from the spatial relationship of the first location relative to the other locations. The weight of the respective location may decrease with the spatial distance of the location from the first location. Thus, locations that are further from the location of interest may have less impact than locations that are spatially closer.

The system may also be configured to iteratively update the weights in time increments based on the data received from the sensors. This may improve the accuracy of the learned model.

The model may be a neural network. This may be a convenient implementation.

The location may be a corresponding traffic intersection. The system may also be used to generate a time plan for a plurality of sets of traffic lights located at the intersection. Thus, the system may be used to predict and control traffic at multiple intersections.

According to a second aspect, there is provided a method for predicting road traffic flow from data received from a plurality of sensors, each sensor of the plurality of sensors for acquiring road traffic data at a respective traffic location, wherein the method comprises: receiving data from the sensor; implementing a machine learning algorithm to learn a model of a relationship between traffic flow at a first location and traffic flow at other locations where the sensor is located according to the time series of the received data; implementing the model to predict a time series of traffic flows at the first location from other data received from the sensors.

The method may further include iteratively updating the weights in time increments as a function of the data received from the sensors. This may improve the accuracy of the learned model.

The model may be a neural network. This may be a convenient implementation.

The location may be a corresponding traffic intersection. The method may also include generating a time plan for a plurality of sets of traffic lights located at the intersection. Thus, the method may be used to predict and control traffic at multiple intersections.

According to a third aspect, there is provided a computer program which, when executed by a computer, causes the computer to perform the above method. The computer program may be stored on a non-transitory computer readable storage medium.

Drawings

The invention will now be described by way of example with reference to the accompanying drawings. Wherein:

fig. 1 shows an overview of an example of processing settings for traffic prediction;

FIG. 2 illustrates an exemplary implementation of a system that receives data from cameras used to collect traffic flow time series data at spatially distant intersections in a region;

FIG. 3 illustrates another exemplary implementation of a system for receiving data from a camera for collecting traffic flow time series data at adjacent intersections;

FIG. 4 illustrates an overview of an exemplary architecture of a traffic prediction system;

FIG. 5 illustrates modules of a traffic prediction system;

FIG. 6 illustrates multi-step time learning and prediction in a temporal road traffic flow time series prediction module;

FIG. 7 illustrates spatio-temporal modeling of a time series;

8(a) to 8(c) illustrate spatial modeling of data using a traffic prediction system;

FIG. 9 illustrates nonlinear autoregressive spatio-temporal prediction in a system;

FIG. 10 illustrates the runtime functionality of the traffic prediction system;

11(a) to 11(c) show the practical use of spatiotemporal data using this system;

FIG. 12 illustrates extensibility features of the system;

fig. 13 shows an example of a method for predicting a road traffic flow.

Detailed Description

The present invention relates to a system and method for predicting road traffic flow by aggregating data from spatio-temporal correlated sensor data sources to describe traffic scenarios in urban environments. For example, from one intersection level, to a series of intersection levels, and to an area formed by a plurality of intersections. Traffic flow may be defined as the number of vehicles passing a particular point during a particular time period (e.g., number of vehicles per minute), or the number of vehicles passing through an intersection per traffic light cycle (i.e., the number of vehicles passing during a time period when the traffic light at the intersection is green). Thus, the time resolution for evaluating the number of vehicles may be a unit time period, as well as the duration of a traffic light cycle.

FIG. 1 illustrates a typical scenario (i.e., an intersection of two or more roads) depicting an urban traffic environment. At the intersection 101, given the traffic flow data 102 (i.e., the number of vehicles passing per lane) for the roads S1-S4 measured by the traffic flow sensors 103-106, the traffic prediction system may predict the future traffic flow 107 and the control time 108 of the traffic lights at the intersections of the roads S1-S4 to maximize the flow of traffic through the intersection.

Using camera data for traffic flow prediction only at each location, edge (i.e., a road connecting two intersections and intersecting another road to form an intersection), or intersection is not only costly, but also requires additional infrastructure, such as specialized image processing software. Therefore, large scale installation of such systems is not cost effective. In addition, privacy concerns are also a barrier. It is also not sufficient to use camera data of only one edge of the intersection, since all edges are correlated in space and time. However, at the city level, each edge of the intersection is connected to other edges that may affect the local road traffic flow estimation.

The system described herein utilizes local temporal correlation, global spatial correlation, and connectivity between locations, and thus can help improve local and global predictions as a whole. As an inherently complex process, road traffic flow may be modeled for prediction using a spatiotemporal model that assumes data in the form of a spatially distributed time series (i.e., a sequence of time index values). The system can receive traffic data time series, extract spatiotemporal correlations between available sensor values, and finally predict future traffic flows within any time range.

As shown in the exemplary embodiment of fig. 2, urban area 200 includes: a series of roads along which vehicles may travel in at least one direction; and intersections or crossroads between intersecting roads. A first traffic flow sensor 201 is located at a first position at a traffic intersection 202. The first sensor is used to acquire data that provides a measure of the number of vehicles at the intersection, thereby determining the flow of traffic therefrom. Preferably, the first sensor 201 is a camera for collecting images of the number of vehicles at the intersection, from which the traffic flow can be determined. Alternatively, the first traffic flow sensor may be an inductive loop. The second traffic flow sensor 204 is located at a second location of the intersection 203 that is spatially distant from the first intersection 202. A third traffic flow sensor 206 is located at a third location spatially distant from the intersection 205 of the first and second intersections.

The road traffic flow prediction system is used to predict road traffic flow based on data received from the

sensors

201, 204, 206. As will be described in greater detail below, the system implements a machine learning algorithm to learn a model of the relationship between traffic flow at a first location (intersection 202) and traffic flow at other locations (intersections 203, 205) at which the sensors are located, from a time series of data received from the sensors. The system implements the model to predict a time series of traffic flows at a first location (i.e., intersection 201) from other data received from the sensors. The system may predict a time series of traffic flows at a plurality of locations in the area. In another embodiment, the system predicts a time series of traffic flows at all locations in the area. In this example, traffic flow sensors are located at intersections of the area. However, the system may receive data from sensors located at more or fewer locations or intersections of the area. The predicted traffic flow time series of intersections may be used to control the series of traffic lights, as indicated at 207, 208, 209, located at the respective intersections to improve road traffic management at the intersections.

The model assigns weights to the data received from each sensor. The weight applied to data received from the respective sensor decreases with the distance of the location at which the respective sensor is located from the first location (the particular location at which traffic flow is predicted). For example, in fig. 2, when predicting the traffic flow at the intersection 202, data from a sensor located at the intersection 203 (spatially closer to the intersection 202 than the intersection 205) is given a higher weight than data from a sensor at the intersection 205.

Optimization of the model is continued in conjunction with historical data and incoming streams of current traffic data (e.g., number of vehicles, vehicle speed, occupancy at traffic lights, etc.).

In another example, as shown in fig. 3, the system uses traffic flow graphs recorded at intersection a (302), intersection B (303), intersection C (304), intersection D (305), and intersection O (301) within the last 5 days to generate a multi-step (future time horizon, i.e., minutes, hours, etc.) prediction of the traffic flow through intersection O shown at 301.

A multi-step prediction of traffic flow through intersection O is shown at 306. The method may be used to predict traffic flow at each or more or all of intersections A, B, C, D and O using recorded traffic flow graphs from other intersections. Intersections A, B, C and D are neighbors of intersection O. "adjacent" means an apparent connection (through an edge) between two intersections. The intersection in fig. 3 is adjacent at a spatial lag of 1 (1 jump), which means directly adjacent to the intersection O. The farther an intersection is from the reference intersection (O in this example), the higher the spatial lag.

The predicted traffic flow time series for each intersection may be used to control traffic light series to improve road traffic management at the intersections. The system is a component that interacts with the traffic control system of each intersection. Thus, the system is located at an intersection between traffic estimation and modeling and a traffic signal sequence control assembly.

Fig. 4 presents the overall architecture and data flow of the system. As shown in the flow chart at 400, the system includes several sections that cooperate to process incoming sensor time series data, as shown at 401.

The system utilizes a recursive temporal prediction module 402 that is capable of learning the non-linearity between traffic flow time series samples 401 in order to predict traffic flow at any future time horizon. The spatiotemporal correlation fusion module 403 utilizes the spatial adjacency and temporal correlation of multiple traffic flow predictions at each location to extract a weighting mechanism for the impact of the time series of a particular location on its neighbors. Given the weight of each traffic flow time series extracted from each intersection and the prediction in time range for a single intersection, the recursive spatiotemporal prediction module 404 infers a more accurate prediction of the traffic flow predictions for the various intersections.

The overall system can predict the road traffic flow for a single or multiple locations and/or intersections (i.e., in an area) by embedding spatial neighborhood data and learning spatiotemporal correlations of arbitrary prediction size.

The system allows time series predictions of highly non-stationary problems (such as traffic flow predictions) to be improved at a large range, regional and urban level, using a time-dependent learning mechanism that can capture sensor time series shape and time distribution, as well as a spatio-temporal dependent learning and weighting mechanism responsible for fusing data and a powerful multivariate reasoning mechanism based on simple operations.

The proposed system has a flexible infrastructure that can exploit road traffic sensor correlations and extract understandable spatiotemporal correlations to improve traffic predictions. The system is a lightweight learning system, and can extract basic sensor statistical data describing road traffic by using a recursive computation layer. In addition, the prediction system has a simple modular structure and interpretable output, and is suitable for double-sensor, triple-sensor and multi-sensor scenes (such as traffic flow prediction of regional or city levels). Due to the strong learning capability of the system, explicit coding of sensor associations and sensor models is not required. The system employs an efficient mechanism and has the inherent ability to infer traffic flow at any spatially connected intersection given the available traffic flow time series, thus minimizing the associated cost of equipping all intersections with cameras. From a computational perspective, the system does not require fixed or large amounts of historical data to learn and can provide arbitrary prediction time ranges due to high efficiency.

Thus, the system provides a pipelined processing mechanism that is able to predict traffic flow time series by exploiting the inherent time correlation of each individual road traffic intersection or edge, fuse the individual predictions by exploiting the spatio-temporal correlation in a systematic way, and predict all time series with the fused view angles to improve the prediction for any time horizon.

This functionality is embedded in the constituent modules of the proposed system, as shown in fig. 5. The exemplary system has three subsystems. Each element and how they positively affect the performance of the overall system in actual traffic prediction will now be described.

An input time series data set from sensors located at n intersections is shown at 500. Modeling of non-stationary and deterministic time series requires the use of specialized mechanisms to interpret their characteristics. Furthermore, processing multiple such correlated time series results in an appropriate layer to describe its covariance. The temporal road traffic flow time series prediction module 501 uses a Nonlinear Auto Regression (NAR) neural network to predict individual traffic flow time series for each location or intersection. The module uses a feed-forward neural network to regress a function of the time relationship between the sensor readings and a feedback loop for multi-step prediction. Such systems are easily trained and quickly converge to accurate predictions, even if the historical data is not large and is within an arbitrary time frame. The module uses historical data of road traffic (e.g., the number of vehicles at the intersection as determined by the camera) to predict traffic flow for a certain time horizon (e.g., minutes, hours, days, etc.) in the future. The system learns non-linear time correlations between traffic flow samples recorded at each location and uses recursion to predict traffic flow for time ranges of arbitrary size.

The core assumption is that the connection locations (e.g., intersections) in an area are spatio-temporal in nature. This correlation is considered visible in its time-of-day traffic flow pattern. To enable multi-step prediction (i.e., any future time range) from a single time series of historical data, a recurrent neural network is employed. The neural network is the Nonlinear AutoRegressive (NAR) neural network input defined by the following equation:

y(t)＝f(y(t-1),y(t-2),…,y(t-n)),y–1D (1)

wherein the next value of the output signal y (t) is regressed according to the previous value of the output signal. The model is implemented by: and (3) performing multi-step prediction on any time range by utilizing a feedforward neural network approximation function f and utilizing a recursion relation.

The respective time series data from the sensors are fed to respective time predictors which use the time correlations within each time series to output respective traffic flow predictions, as shown in fig. 6. During the training phase, the system is fed with historical data recorded over a period of time, while during the reasoning phase, the system outputs expected (arbitrary) predictions of the respective road traffic flow time series.

Thus, the module predicts individual traffic flow time series from the temporal relationships learned from historical data for individual intersections.

The spatiotemporal road traffic flow time series fusion module 502 uses a universal multivariate time series spatiotemporal learning model based on spatiotemporal autoregression (STARMA). The model is applied between the observations at each intersection and the observations at adjacent intersections, as shown in FIG. 7. The module learns a weighting scheme using spatiotemporal correlations between intersections to determine the effect of traffic flow at each intersection on traffic flow at adjacent intersections.

The system automatically extracts from the traffic flow time series the best values of past samples to be considered in the time estimation and learns parameters that weight the impact of each adjacent traffic flow time series on the local prediction.

The spatial layout of intersections and the directions in which vehicles can travel are encoded in the problem, and the system learns the contribution (i.e., weight) of each traffic flow estimator. The weighting mechanism considers the spatial proximity of intersections (large scale) and the direction of each intersection (fine scale), as shown by the spatial relationship and spatial relationship matrix in fig. 8(a) and 8(b), respectively. Fig. 8(c) shows an example of a spatial lag weight matrix. At lag1, all 32 directions (8 intersections and 4 directions) have symmetric connections with the nearest two directions, so the weight strength of each adjacent direction is 0.5. At lag 2, each direction has an asymmetric connection due to the arrangement. For each of the 32 directions, the weight is equal to 1/n, where n is the number of neighbors at the lag.

Thus, the system uses a weighting mechanism based on topological (spatial) distances to fuse temporal (time-horizon) predictions from all spatially connected locations or intersections. Locations or intersections that are spatially closer to the predicted traffic flow are given a higher weight than locations or intersections that are further away. This fusion weighting scheme is efficient, lightweight, and extensible for arbitrary intersection configurations in the area. The resulting weights are then used for global spatio-temporal prediction in an arbitrary time horizon.

The spatiotemporal road traffic flow time series prediction system is shown as 503 in fig. 5. The module performs multi-step predictive using a multivariate neural network autoregressive. The weighting scheme is combined with individual non-linear autoregressive neural network predictions in a non-linear autoregressive neural Network (NARX) with external inputs that is capable of simultaneously generating predictions of arbitrary time-horizon traffic flow predictions for the entire region. The mechanism operates similar to a non-linear autoregressive neural network by regressing a function of the time relationship between the sensor time series of one location and the other locations connected by learned weights and feedback loops for multi-step prediction. These mechanisms can be extended to large-scale hierarchies (i.e., regions in cities) through the same functional modules. The system combines the temporal (time range) prediction with spatiotemporal weights to obtain a final spatiotemporal (time range) prediction, as shown at 504.

The recurrent neural network is used in order to achieve multi-step prediction (i.e., for an arbitrary time horizon) of multiple time series from spatio-temporal correlated multivariate inputs (i.e., multiple time series from adjacent intersections).

In one example, the neural network is a non-linear autoregressive neural Network (NARX) with an external input defined by the equation:

y(t)＝f(y(t-1),y(t-2),…,y(t-n)),x(t-1),x(t-2),…,x(t-n)),y–1D,x–nD, (2)

where the next value of the output signal y (t) is regressed based on the previous value of the output signal and the previous value of the external input signal (spatially correlated time series). The NARX model is implemented by: using a feedforward neural network approximation function f, a multi-step prediction is performed using a recursive relationship (i.e., arbitrary time ranges: minutes, hours, days, etc.).

As shown in fig. 9, during the training phase, the system is fed historical data recorded over a period of time along with learned spatiotemporal fusion weights. In the inference stage, the system outputs expected (arbitrary) predicted values for each road traffic flow time series using spatiotemporal correlations between the time series.

Thus, the system uses recursion to learn a non-linear spatiotemporal correlation between each traffic flow at each intersection and each traffic flow at (spatially) adjacent intersections to predict traffic flow samples of arbitrary size time ranges. The system predicts the traffic flow time series for each location or intersection by considering the effects of other spatially connected locations and their own historical data to obtain a more accurate traffic flow prediction for each location.

All of the above components cooperate in the operation of the system and correspond to the functional blocks previously described, as shown in fig. 10. The system runtime sequence and operation of the corresponding functional modules (blocks) 1001 and 1003 will now be summarized.

In block 1001, an arbitrary time range prediction for a single time series is calculated from historical data 1000 by using a learning system that is able to extract time dependencies through a recursive relationship. In block 1002, spatiotemporal correlation between locations is exploited by learning a weighting scheme. In block 1003, time range prediction is combined with a spatio-temporal weighting scheme to make more accurate arbitrary time range prediction from historical data from multiple time series by using a learning system that can extract time correlations through a recursive relationship. Multiple time series of predictions can be performed in parallel, supporting hierarchical representations (regions).

Thus, the system provides a method of predicting road traffic flow over an arbitrarily long time horizon for a plurality of locations by exploiting spatiotemporal correlations between locations. The system is capable of implementing accurate multivariate prediction of road traffic flow (i.e., without the need to install a camera at each location) by embedding spatial knowledge in the system, learning temporal correlation in historical traffic flow time series, and learning spatial correlation between sensed road traffic at multiple locations, using efficient learning and spatio-temporal correlation according to fast and efficient calculations, and fixed time and resource budgets.

The system can be deployed for a variety of scenarios, independent of location, road geometry, size and configuration, and available sensors. The spatial layout may be directly embedded and used in the fusion module 502 to determine a weighting scheme for improved traffic flow time series predictions. Thus, the system can perform fast and efficient modeling, representation, learning, and prediction with a fixed resource budget. The control unit belongs to a resource light-weight system in the aspects of memory, size and calculation.

The methods described herein are particularly well suited for traffic optimization, which requires modeling, prediction, and rapid adaptation to sensor time series. As shown by the modular structure of the system, the system adopts efficient single time sequence prediction, and can embed and utilize space-time learning to formulate a weighting scheme fusing the contributions of all traffic flow time sequences, so that more accurate single-intersection time sequence prediction is carried out in a unified computing unit. Such a system minimizes costs by utilizing learned spatiotemporal correlations between all available traffic flow time series in a region to reduce the need to install cameras at each location or intersection. These mechanisms can be extended to large-scale hierarchies (i.e., regions in cities) through the same functional modules.

Fig. 11(a) to 11(c) show examples performed on actual data sets of a single intersection. Fig. 11(a) shows modeling of spatial proximity for time series. The spatial lag weight of the model is used to determine the direct direction with a particular correlation direction. For example, 1N is linked to 1V, 1E, 1S and 4S. Fig. 11(b) shows a single traffic flow prediction using time correlation. The system makes a multi-step prediction (i.e., one full day in the future) of a single time series (i.e., direction 1N) from a spatio-temporally correlated multivariate input (i.e., multiple time series, for

directions

1V, 1E, 1S, and 4S at spatial lags 1and 2). Employing a recurrent neural Network (NARX) such that:

Flow 1N(t)＝f(Flow1N(t-1),Flow1N(t-2),…,Flow1N(t-n),FlowNeighborsLag1and2(t-1),FlowNeighborsLag1and2(t-2),…,FlowNeighborsLag1and2(t-n)) (3)

where Flow1N is the 1D value and Flow neighbor bag 1and2 are nD vectors of neighbor values of 1N at spatial lags of 1and 2. The next value of the output signal Flow1N (t) is regressed based on the previous value of the Flow1N signal and the previous value of the Flow neighborslag1and 2. The NARX model is implemented by: and (3) performing multi-step prediction by using a feedforward neural network approximation function f and a recursion relation. Fig. 11(c) shows the spatiotemporal fusion of the improved future 24-hour predictions that predict day 55 traffic at an intersection based on five adjacent directions using two days of historical data (using data acquired from day 52 to day 54).

To emphasize the scalable nature of the system, FIG. 12 introduces the basic mechanism to achieve region or city level prediction using spatio-temporal correlation of individual intersections. The same neural network predictor can be used for arbitrary settings of the relevant direction in order to extrapolate the prediction to the area. For example, a total regional traffic flow prediction over 24 hours may be calculated using a prediction of all traffic flows for a combined direction in a region, taking into account the overall geographic setting. The traffic flow can be used to create planning and adjusting total cycles in real time control of the STARMA solution.

FIG. 13 illustrates an example of a method for predicting road traffic flow from data received from a plurality of sensors, each sensor for acquiring road traffic data at a respective traffic location. The method comprises the following steps: in step 1301, data is received from a sensor. In step 1302, the method comprises: a machine learning algorithm is implemented to learn a model of the relationship between traffic flow at a first location and traffic flow at other locations where the sensor is located, from the time series of received data. In step 1303, the method comprises: the model is implemented to predict a time series of traffic flows at the first location from other data received from the sensors.

The system described herein may include a processor and a non-volatile memory. The system may include a plurality of processors and a plurality of memories. The memory may store data executable by the processor. The processor may be configured to operate in accordance with a computer program stored in a non-transitory form on a machine-readable storage medium. The computer program may store instructions for causing the processor to perform its methods in the manner described herein. These components may be implemented in physical hardware, or may be deployed on various edge or cloud devices.

The system is able to learn cross-location correlations (i.e., pre-local decisions based on a global understanding of traffic conditions) that can support traffic congestion mitigation. The system is capable of predicting a traffic flow for any time range of the location by learning traffic flow dynamics for the location, and may also predict traffic flows for a plurality of spatially connected locations according to the learned weighting scheme. The system automatically learns the underlying mathematical relationships that are best suited to historical samples of traffic sensor data without any prior information about underlying statistics and correlations to obtain a more accurate prediction of traffic flow for each location. With such calculations, the system is able to minimize the cost of the traffic control system by inferring the traffic flow at a location from a plurality of other spatially connected locations. Thus, the need for a commonly used camera is reduced. Thus, the system can take advantage of spatio-temporal correlations inherent in the urban infrastructure to process multi-temporal sequence predictions for large-scale urban-level metropolitan groups.

The proposed system overcomes the prior art methods of complex analytical flow models based on differential equations, numerical methods and empirical methods, which are resource consuming, computationally expensive and complex, has a limited number of free parameters, relies to a large extent on extracting these parameters from the data without prior intervention by the system designer.

The output of the system can be applied to a traffic control unit, updating traffic lights to maximize traffic flow. The proposed unit can process data from multiple sensors simultaneously. The system is supported by flexible instrumentation, ensuring updates with low latency, high incoming event rates, and fixed resource budgets. Further, the system can be deployed at any type of location or intersection without prior training, regardless of intersection layout, size, and number of available sensors. This provides a significant advantage in reducing deployment costs.

The system combines the learning ability of the neural network in effectively representing the time sequence and an efficient space-time correlation learning and fusion mechanism. This combination allows real-time learning and relearning in any time frame.

The system continuously learns the temporal correlation of traffic data, exploits the spatial correlation between road traffic time series from different sensors, and adapts to changes to improve road traffic predictions. This is important in traffic flow prediction and control where the control unit must estimate and adapt to changes in the data distribution and provide accurate predictions and intelligent control actions, such as traffic light green time. Thus, the simple operations performed by the system are also advantageous because the computation requires a limited time span to process, resource allocation and execution time when the data enters the system.

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 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 invention.

Claims

1. A road traffic flow prediction system for predicting road traffic flow from data received from a plurality of sensors, each sensor of the plurality of sensors for acquiring road traffic data at a respective location, the system being operable to:

receiving data from the sensor;

implementing a machine learning algorithm to learn a model of a relationship between traffic flow at a first location and traffic flow at other locations where the sensor is located according to the time series of the received data;

implementing the model to predict a time series of traffic flows at the first location from other data received from the sensors.

2. The system of claim 1, wherein at least one of the sensors is a camera for capturing one or more images of a vehicle at the respective location.

3. The system of claim 1 or claim 2, wherein the model comprises weights determined from spatial relationships of the first location relative to the other locations.

4. The system of claim 3, wherein the weight of the respective location decreases with spatial distance of the location from the first location.

5. The system of claim 3 or claim 4, further configured to iteratively update the weights in time increments as a function of the data received from the sensor.

6. The system of the preceding claim, wherein the model is a neural network.

7. The system of any preceding claim, wherein the location is a corresponding traffic intersection.

8. The system of claim 7, further configured to generate a time plan for a plurality of sets of traffic lights located at the intersection.

9. A method for predicting roadway traffic flow from data received from a plurality of sensors, wherein each sensor of the plurality of sensors is configured to acquire roadway traffic data at a respective location, the method comprising:

receiving data from the sensor;

10. The method of claim 9, wherein at least one of the sensors is a camera for capturing one or more images of a vehicle at the respective location.

11. A method according to any one of the preceding claims 9 or 10, wherein the model comprises weights determined from the spatial relationship of the first location relative to the other locations.

12. The method of claim 11, wherein the weight of the respective location decreases with spatial distance of the location from the first location.

13. The method of claim 9 or claim 10, further comprising iteratively updating the weights in time increments as a function of the data received from the sensor.

14. The method of any one of claims 9 to 13, wherein the model is a neural network.

15. The system of any preceding claim, wherein the location is a corresponding traffic intersection.

16. The method of any one of claims 9 to 15, further comprising generating a time plan for a plurality of sets of traffic lights located at the intersection.

17. A computer program, characterized in that it causes a computer to carry out the method according to any one of claims 9 to 16 when said computer program is carried out by said computer.