Intelligent traffic light control method and system based on spatiotemporal neural architecture search

Abstract

The present invention belongs to the technical field of traffic control, and relates to an intelligent traffic light control method and system based on spatiotemporal neural architecture search, which firstly obtains traffic flow information and traffic intersection distribution space information and constructs a road condition state diagram; then constructs a temporal semantic encoder search space and a spatial semantic encoder search space to extract dynamic traffic map features; then constructs a spatiotemporal feature learning network, designs search strategies to learn different search spaces, and optimizes network structure parameters and weight parameters; finally, fine-tunes the weight parameters of the spatiotemporal feature learning network, adds a traffic light control layer, outputs traffic signal adjustment results, and improves the accuracy of traffic light control, without the need to manually design a network space architecture, with rich spatiotemporal semantics and high control accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic control, and in particular, relates to an intelligent traffic light control method and system based on spatiotemporal neural architecture search. Background Art

[0002] The construction of smart cities relies on the widespread application of artificial intelligence technology in traffic flow management and control, especially in the control of urban traffic lights. One of the core tasks of intelligent transportation systems is to optimize the control of traffic lights to reduce traffic congestion and improve road traffic efficiency. The goal of intelligent traffic light control systems is to dynamically adjust the duration and switching strategy of traffic lights based on real-time traffic flow data, thereby achieving balanced distribution of traffic flow and efficient passage.

[0003] At present, the main methods of controlling urban traffic lights are, on the one hand, based on preset timetables and fixed timing schemes, but they cannot adapt to real-time changes in traffic flow, which can easily lead to traffic congestion during peak hours or waste of road resources during non-peak hours; on the other hand, they use sensing equipment installed at intersections (such as vehicle detectors, infrared sensors, etc.) to sense traffic flow in real time and adjust the duration and switching strategy of traffic lights accordingly. However, the deployment and maintenance costs of sensor equipment are high, and they rely on local data from a single detection point, making it impossible to globally optimize traffic flow. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides an intelligent traffic light control method and system based on spatiotemporal neural architecture search, constructs time-adaptive and space-adaptive search spaces based on the spatiotemporal neural architecture search algorithm, fully considers the correlation of time and space dimensions, improves the accuracy of traffic light control, realizes more efficient traffic management, reduces the risk of traffic congestion, and improves the overall operation efficiency of the urban transportation system.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides an intelligent traffic light control method based on spatiotemporal neural architecture search, comprising the following steps:

[0007] S1. Obtaining traffic flow information and traffic intersection distribution spatial information in a preset traffic intersection area over a period of time, and constructing a basic traffic state diagram, a time sequence traffic state diagram, and a spatial traffic state diagram based on the acquired information;

[0008] S2, based on spatiotemporal neural architecture search, construct temporal semantic encoder search space and spatial semantic encoder search space to extract dynamic traffic map features;

[0009] S3. Construct a spatiotemporal feature learning network, design search strategies to learn different search spaces, and use differentiable methods to optimize the network structure parameters and weight parameters of the temporal semantic encoder search space and the spatial semantic encoder search space;

[0010] S4. Fine-tune the weight parameters of the spatiotemporal feature learning network, add a traffic light control layer, and output the traffic signal adjustment results.

[0011] As a further technical solution of the present invention, the traffic flow information in step S1 is obtained from traffic sensors or monitoring equipment, including pedestrian data, bicycle data and motor vehicle data.

[0012] As a further technical solution of the present invention, step S1 constructs a basic road condition state diagram according to static traffic intersection distribution information, constructs a time sequence road condition state diagram according to dynamic time segment association information, and constructs a spatial road condition state diagram according to dynamic traffic flow change information. The constructed basic road condition state diagram, time sequence road condition state diagram, and spatial road condition state diagram are: ,in Represents the traffic network composed of preset traffic intersections nodes, Represents the adjacency relationship between nodes in the graph. Represents a node in a transportation network dimensional feature representation, similarly represents the edge The dimension feature indicates that if any two nodes in the traffic network and There is no adjacency relationship between , then in The value between the two nodes is set to 0, indicating that there is no adjacency between the two nodes. To represent the set of adjacent nodes of any node, when two nodes , If there is an adjacency relationship between them, add them to the set , the three state diagrams are represented as , and .

[0013] As a further technical solution of the present invention, the temporal semantic encoder search space in step S2 includes a long short-term memory network, a gated neural unit and a Transformer network, and the spatial semantic encoder search space includes a graph convolutional neural network, a graph attention neural network and a convolutional neural network. Based on the temporal semantic encoder search space and the spatial semantic encoder search space, a candidate operation set is constructed. , and perform weighted fusion based on the basic traffic state diagram, the temporal traffic state diagram, and the spatial traffic state diagram to obtain advanced dynamic traffic map features: ,in is the dynamic weight, As the bias, the dynamic traffic map features are used Normalize using the softmax algorithm and get:

[0014] ,

[0015] In the formula For Node and nodes In the Correlation on dimensional features, Representative Node and nodes The weight of the edge between the candidate operation sets is a predefined temporal semantic encoder and a spatial semantic encoder. Then the normalized dynamic traffic feature map is and traffic flow information Input into the spatiotemporal search space to obtain the hidden state , calculated as follows: .

[0016] As a further technical solution of the present invention, the process of step S3 is: constructing the entire spatiotemporal feature learning network as a directed acyclic graph, the operation between each node in the spatiotemporal feature learning network constructed in the initial stage is uncertain, and a mixed candidate operation set is defined on each edge to continuously relax the time or space semantic encoder search space, and by solving the two-level optimization problem, the directed acyclic graph network weights and the mixing probability are jointly optimized to obtain the optimal network model architecture through the learned mixing probability, wherein the two-level optimization problem is expressed as:

[0017] ,

[0018] in The architecture parameters of the network are learned for spatiotemporal features, Learning the model parameters of the network for spatiotemporal features, and represents the validation loss and training loss function; the high-level traffic features calculated are : , is an edge in a directed acyclic graph<i,j> The weight of operation o.

[0019] As a further technical solution of the present invention, the process of step S4 is: the intersection is modeled by traffic simulation software SUM0, the road intersection in the intersection is composed of four roads: east, south, west and north, and a traffic light is set at the intersection. The intersection is a two-way lane. Along the driving direction of the vehicle, the left lane is a left turn lane, the middle lane is a straight lane, and the right lane is a straight and right turn lane; the control signals of the traffic lights are: straight in the north-south direction, left turn in the north-south direction, straight in the east-west direction, and left turn in the east-west direction. A yellow signal light is designed for transition between the four phase switches to ensure that the vehicle passes through the intersection safely, and different traffic intersection semantic information is expressed as features. : ,in The spatial language features of all traffic signal controls, Represents the linear layer operation, and uses the cross attention mechanism to fuse features with high-level traffic features: ,in represent Characteristics change, and It represents Feature transformation, Representative Matrix The degree, represents the attention mechanism layer, express Activation function, after feature fusion, the flow information of the preset intersection and the semantic information of the traffic light are fully combined, and the control information of the traffic light is output through a multi-layer perceptron.

[0020] In a second aspect, the present invention provides an intelligent traffic light control system based on spatiotemporal neural architecture search, comprising:

[0021] A data acquisition module is used to acquire the traffic flow data and spatial distribution of a preset traffic intersection over a period of time, and construct a basic road condition state diagram, a time sequence road condition state diagram and a spatial road condition state diagram;

[0022] A spatiotemporal search space construction module, used to construct a temporal semantic encoder search space and a spatial semantic encoder search space;

[0023] The spatiotemporal search space optimization module is used to design search strategies to learn different search spaces, optimize the network structure parameters and weight parameters of the temporal semantic encoder search space and the spatial semantic encoder search space using differentiable methods, and extract high-level features;

[0024] The traffic light control module is used to fuse the obtained high-level features and spatial semantic embeddings to generate and output the traffic signal status.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] According to the traffic flow characteristics and preset traffic intersection information, the time detector search space and spatial semantic encoder search space are constructed based on the idea of ​​spatiotemporal neural architecture search, and a gradient-based differentiable search algorithm is used to automatically design the spatiotemporal network architecture. Combined with the traffic light control signal layer, the intelligent traffic light control model has rich spatiotemporal data characteristics, thereby improving the accuracy of traffic light control. There is no need to manually design the network space architecture, with rich spatiotemporal semantics and high control accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0028] Figure 1 Flowchart of the intelligent traffic light control method based on spatiotemporal neural architecture search provided by the present invention.

[0029] Figure 2 Flowchart of the neural architecture search algorithm provided by the present invention.

[0030] Figure 3 This is a structural block diagram of the intelligent traffic light control system based on spatiotemporal neural architecture search provided by the present invention. DETAILED DESCRIPTION

[0031] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0032] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0033] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0034] Introduction to Neural Architecture Search Technology: Neural architecture search is a method for automatically designing neural network architectures. Manually designing neural network architectures relies on the experience and intuition of experts, and gradually improving the network architecture through repeated trials requires high trial and error costs. The neural architecture search algorithm aims to automatically find the optimal network architecture to achieve the best performance on a specific task. It is mainly divided into evolutionary algorithms, reinforcement learning, and gradient-based methods. Among them, the gradient-based method optimizes differentiable architecture parameters to perform architecture search. This method significantly reduces search time and computational costs, making large-scale neural architecture search possible.

[0035] Embodiment 1:

[0036] Urban traffic light control aims to optimize the configuration of traffic lights according to the observed real-time traffic flow (vehicle flow, pedestrian flow, bicycle flow), adaptively adjust the state of traffic lights, and improve the efficiency and safety of the traffic system. This embodiment discloses an intelligent traffic light control method based on spatiotemporal neural architecture search, which uses neural architecture search technology to automatically find the best network architecture for the adaptation task, designs the temporal semantic encoder search space and the spatial semantic encoder search space according to the spatiotemporal characteristics of traffic flow data and traffic intersections, and uses a differentiable search algorithm to calculate the best structural parameters and network in different spaces. Network parameters. Specifically, traffic flow data has spatiotemporal characteristics. From a spatial perspective, in an urban area, different traffic intersections have strong correlations. They are geographically distributed differently but are related to each other. From a temporal perspective, traffic flow data has periodicity and proximity. Based on the traffic flow information at the past and current moments, the traffic flow information of the preset intersections at future moments is accurately captured, thereby effectively assisting the control of traffic lights. After obtaining the weight parameters and network parameters, the model is fine-tuned in combination with the semantic information of the traffic intersection, and a traffic light control layer is added to finally output the traffic signal value information of the predicted intersection, such as Figure 1 As shown, the method specifically comprises the following steps:

[0037] S1. Data acquisition:

[0038] Obtain traffic flow information for a period of time in a preset traffic intersection area and traffic intersection distribution space information, and construct a basic road condition state diagram, a time sequence road condition state diagram, and a spatial road condition state diagram based on the above information, wherein the traffic flow information is obtained from traffic sensors or monitoring equipment, including pedestrian data, bicycle data, and motor vehicle data, and pre-processes the data. Specifically, according to the preset traffic intersection connection information, the basic road condition state diagram, the time sequence road condition state diagram, and the spatial road condition state diagram are constructed as follows: ,in Represents the traffic network composed of preset traffic intersections nodes, Represents the adjacency relationship between nodes in the graph. Represents a node in a transportation network dimensional feature representation, similarly represents the edge The dimension feature indicates that if any two nodes in the traffic network and There is no adjacency relationship between , then in The value between the two nodes is set to 0, indicating that there is no adjacency between the two nodes. To represent the set of adjacent nodes of any node, when two nodes , If there is an adjacency relationship between them, add them to the set In this embodiment, the three state diagrams are respectively represented as , and .

[0039] Based on the above description, this embodiment is based on three state diagrams. , , and historical traffic flow data through preset traffic intersections , generate traffic signal control for preset traffic intersections .

[0040] S2. Construction of spatiotemporal search space:

[0041] Design the temporal semantic encoder search space and the spatial semantic encoder search space to extract high-level features from the input information. The basic road condition state diagram, the time series road condition state diagram, the spatial road condition state diagram and the traffic flow data have strong spatiotemporal correlation, so extracting high-level spatiotemporal features requires sophisticated design. This embodiment uses a spatiotemporal neural architecture search algorithm to automatically generate the best network structure, such as Figure 2 As shown, the neural architecture algorithm is mainly composed of a search space, a search strategy, and a performance evaluation strategy. Specifically, this embodiment designs a temporal semantic encoder search space and a spatial semantic encoder search space, wherein the temporal semantic encoder search space includes a long short-term memory network, a gated neural unit, and a Transformer, and the spatial semantic encoder search space includes a graph convolutional neural network, a graph attention neural network, and a convolutional neural network. Based on the above encoders, this embodiment constructs a candidate operation set .

[0042] In this embodiment, the basic traffic state diagram, the time sequence traffic state diagram and the spatial traffic state diagram are weightedly fused to obtain advanced dynamic traffic diagram features: ,in is the dynamic weight, As the bias, the dynamic traffic map features are used Normalize using the softmax algorithm and get:

[0043] ,

[0044] In the formula For Node and nodes In the Correlation on dimensional features, Representative Node and nodes The weight of the edge between the candidate operation sets is a predefined temporal semantic encoder and a spatial semantic encoder. Then the normalized dynamic traffic feature map is and traffic flow information Input into the spatiotemporal search space to obtain the hidden state , calculated as follows: .

[0045] S3, spatiotemporal search space optimization:

[0046] Design as Figure 2 The search strategy shown learns different search spaces, and uses a differentiable method to optimize the network structure parameters and weight parameters. This embodiment uses a differentiable method to optimize the search space. Specifically, the entire spatiotemporal feature learning network is constructed as a directed acyclic graph. The operations between each node in the spatiotemporal feature learning network constructed in the initial stage are uncertain. A mixed candidate operation set is defined on each edge to continuously relax the time or space semantic encoder search space. By solving the two-level optimization problem, the directed acyclic graph network weights and the mixing probability are jointly optimized to obtain the optimal network model architecture through the learned mixing probability, where the two-level optimization problem is expressed as:

[0047] ,

[0048] in The architecture parameters of the network are learned for spatiotemporal features, Learning the model parameters of the network for spatiotemporal features, and represents the validation loss and training loss function; the high-level traffic features calculated are : , is an edge in a directed acyclic graph<i,j> The weight of operation o.

[0049] S4, Traffic light control:

[0050] The weight parameters of the model are fine-tuned, and a traffic light control layer is added to output the traffic signal adjustment result. In this embodiment, the advanced flow features are embedded and integrated with the traffic intersection features, and finally the control information of the traffic light is output through a multi-layer perceptron. Specifically, the intersection is modeled by the traffic simulation software SUM0. The road intersection in the intersection consists of four roads: east, south, west, and north. A traffic light is set at the intersection. In this embodiment, the intersection is a two-way 6-lane road. Along the driving direction of the vehicle, the left lane is the left turn lane, the middle lane is the straight lane, and the right lane is the straight and right turn lane; the control signals of the traffic lights are: north-south straight, north-south left turn, east-west straight, and east-west left turn. A yellow signal light is designed for transition between the four phase switches to ensure that the vehicle passes through the intersection safely. Based on the above description, different traffic intersection semantic information is expressed as features. , calculated as follows: ,in The spatial language features of all traffic signal controls, Represents the linear layer operation, and uses the cross attention mechanism to fuse features with high-level traffic features: ,in represent Characteristics change, and It represents Feature transformation, Representative Matrix The degree, represents the attention mechanism layer, express Activation function, after feature fusion, the flow information of the preset intersection and the semantic information of the traffic light are fully combined, and the control information of the traffic light is output through a multi-layer perceptron.

[0051] Based on the above steps, a model for intelligent traffic light control can be obtained, which can effectively improve control accuracy and alleviate traffic pressure.

[0052] Embodiment 2:

[0053] like Figure 3 As shown, this embodiment provides an intelligent traffic light control system based on spatiotemporal neural architecture search, including:

[0054] A data acquisition module is used to acquire the traffic flow data and spatial distribution of a preset traffic intersection over a period of time, and construct a basic road condition state diagram, a time sequence road condition state diagram and a spatial road condition state diagram;

[0055] A spatiotemporal search space construction module, used to construct a temporal semantic encoder search space and a spatial semantic encoder search space;

[0056] The spatiotemporal search space optimization module is used to design search strategies to learn different search spaces, optimize the network structure parameters and weight parameters of the temporal semantic encoder search space and the spatial semantic encoder search space using differentiable methods, and extract high-level features;

[0057] The traffic light control module is used to fuse the obtained high-level features and spatial semantic embeddings to generate and output the traffic signal status.

[0058] In more embodiments, there is also provided: an electronic device including a memory and a processor, and equipped with computer instructions running on the processor. When these instructions are executed by the processor, the method steps described in Example 1 can be accurately implemented. For the sake of simplicity, the detailed method content will not be repeated here.

[0059] It should be clear that the processor in this embodiment is not limited to the central processing unit CPU, it can also be a general-purpose processor core, a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic device PLD, a discrete gate circuit or transistor logic component, a discrete hardware module, etc. In addition, the general-purpose processor can be a microprocessor MPU, or any conventional computing core. The memory can include a read-only memory and a random access memory, and provide instructions and data to the processor. A part of the memory can also include a non-volatile random access memory. For example, the memory can also store information about the type of device.

[0060] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is completed.

[0061] The method in Example 1 can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it is not described in detail here.

[0062] The algorithms, systems, and network structures not described in detail in the present invention are all general technologies in the art.

[0063] Those skilled in the art will appreciate that the units, i.e., algorithm steps, of the various examples described in conjunction with this embodiment can be implemented in electronic hardware or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

Claims

1. An intelligent traffic light control method based on spatiotemporal neural architecture search, characterized in that: The following steps are involved: S1. Obtain traffic flow information and traffic intersection distribution spatial information in a preset traffic intersection area within a period of time, and construct a basic road condition state diagram, a time-series road condition state diagram, and a spatial road condition state diagram based on the obtained information; wherein the traffic flow information is obtained from traffic sensors or monitoring equipment, including pedestrian data, bicycle data, and motor vehicle data, and a basic road condition state diagram is constructed based on static traffic intersection distribution information, a time-series road condition state diagram is constructed based on dynamic time segment association information, and a spatial road condition state diagram is constructed based on dynamic traffic flow change information. The constructed basic road condition state diagram, time-series road condition state diagram, and spatial road condition state diagram are: ,in Represents the traffic network composed of preset traffic intersections nodes, Represents the adjacency relationship between nodes in the graph. Represents a node in a transportation network dimensional feature representation, similarly represents the edge The dimension feature indicates that if any two nodes in the traffic network and There is no adjacency relationship between , then in The value between the two nodes is set to 0, indicating that there is no adjacency between the two nodes. To represent the adjacent nodes of any node When two nodes , If there is an adjacency relationship between them, add them to the set , the three state diagrams are represented as , and ; S2. Based on the spatiotemporal neural architecture search, a temporal semantic encoder search space and a spatial semantic encoder search space are constructed to extract dynamic traffic map features; the temporal semantic encoder search space includes a long short-term memory network, a gated neural unit, and a Transformer network, and the spatial semantic encoder search space includes a graph convolutional neural network, a graph attention neural network, and a convolutional neural network. Based on the temporal semantic encoder search space and the spatial semantic encoder search space, a candidate operation set is constructed. , and perform weighted fusion based on the basic traffic state diagram, the temporal traffic state diagram, and the spatial traffic state diagram to obtain advanced dynamic traffic map features: ,in is the dynamic weight, As the bias, the dynamic traffic map features Using the softmax algorithm for standardization, we get: ， In the formula For Node and nodes In the Correlation on dimensional features, Representative Node and nodes The weight of the edge between them, each candidate operation set is a predefined type of time semantic encoder and space semantic encoder, and then the normalized dynamic traffic feature map and traffic flow information Input into the spatiotemporal search space to obtain the hidden state , calculated as follows: = ; S3. Construct a spatiotemporal feature learning network, design search strategies to learn different search spaces, and use differentiable methods to optimize the network structure parameters and weight parameters of the temporal semantic encoder search space and the spatial semantic encoder search space; the specific process is: construct the entire spatiotemporal feature learning network as a directed acyclic graph. The operations between each node in the spatiotemporal feature learning network constructed in the initial stage are uncertain. A mixed set of candidate operations is defined on each edge to continuously relax the temporal or spatial semantic encoder search space. By solving the two-level optimization problem, the directed acyclic graph network weights and mixing probabilities are jointly optimized to obtain the optimal network model architecture through the learned mixing probability. The two-level optimization problem is expressed as: ， in The architecture parameters of the network are learned for spatiotemporal features, Learning the model parameters of the network for spatiotemporal features, and represents the validation loss and training loss function; the high-level traffic features calculated are : , is an edge in a directed acyclic graph<i,j> The weight of operation o; S4. Fine-tune the weight parameters of the spatiotemporal feature learning network, add a traffic light control layer, and output the traffic signal adjustment results. The specific process is: model the intersection through the traffic simulation software SUM0. The road intersection in the intersection consists of four roads: east, south, west, and north. There is a traffic light at the intersection. The intersection is a two-way lane. Along the direction of vehicle travel, the left lane is the left turn lane, the middle lane is the straight lane, and the right lane is the straight and right turn lane; the control signals of the traffic lights are: straight in the north-south direction, left turn in the north-south direction, straight in the east-west direction, and left turn in the east-west direction. A yellow signal light is designed for transition between the four phase switches to ensure that vehicles pass through the intersection safely. Different traffic intersection semantic information is expressed as features : = ,in The spatial language features of all traffic signal controls, Represents the linear layer operation, and uses the cross attention mechanism to fuse features with high-level traffic features: ,in represent Characteristics change, and It represents Feature transformation, represent The degree of the matrix, represents the attention mechanism layer, express Activation function, after feature fusion, the flow information of the preset intersection and the semantic information of the traffic light are fully combined, and the control information of the traffic light is output through a multi-layer perceptron.

2. An intelligent traffic light control system based on spatiotemporal neural architecture search, characterized in that: The method according to claim 1 is performed, comprising: A data acquisition module is used to acquire the traffic flow data and spatial distribution of a preset traffic intersection over a period of time, and construct a basic road condition state diagram, a time sequence road condition state diagram and a spatial road condition state diagram; A spatiotemporal search space construction module, used to construct a temporal semantic encoder search space and a spatial semantic encoder search space; The spatiotemporal search space optimization module is used to design search strategies to learn different search spaces, optimize the network structure parameters and weight parameters of the temporal semantic encoder search space and the spatial semantic encoder search space using differentiable methods, and extract high-level features; The traffic light control module is used to fuse the obtained high-level features and spatial semantic embeddings to generate and output the traffic signal status.

Images