CN114548298A

CN114548298A - Model training method, traffic information processing method, device, equipment and storage medium

Info

Publication number: CN114548298A
Application number: CN202210178922.6A
Authority: CN
Inventors: 李福樑
Original assignee: Apollo Zhilian Beijing Technology Co Ltd; Apollo Zhixing Technology Guangzhou Co Ltd
Current assignee: Apollo Zhilian Beijing Technology Co Ltd; Apollo Zhixing Technology Guangzhou Co Ltd
Priority date: 2022-02-25
Filing date: 2022-02-25
Publication date: 2022-05-27
Anticipated expiration: 2042-02-25
Also published as: CN114548298B

Abstract

The disclosure provides a model training method, a traffic information processing device, a traffic information processing equipment and a storage medium, and relates to the technical field of artificial intelligence such as intelligent traffic and deep learning. The specific implementation scheme is as follows: acquiring a collected sample collected at a road intersection; the collected samples comprise phase information of traffic lights at the road intersections, traffic flow information at the intersections and facility information of the roads at the intersections; converting the collected sample into collected sample input data according to a preset information data conversion mode; expanding the collected sample input data to generate training sample input data; and training a prediction model to be trained according to the input data of the training sample to obtain the trained prediction model, wherein the prediction model is used for outputting signal lamp phase recommendation information according to the intersection traffic information and the intersection road facility information. The prediction model generated by the embodiment of the disclosure can provide efficient and accurate signal lamp phase recommendation information for road intersections.

Description

Model training method, traffic information processing method, device, equipment and storage medium

Technical Field

The present disclosure relates to the field of computer technology, and more particularly to the field of artificial intelligence technologies such as intelligent transportation and deep learning.

Background

Traffic signal control is an important means for urban road intersection management and control, and plays a key role in maintaining road traffic order, guaranteeing traffic safety and improving traffic operation efficiency. The design of a signal phase scheme of a traffic signal lamp is an important content of the establishment work of an intersection signal control scheme. The phase setting scheme is a structural framework of the intersection signal control scheme, and not only determines the running order of traffic flow in the intersection, but also plays a decisive role in the traffic efficiency of the intersection.

Disclosure of Invention

The disclosure provides a model training method, a traffic information processing method, a model training device, a traffic information processing equipment and a storage medium.

According to an aspect of the present disclosure, there is provided a model training method, including:

acquiring a collected sample collected at a road intersection; the collected samples comprise phase information of traffic lights of the road intersection, traffic flow information of the road intersection and facility information of the road intersection;

converting the collected sample into collected sample input data according to a preset information data conversion mode;

expanding the collected sample input data to generate training sample input data;

and training a prediction model to be trained according to the training sample input data to obtain the trained prediction model, wherein the prediction model is used for outputting signal lamp phase recommendation information according to intersection traffic flow information and intersection road facility information.

According to another aspect of the present disclosure, there is provided a traffic information processing method including:

acquiring intersection traffic flow information and intersection road facility information of a road intersection;

and inputting the intersection traffic flow information and the intersection road facility information into a prediction model to obtain phase recommendation information, wherein the prediction model is a trained prediction model provided by any one embodiment of the disclosure.

According to another aspect of the present disclosure, there is provided a model training apparatus including:

the acquisition sample module is used for acquiring an acquisition sample acquired at a road intersection; the collected samples comprise phase information of traffic lights of the road intersection, traffic flow information of the road intersection and facility information of the road intersection;

the conversion module is used for converting the collected sample into collected sample input data according to a preset information data conversion mode;

the training sample input data module is used for expanding the acquired sample input data to generate training sample input data;

and the training module is used for training a prediction model to be trained according to the training sample input data to obtain the trained prediction model, and the prediction model is used for outputting signal lamp phase recommendation information according to intersection traffic flow information and intersection road facility information.

According to another aspect of the present disclosure, there is provided a traffic information processing apparatus including:

the information acquisition module is used for acquiring intersection traffic flow information and intersection road facility information of the road intersection;

and the phase prediction module is used for inputting the intersection traffic flow information and the intersection road facility information into a prediction model to obtain phase recommendation information, wherein the prediction model is a trained prediction model provided by any one embodiment of the disclosure.

According to another aspect of the present disclosure, there is provided an electronic device including:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a method according to any one of the embodiments of the present disclosure.

According to another aspect of the present disclosure, there is provided a non-transitory computer readable storage medium having stored thereon computer instructions for causing a computer to perform a method in any of the embodiments of the present disclosure.

According to another aspect of the present disclosure, there is provided a computer program product comprising computer programs/instructions which, when executed by a processor, implement the method in any of the embodiments of the present disclosure.

According to the technology disclosed by the invention, the training of the prediction model can be carried out according to the actually collected samples of the intersection with the road, so that the phase recommendation information of the traffic signal lamp can be generated, and a more reasonable phase scheme is provided.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.

Drawings

The drawings are included to provide a better understanding of the present solution and are not to be construed as limiting the present disclosure. Wherein:

FIG. 1 is a schematic diagram of a model training method according to an embodiment of the present disclosure;

FIG. 2 is a schematic diagram of a model training method according to another embodiment of the present disclosure;

FIG. 3 is a schematic diagram of a model training method according to yet another embodiment of the present disclosure;

FIG. 4 is a schematic diagram of a model training method according to yet another embodiment of the present disclosure;

FIG. 5 is a schematic diagram of a model training method according to yet another embodiment of the present disclosure;

FIG. 6 is a schematic diagram of a model training method according to yet another embodiment of the present disclosure;

FIG. 7 is a schematic diagram of a model training method according to yet another embodiment of the present disclosure;

FIG. 8 is a schematic diagram of a model training method according to yet another embodiment of the present disclosure;

FIG. 9 is a schematic diagram of a traffic information processing method according to an embodiment of the present disclosure;

FIG. 10 is a schematic diagram of a model training method according to yet another embodiment of the present disclosure;

FIG. 11 is a schematic diagram of a model training method according to an example of the present disclosure;

FIG. 12 is a schematic illustration of an intersection according to an example of the present disclosure;

FIG. 13 is a schematic illustration of an intersection according to another example of the present disclosure;

FIG. 14 is a schematic diagram of a model training method according to another example of the present disclosure;

FIG. 15 is a schematic illustration of an intersection road numbering plan according to an example of the present disclosure;

FIG. 16 is a schematic illustration of intersection road numbering according to another example of the present disclosure;

FIG. 17 is a schematic diagram of a model training method according to yet another example of the present disclosure;

FIG. 18 is a schematic diagram of a generative model structure according to an example of the present disclosure;

FIG. 19 is a schematic view of a discriminant model structure according to an example of the present disclosure;

FIG. 20 is a schematic diagram of a decision tree generation flow according to another example of the present disclosure;

FIG. 21 is a schematic illustration of a pruning process according to another example of the present disclosure;

FIG. 22 is a schematic diagram of a model training and prediction process according to another example of the present disclosure;

FIG. 23 is a schematic illustration of a process for expanding capacity according to another example of the present disclosure;

FIG. 24 is a schematic diagram of a model training apparatus according to an embodiment of the present disclosure;

FIG. 25 is a schematic diagram of a model training apparatus according to another embodiment of the present disclosure;

FIG. 26 is a schematic view of a model training apparatus according to yet another embodiment of the present disclosure;

FIG. 27 is a schematic view of a model training apparatus according to yet another embodiment of the present disclosure;

FIG. 28 is a schematic view of a model training apparatus according to yet another embodiment of the present disclosure;

FIG. 29 is a schematic view of a model training apparatus according to yet another embodiment of the present disclosure;

FIG. 30 is a schematic view of a traffic information processing apparatus according to yet another embodiment of the present disclosure;

FIG. 31 is a schematic diagram of a post-pruning decision tree according to an example of the present disclosure;

FIG. 32 is a block diagram of an electronic device for implementing a model training method of an embodiment of the present disclosure.

Detailed Description

Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, in which various details of the embodiments of the disclosure are included to assist understanding, and which are to be considered as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

The embodiment of the present disclosure first provides a model training method, as shown in fig. 1, including:

step S11: acquiring a collected sample collected at a road intersection; the collected samples comprise phase information of traffic lights at the road intersections, traffic flow information at the intersections and facility information of the roads at the intersections;

step S12: converting the collected sample into collected sample input data according to a preset information data conversion mode;

step S13: expanding the collected sample input data to generate training sample input data;

step S14: and training a prediction model to be trained according to the input data of the training sample to obtain the trained prediction model, wherein the prediction model is used for outputting signal lamp phase recommendation information according to the intersection traffic information and the intersection road facility information.

In this embodiment, the collected samples collected at the intersections may be collected information at intersections of actual urban roads to obtain collected samples, or collected information at any intersection having a traffic light to obtain collected samples.

In this embodiment, a real intersection may correspond to one collection sample. Alternatively, a real intersection may correspond to more than one collected sample.

In this embodiment, the intersection road may include an intersection where two or more roads intersect, such as a crossroad, a herringbone intersection, and a T-intersection (T-shaped intersection). In this embodiment, for an intersection, one collected sample may be obtained for each straight entrance lane. For example, as shown in fig. 12, the intersection includes a north-south entrance lane and an east-west entrance lane, and a single collected sample can be obtained for each of the north-south entrance lane and the east-west entrance lane.

For the T-shaped intersection shown in fig. 13, one collected sample can be obtained for each of the three inlet channels, or the collected samples of the three inlet channels can be integrated into one collected sample.

The intersection road facility information may include any facility information related to road traffic provided at the intersection. Such as road channelizing information, traffic light facility information, etc. The canalized traffic is to draw lines on the road or to separate lanes by green zones according to traffic volume, so that vehicles with different properties and different speeds can run along the specified direction without mutual interference like water flow in the canal. The road channelized information can be information such as traffic signs, marking lines and/or traffic islands arranged at plane crossings, and can guide traffic flow and pedestrians to respectively travel their roads.

In this embodiment, the collected sample may be originally collected data, including video, website data, document data, image data, and the like. Or the data obtained by converting the originally collected data, such as text information or other types of characters obtained by converting videos, website data, document data, image data, and the like. For example, a collection sample table can be generated by summarizing images, videos and other data of a road intersection, wherein the table includes traffic signal lamp phase information, intersection traffic flow information, intersection road facility information, corresponding sub-categories and specific contents corresponding to the sub-categories.

In this embodiment, the collected sample is converted into the collected sample input data according to a preset information data conversion mode, and may be converted into data that can be identified by the prediction model according to a set conversion mode.

The method comprises the steps of expanding the collected sample input data to generate training sample input data, wherein the data volume of the collected sample input data can be expanded, the number of data is increased, and more training sample input data are constructed according to the limited collected sample input data.

The method for training the prediction model to be trained according to the training sample input data to obtain the trained prediction model can comprise the steps of constructing the prediction model to be trained according to the training sample input data, and then training the prediction model to be trained by utilizing the training sample input data to obtain the trained prediction model.

In the embodiment, the training of the prediction model can be performed according to the collected samples of the road intersections collected actually, so that the phase recommendation information of the traffic signal lamps can be generated, and a more reasonable phase scheme is provided.

In one embodiment, as shown in fig. 2, the preset information data conversion manner includes a preset numbering manner and a preset calculation manner; according to the predetermined information data conversion mode, will gather the sample and convert into and gather sample input data, include:

step S21: converting the phase information of the traffic signal lamps and the information of the road facilities at the intersections into serial number data according to a preset serial number mode;

step S22: calculating traffic flow index data according to a preset calculation mode and traffic flow information at the intersection;

step S23: and acquiring the input data of the collected samples according to the serial number data and the traffic flow index data.

In this embodiment, the converting the phase information of the traffic signal lamp and the information of the intersection road facility into the number data according to the preset number mode may include converting the phase information of the traffic signal lamp into the number data according to the preset number mode; and converting the intersection road facility signal information into serial number data according to a preset serial number mode. The method also comprises the step of converting the relevant combination of the traffic signal lamp phase information and the intersection road facility information into number data according to a preset number mode.

The number data can be integer data, decimal data and special symbol data.

In a specific implementation manner, step S21 and step S22 may be executed in any order, or may be executed simultaneously.

In this embodiment, the obtaining of the collected sample input data according to the serial number data and the traffic flow index data may include taking the serial number data and the traffic flow index data as the collected sample input data.

In actual operation, each actual intersection can acquire relevant traffic light phase information, traffic flow information and intersection specific road facility information. Thus, the collected sample input data may include a plurality of sets, and each set of collected sample input data may correspond to a specific intersection, i.e., to a combination of a set of traffic stream indicator data, intersection asset information number data, and traffic light phase information number data.

In this embodiment, the collected samples are converted into collected sample input data, so that the collected samples of the non-data type can be converted into input data of the data type, and training of the prediction model to be trained is realized.

In one embodiment, calculating the traffic flow index data according to the intersection traffic flow information in a preset calculation mode includes:

determining intersection type information corresponding to the intersection traffic flow information;

and calculating the traffic flow index data according to the intersection type information and the intersection road and intersection traffic flow information corresponding to the intersection traffic flow information.

In this embodiment, the intersection type information may include a specific type of the intersection, for example, in actual road traffic, an intersection, a t-junction, and the like may generally exist. For further specific distinction, the crossroad, the T-intersection and the like can be divided into more subcategories according to angles.

In other implementations, the T-intersection may be further divided, for example, into Y-intersections or T-intersections.

In another possible implementation, the crossroad or the t-intersection may also be classified as to whether it extends in one of the north, east, west and north directions. For example, intersections that are north, south, west, and north may be classified into a first category, and intersections that are not north, south, west, and north may be classified into a second category. The direction division can be carried out according to the angle of the road for the T-junction. For example, a T-junction in which three roads are exactly adjacent to each other and the included angle is about 120 degrees can be classified into a third category, two of the three roads form a straight line, and the other two roads are perpendicular to each other, so that the T-junction formed by the three roads is classified into a fourth category.

In this embodiment, the intersection roads corresponding to the intersection traffic information may include categories into which the individual intersection access channels are divided according to the direction information of the intersection roads. For example, for a crossroad, a first number may be assigned to the north-south entry and another number may be assigned to the east-west entry.

In this embodiment, the traffic data is converted into the traffic index, so as to perform unified expression on the traffic flow value, and perform dimension unification with other types of collected sample information.

In one embodiment, the intersection asset information comprises: the system comprises intersection type information, left-turn lane widening information, right-turn lane channelizing information, left-turn waiting area information, pedestrian street crossing facility information, left-turn motor vehicle signal lamp group type, right-turn motor vehicle signal lamp group type information and non-motor vehicle signal lamp group type; as shown in fig. 3, converting the phase information of the traffic signal lamp and the information of the intersection road facility into number data according to a preset number mode includes:

step S31: converting intersection road facility information into a digital number and converting traffic signal lamp phase information into a phase vector according to a preset numbering mode;

step S32: forming a facility vector by the number;

step S33: and splicing the phase vector and the facility vector, and taking the obtained spliced vector as the serial number data.

In this embodiment, the preset number manner may be a preset number table.

Converting the intersection road facility information into a digital number according to a preset numbering mode, and converting the traffic signal lamp phase information into a phase vector, which may include converting the intersection road facility information into a digital number according to a preset numbering mode, and converting the traffic signal lamp phase information into a phase vector according to a preset or randomly selected coding mode.

In practical applications, the traffic signal phase information may include various types. The numbering may be in the range of 1-100 or any range of numbers for a variety of traffic signal phase information. If the phase information of each traffic signal lamp is an integer, the phase information of the traffic signal lamp can be converted into binary numbers, octal numbers, hexadecimal numbers and the like through a binary coding mode, an octal coding mode, a hexadecimal coding mode and the like, and the converted numerical values are directly used as phase vectors.

In this embodiment, the facility vector is formed by the number numbers, and the number numbers may be converted into vectors according to a given encoding method. Or directly numbering the number as one of the multidimensional vector numbers. For example, if the multidimensional vector includes [ x1, x2, x3 … ], the number may be at least one of x1, x2, and x 3.

In the embodiment, the intersection road facility information and the traffic signal lamp phase information are subjected to vectorization expression, so that the prediction model to be trained can learn effective information of an actual sample from a vector, and a more accurate recommendation information prediction result can be generated after the training process is finished.

In one embodiment, as shown in fig. 4, augmenting the collected sample input data to generate training sample input data comprises:

step S41: carrying out dimension expansion processing on the serial number data to obtain a first vector;

step S42: performing linear rectification calculation on the first vector to obtain a second vector;

step S43: from the second vector, training sample input data is generated.

The dimension expansion processing may be performed on the number data by converting one-dimensional or two-dimensional number data into a multidimensional vector.

The first vector is subjected to a Linear rectification calculation, which may be a calculation using a Linear rectification function (ReLU). The linear rectification function, also called a modified linear unit, is an activation function (activation function) commonly used in artificial neural networks, and may generally refer to a non-linear function represented by a ramp function and a variation thereof.

In this embodiment, generating the training sample input data according to the second vector may include using the second vector as the training sample input data, or using the second vector as part of the training sample input data.

In the embodiment, through dimension expansion and linear rectification calculation, the content of the expansion data can be controlled while the data size is expanded, so that the expanded data conforms to the randomness requirement and keeps certain consistency with original real sample data.

In one embodiment, as shown in FIG. 5, generating training sample input data from the second vector comprises:

step S51: performing linear activation processing on the second vector to obtain a third vector;

step S52: sampling the content corresponding to the phase information of the traffic signal lamp in the third vector to obtain a fourth vector;

step S53: and determining training sample input data according to the fourth vector and the collected sample input data.

In this embodiment, performing linear activation processing on the second vector may include calculating the second vector by using a linear activation function to implement the linear activation processing.

In a specific implementation, the linear activation function used may include: at least one of Sigmoid (S-shaped growth curve) function, Tanh (hyperbolic tangent) function, ReLU function, and the like.

In this embodiment, through linear activation and sampling operations, data meeting requirements can be selected from new data generated in the expansion process to form data of an expanded sample, thereby increasing the sample data volume.

In this embodiment, the content corresponding to the traffic signal phase information in the third vector is sampled, and a random or other set manner sampling operation may be performed on the generated third vector that is larger than the required dimensional data to determine the fourth vector, where the content targeted by the sampling operation is the content corresponding to the traffic signal phase information.

In one embodiment, the predictive model is a decision tree model; as shown in fig. 6, the prediction model is a decision tree model; the model training method further comprises the following steps:

step S61: obtaining a sample subset according to the training sample input data, wherein the sample subset comprises at least one of the training sample input data;

step S62: constructing nodes of a decision tree model according to the sample subsets;

step S63: and obtaining a prediction model to be trained according to the nodes of the decision tree model.

Deriving the sample subset from the training sample input data may comprise using a portion of the training sample input data as the sample subset.

In this embodiment, the sample subset includes at least one of the training sample input data, which may mean that training itself includes at least one group of the training sample input data, where each group corresponds to information of related facilities, traffic flows, phases, and the like of an intersection.

And constructing nodes of the decision tree model according to the sample subset, wherein the nodes can be constructed by taking the intersection road type corresponding to the sample subset and the traffic signal lamp phase information corresponding to the intersection traffic flow information as decision references of the nodes of the decision tree model.

For example, if a sample subset contains a set of training sample input data, the type of an intersection road is a, the intersection traffic flow information is B, and the traffic signal light phase information is C, the constructed node can give a decision suggested by the traffic signal light phase information of C for an intersection with the intersection road type of a and the intersection traffic flow information of B according to the sample subset.

In the embodiment, the nodes of the decision tree model are constructed by training the sample subset of the sample input data, so that the decision tree model from the root node to the leaf nodes can be formed, and an accurate prediction result is provided for phase prediction of the traffic signal lamp.

In one embodiment, as shown in fig. 7, obtaining a sample subset from training sample input data includes:

step S71: obtaining a plurality of characteristics according to training sample input data; each feature corresponds to traffic signal lamp phase information or intersection road facility information;

step S72: determining an information gain value of each of the plurality of features on training sample input data; the information gain value is determined according to the characteristic and the conditional empirical entropy of the training sample input data;

step S73: and according to the information gain value, segmenting the input data of the training sample to obtain a sample subset.

Through the information gain value, the to-be-trained prediction model can learn the relation between the traffic information and the phase information from the actual sample, meanwhile, the influence of the actual phase information on the traffic can be learned from the traffic information, and a more accurate prediction result is provided in a prediction stage.

In the embodiment, the sample subset is determined according to the information gain value, and then the prediction model to be trained is constructed according to the sample subset, so that the prediction model can fully keep information which is useful for a prediction result in input data of a training sample, and the training effect of the prediction model to be trained is improved.

In one embodiment, as shown in fig. 8, segmenting the training sample input data according to the information gain value to obtain a sample subset includes:

step S81: determining target characteristics according to the information gain value;

step S82: determining a segmentation point of input data of a training sample according to the target characteristics;

step S83: and according to the segmentation points, segmenting the input data of the training sample to obtain a sample subset.

In this embodiment, the division point of the training sample input data may include a division point of the training sample input data. I.e. the dividing points used to divide the full set of training sample input data into subsets.

In one possible implementation, the target feature may be present in multiple forms. According to the target features, the segmentation points of the input data of the training samples are determined, and the segmentation points of the input data of the training samples can be determined by combining multiple target features.

Specifically, for example, the target features may include a-feature of the intersection road type (e.g., a cross intersection in the north-south direction), b-feature of the intersection infrastructure information, c-feature of the intersection traffic information, and d-feature of the intersection phase information, and then a, b, c, and d may be combined into a set of dividing points. The object features are classified into one group, and the non-object features are classified into another group. Features corresponding to the number (or index) of one side of the target feature may be divided into one group, and features corresponding to the number (or index) of the other side of the target feature may be divided into another group.

In the embodiment, the training sample input data is segmented according to the target characteristics, so that a binary decision tree is constructed according to the training sample input data, and the prediction effect of the prediction model is improved.

In one embodiment, training a prediction model to be trained to obtain a trained prediction model includes:

calculating loss values corresponding to leaf nodes of the decision tree;

and deleting the leaf nodes under the condition that the loss value meets the preset pruning condition.

In this embodiment, the pruning conditions may correspond to optimization conditions of the decision tree. If the decision given by the leaf node corresponding to the decision tree does not accord with the given operation efficiency of the intersection phase information, the leaf node can be deleted.

Deleting the leaf node when the loss value meets the preset pruning condition may include deleting content corresponding to the pruning condition in the leaf node when the loss value meets the preset pruning condition. For example, in a leaf node of a decision tree, three phase suggestions are given, and if one of the three phase suggestions meets the pruning condition, the phase suggestion corresponding to the pruning condition in the leaf node is deleted.

In this embodiment, nodes unsuitable for prediction in the decision tree model are deleted by pruning, so as to optimize the decision tree model.

In one embodiment, the loss value is calculated based on the number of leaf nodes of the decision tree, the number of training sample input data corresponding to the leaf nodes for each traffic signal phase information, and the number of training sample input data corresponding to the leaf nodes.

In this embodiment, the number of leaf nodes of the decision tree, the number of training sample input data corresponding to each traffic signal phase information of the leaf nodes, and the number of training sample input data corresponding to the leaf nodes can reasonably determine a loss value, so as to effectively train a prediction model to be trained.

An embodiment of the present disclosure further provides a traffic information processing method, as shown in fig. 9, including:

step S91: acquiring intersection traffic flow information and intersection road facility information of a road intersection;

step S92: and inputting the intersection traffic flow information and the intersection road facility information into a prediction model to obtain phase recommendation information, wherein the prediction model is a trained prediction model provided by any one embodiment of the disclosure.

In this embodiment, the step of inputting the intersection traffic information and the intersection facility information into the prediction model to obtain the phase recommendation information may include searching for a corresponding node in the decision tree according to the intersection traffic information and the intersection facility information, and finally obtaining the child node according to each level of nodes.

For example, the intersection traffic flow information corresponds to X index data, each index data corresponds to a first-level child node, the intersection road facility information corresponds to Y seed information, and each seed information corresponds to a first-level child node. The method comprises the steps of searching for a corresponding child node according to traffic information corresponding to a certain road intersection, searching for a series of child nodes according to the child nodes corresponding to intersection road facility information, searching for a search path from a root node to a leaf node formed by each child node in a decision tree, and finally searching for the leaf node to obtain phase recommendation information recorded in the leaf node to serve as the phase recommendation information corresponding to the road intersection.

In this embodiment, the trained prediction model provided by the embodiment of the present disclosure can be used to perform prediction, and more reasonable phase recommendation information is provided according to the displayed road traffic facility information and traffic flow information.

In an example of the present disclosure, a method for recommending an urban road intersection traffic signal control phase setting scheme is provided, as shown in fig. 10, including the following steps:

s101: and collecting a certain number of samples of the actual intersections of the urban roads. The sample information comprises four parts of intersection traffic organization channelizing, traffic signal control facilities, traffic flow characteristics and the current signal phase setting scheme.

S102: and carrying out classified marking on the traffic organization channelized information and the traffic signal control facility information of each sample intersection.

S103: and calculating the traffic flow characteristic index of the sample intersection.

S104: and carrying out classification numbering on the signal phase setting scheme of each sample intersection. Considering the common 10 phase setting scheme types, the collected intersection samples will be classified according to the phase setting scheme in table 3 below, where the right-turn arrow indicates that the right-turn traffic is controlled by a dedicated direction indicator light group.

TABLE 3 intersection phase setup scheme type and numbering

S105: based on the collected intersection sample set, the sample capacity is expanded by using a generation countermeasure network (GAN) algorithm.

S106: training and generating a binary decision tree for phase setting scheme recommendation based on the real intersection samples and simulated intersection samples generated by the GAN.

S107: and pruning the generated decision tree T. The number of leaf nodes of the tree T is | T |, T is the leaf node of the tree T, the leaf node is provided with Nt intersection samples, wherein Ntk samples with the type of phase scheme Ck are provided, K is 1,2, …, K, alpha is more than or equal to 0 as regularization parameters, and then the loss function of the decision tree learning is

S108: and recommending the intersection phase setting scheme by using the pruned decision tree T alpha.

In this example, the type of the intersection phase setting scheme in table 3 may be determined according to an actual lane line in the road. The dashed lines may represent default schemes.

Further, the intersection features are split into two parts, namely organization channelizing, static features of signal control facilities and dynamic features of traffic flow characteristics in steps S102 and S103, wherein the static features limit the selection space of the signal phase scheme, and the dynamic features play a role in restricting the preference of the phase scheme.

Further, in step S105, the generation countermeasure network algorithm is used to expand the sample capacity, so that on one hand, enough samples are ensured to train an effective decision tree; on the other hand, the intersection feature values of the sample set are enriched, and the distribution of samples of different phase scheme types is balanced through sample screening, so that the problem that the recommendation accuracy of the decision tree phase scheme is reduced due to the fact that part of samples of the phase scheme types are distributed too intensively is solved.

In an example of the present disclosure, as shown in fig. 11, collecting a number of samples of an actual intersection of an urban road includes:

s111: the method comprises the steps of collecting intersection traffic organization channelized information, wherein the intersection traffic organization channelized information comprises an intersection type, a left-turn lane type, a right-turn organization type, a left-turn waiting area setting condition and a pedestrian crossing organization condition.

S112: the method comprises the steps of collecting traffic signal control facility information of a cross, wherein the traffic signal control facility information comprises a left-turning motor vehicle signal lamp group type, a right-turning motor vehicle signal lamp group type and a non-motor vehicle signal lamp group type.

S113: and collecting intersection traffic flow characteristic information, including lane flow and average saturated headway data of each traffic flow in a certain period of time.

S114: and acquiring an intersection signal phase setting scheme, including a signal control phase and phase sequence scheme in a corresponding time period.

In another embodiment of the present disclosure, the classifying and marking the traffic organization channelized information and the traffic signal control facility information of each sample intersection includes:

considering two types of common crossroads and T-shaped crossroads, the crossroad is split into two independent samples according to orthogonal angles, as shown in fig. 12, including sample 1: corresponding to east-west entry lane and sample 2: corresponding to the north-south entry. The T-junction is treated as a single whole sample, and 3 inlet lanes are numbered respectively, as shown in fig. 13, including inlet lane 1, inlet lane 2, and inlet lane 3. And carrying out 0-1 classification marking on the traffic organization channelized information and the traffic signal control facility information of the sample intersection according to the following table 1.

TABLE 1 intersection traffic organization channeling and traffic signal control facility information classification labels

In one example, calculating a sample intersection traffic flow characteristic indicator, as shown in fig. 14, includes:

s141: and respectively numbering key traffic flows of the crossroads and the T-shaped intersections corresponding to the collected samples. The numbering can be as shown in fig. 15 and fig. 16 respectively, the number of the crossroad can be m1-m4, and the number corresponding to each entrance (Leg) of the T-shaped intersection can be m1-m 3.

S142: and calculating the traffic flow ratio of each important traffic flow of each intersection sample according to the data in the collected samples. When the traffic flow qi is the lane flow (pcu/h/lane) of the traffic flow mi, hi is the saturated headway(s) of the traffic flow mi, the traffic flow ratio yi of the traffic flow mi is qihi/3600. Where i may correspond to the numbers in the m subscripts in fig. 15, 16.

S143: and calculating the traffic flow index according to a preset formula. Defining the Difference (DR) indexes of the key traffic demands of each strand, namely bidirectional straight traffic DR1, bidirectional left-turning traffic DR2, opposite straight left-turning conflict traffic DR3, and same-direction straight left-turning traffic DR4 and DR5, wherein the values of the indexes can be calculated according to the following table 2:

TABLE 2 difference index calculation method for key traffic flow passing demands

In one example, as shown in fig. 17, expanding the sample volume with a generative countermeasure network (GAN) algorithm based on the collected intersection sample set includes:

s171: and constructing a feature vector of the collected real intersection sample. Specifically, according to the sequence listed in table 1 and table 2, each intersection sequentially combines the intersection information classification labels and the traffic flow characteristic index values into a 13-dimensional feature vector Xi ═ Xi (1), Xi (2), …, Xi (13) ], where Xi represents the feature vector of the intersection i.

S172: and constructing a phase scheme type vector of the collected real intersection sample. Each intersection is one-hot encoded according to the phase setting scheme type number in table 3, specifically, the phase scheme type is represented by a 10-dimensional vector Yi ═ Yi (1), Yi (2), …, Yi (10) ], for example, a vector [1,0,0, …,0] represents a phase scheme type 1, a vector [0,1,0, …,0] represents a phase scheme type 2, and so on, where Yi represents a phase scheme type vector of the intersection i.

S173: and transversely combining the feature vector and the phase scheme type vector of each intersection to form an intersection sample 23-dimensional information vector. That is, Zi is [ xi (1), xi (2), …, xi (13), yi (1), yi (2), …, yi (10) ], where Zi represents the sample information vector of intersection i [ Zi (1), Zi (2), …, Zi (23) ]. Here, i may represent the number of the intersection. For example, data of 10 ten thousand intersections are collected, and the number of the intersections can be 1-100000.

S174: a generative confrontation network (GAN) model is constructed. The GAN model comprises a generation model and a discriminant model.

The structure of the generated model is shown in fig. 18, the generated model is input as m-dimensional random prior noise, is expanded into j-dimension by the full connection layer 1, is subjected to a linear rectification function ReLU, is input into the full connection layer 2, outputs a 23-dimensional tensor, is activated by a nonlinear activation function Tanh, is split into a sample eigenvector and a phase scheme type vector, samples data representing the phase scheme type vector in the tensor through a Gumbel-Softmax (gunbren normalization) layer, and outputs a complete tensor (simulation sample) representing intersection sample information based on the sample eigenvector and a normalized result.

The structure of the discrimination model is shown in fig. 19, the input of the discrimination model is a real intersection sample or a simulation sample tensor generated by a generation model, the simulation sample tensor is expanded into an l-dimensional tensor through the full connection layer 1 and is input into the full connection layer 2 after passing through a linear rectification function ReLU, a 1-dimensional numerical value is output, and phase recommendation and a corresponding probability result PD are output after being processed by a nonlinear activation function Sigmoid. α and β in fig. 18 and 19 indicate intermediate vectors or intermediate data.

S175: and training the GAN model by using the real intersection sample set until the model value function is converged, and then generating a simulated intersection sample.

S176: and eliminating invalid simulated intersection samples, wherein samples of which the phase scheme types are opposite to intersection traffic organization channelizing and signal control facilities in the simulated intersection samples generated by the GAN are invalid samples. The judgment conditions are shown in the following table 4:

table 4 invalid intersection sample judgment condition ("×" indicates that the phase scheme type is invalid)

Further, in step S174, a Gumbel-Softmax sampling layer is added to the GAN generation model to implement one-hot transformation of the generated simulated intersection phase scheme type vector, and backward propagation of the model parameter gradient is not blocked, so that iterative update of the model parameters is ensured.

In one example, simulated intersection samples generated based on real intersection samples and GAN are trained and a binary decision tree for phase setting scheme recommendation is generated, as shown in fig. 20, including:

s201: adding a part of effective simulated intersection samples into a real intersection sample set to form a full sample set U, so that the intersection phase scheme types of the sample set U are distributed as uniformly as possible; and randomly dividing the samples in the U into a training sample set D and a testing sample set F according to a certain proportion, and ensuring that the D and the F both contain samples of all phase scheme types in the U.

S202: a training sample set D is input. Each intersection sample input vector comprises 13 characteristic condition values and phase scheme type numbers, the intersection information characteristic set contained in the sample input vector is B, | D | represents the sample capacity of D, namely the number of samples, the samples in D cover K phase scheme types Ck, K is 1,2, …, K, | Ck | is the number of samples of the phase types Ck, and the empirical entropy of the sample set D is

S203: and if the sample phase type in the D is the same type Ck, the decision tree T is a single node tree, the type Ck is used as the class mark of the node, and the T is returned.

S204: if it is

T is a single node tree, the serial numbers of all types of phase schemes of the samples in D and the sample ratios thereof are used as class marks of the node, and T is returned.

S205: otherwise, calculating the information gain of each characteristic condition pair D, and selecting the characteristic Ag with the maximum information gain. The characteristic conditions are of two types, and the information gains are calculated according to the following methods respectively:

(1) for the traffic organization channelized information and the traffic signal control facility information shown in table 1, which are binary variables of 0-1, D is divided into 2 subsets D1 and D2 according to the value of some characteristic condition a, i Di | is the number of samples of Di, the sample set of phase type Ck in Di is Dik, i Dik | is the number of samples of Dik, and then the conditional empirical entropy of the characteristic condition a on D is

The information gain of the characteristic condition a is g (D, a) ═ H (D) — H (D | a).

(2) For the traffic flow characteristic index shown in table 2, which is a continuous value variable, in order to construct a binary decision tree, the characteristic condition a of a certain continuous value and a certain value s thereof in the sample set D are made to be a segmentation variable and a segmentation point, and the sample set can be divided into two subsets: d1(a, s) { D | a ≦ s }, and D2(a, s) { D | a > s }, then traversing values of the feature condition a in all samples in D, solving argmins [ H (D1) + H (D2) ], wherein s obtained is an optimal segmentation point a, wherein H (Di) is an empirical entropy of a sample set Di, and calculating an information gain of the feature condition a according to the method in (1) according to two subsets segmented by the segmentation point s.

S206: if the information gain of Ag is equal to 0, setting T as a single node tree, taking the serial number of each type of phase scheme of the samples in D and the sample ratio thereof as the class mark of the node, and returning to T.

S207: otherwise, for each value of Ag (for example, a continuous value-taking characteristic condition, a value-taking range based on a cut point) ai (i is 1,2), D is divided into two non-empty subsets D1 and D2, two sub-nodes are constructed, a tree T is formed by the nodes and the sub-nodes, and T is returned.

S208: and (3) recursively calling the steps S202 to S207 by taking Di as a training set and B- { Ag } as a feature set for the ith (i is 1 and 2) child node to obtain a subtree Ti, and returning to the step Ti.

Further, in step S205, for the continuous value-taking characteristics, a heuristic search method is used to find an optimal segmentation point, and the characteristic value is divided into two intervals, thereby implementing the construction of the binary decision tree.

In one embodiment, as shown in fig. 21, the pruning process includes:

s211: the generated overall tree T and the parameter α are input. The overall tree may be a binary decision tree generated by the method shown in fig. 20.

S212: recursively retracting upward from the leaf nodes of the tree. And if the C alpha (TB) is less than or equal to the C alpha (TA), pruning is carried out, namely the father node is changed into a new leaf node, and the classification mark of the leaf node is updated.

S213: and returning and repeating the step S212 until all the leaf nodes do not meet the pruning condition, and obtaining the subtree T alpha with the minimum loss function.

Further, in steps S106 and S108, when the decision tree is generated, the subset in which each leaf node marker is located includes each phase scheme type and sample proportion, the sample proportions of different phase scheme types may be used as confidence levels of different phase schemes selected for intersection samples meeting the feature condition corresponding to the leaf node, so that a plurality of phase alternatives may be recommended to the designer.

After the technical scheme is adopted, the technical scheme provided by the example of the disclosure has at least the following beneficial effects:

the method comprehensively considers static conditions and traffic flow dynamic characteristics of various traffic organization facilities at the intersection, realizes recommendation of an optimal signal control phase setting scheme through analysis and matching, and is more comprehensive in considered intersection information characteristics and phase scheme types compared with the existing phase scheme optimization model method.

According to the method, the intersection sample set is used for generating the decision tree, the input and the output of the decision tree are programmed, a complex mathematical model does not need to be specially established and solved in the actual application process of the phase scheme optimization design, and the optimal phase setting scheme of the intersection can be conveniently obtained under the condition that a user does not have rich signal optimization experience.

According to the method, a plurality of phase alternative schemes and confidence degrees thereof are provided for the intersection, and a user can flexibly select and apply the alternative schemes according to specific requirements, so that a larger operation space is provided for the intersection signal coordination control scheme design and other works.

In another example of the present disclosure, the model training and prediction steps implemented for the data of an actual intersection as shown in fig. 22 are as follows:

s221: 159 groups of actual intersection sample data are collected from cities such as A, B, and the sample information comprises four major parts of intersection traffic organization channelizing, traffic signal control facilities, traffic flow characteristics and the current signal phase setting scheme.

S222: and carrying out classified marking on the traffic organization channelized information and the traffic signal control facility information of each sample intersection.

S223: and calculating the traffic flow characteristic index of the sample intersection.

S224: and carrying out classification numbering on the signal phase setting scheme of each sample intersection. The actual samples collected include 121 crossroads and 38T-intersections, and cover 9 types of phase setting schemes in table 3, and the number of intersection samples for each type of phase scheme is shown in table 5 below.

TABLE 5 number of intersection samples of different phase scheme types

S225: based on the collected intersection sample set, the sample capacity is expanded by using a generation countermeasure network (GAN) algorithm.

S226: training and generating a binary decision tree for phase setting scheme recommendation based on the real intersection samples and simulated intersection samples generated by the GAN. In order to enable the types of the intersection phase schemes of the decision tree training sample set to be distributed more uniformly, 9-17-153 samples are randomly selected from effective intersection samples generated by the GAN model and added into a real intersection sample set to form a full sample set. The number of the selected simulated intersection samples is close to the number of the real samples, and 17 samples are respectively selected for each type of phase scheme. Then, 48 samples are randomly extracted from the full sample set to form a test sample set, and the rest samples form a training sample set.

As shown in fig. 23, the expanding the sample volume by using a generated countermeasure network (GAN) algorithm based on the collected intersection sample set further includes:

s231: and constructing a 13-dimensional feature vector of each acquired real intersection sample.

S232: and constructing a 10-dimensional phase scheme type vector of each acquired real intersection sample.

S233: and transversely combining the feature vector and the phase scheme type vector of each intersection to form an intersection sample 23-dimensional information vector.

S234: and constructing a GAN model. The GAN model can be constructed according to fig. 18 and fig. 19 using a PyTorch or the like tool.

S235: the samples are expanded. The GAN model is trained using the set of real intersection samples until the model cost function converges, and then 159 x 100-15900 simulated intersection samples are generated using the trained generative model.

S236: and eliminating invalid simulated intersection samples. In 15900 simulated intersection samples, 796 invalid samples are available, wherein the types of the phase schemes are contradictory to intersection traffic organization channelizing and signal control facilities, and the effective sample rate is about 95%.

S237: and pruning the generated decision tree. The pruning parameter α is taken as 10, and the decision tree after pruning is shown in fig. 31, in which the leaf nodes are labeled as the numbers of all recommended solution types and their confidence degrees.

S238: and (4) carrying out phase setting scheme recommendation on 48 intersections of the test sample set by using the decision tree after pruning. The results are shown in Table 6 below. The recommended phase alternative schemes of all intersection samples comprise actual current phase schemes, and account for 100%; the phase scheme with the maximum confidence in the recommended phase alternative scheme set of 44 intersection samples is consistent with the existing scheme, accounts for 92%, and has a good testing effect.

Referring to fig. 31, the root node of the decision tree after pruning is the intersection type. The first-level sub-node comprises a left-turn motor vehicle signal lamp group type and a combination of number numbers corresponding to the intersection type. The third level of non-leaf nodes includes a combination of numbers corresponding to the left turn vehicle signal group type for the straight left conflicting traffic (i.e., DR 3). And the fourth-level non-leaf node comprises a non-motor vehicle signal lamp group type and a combination of index ranges corresponding to the direct left collision traffic flows of the object. The fourth-level non-leaf node further comprises a combination of the left-turn lane type and the index range corresponding to the object direct-left conflict traffic flow. And the fifth-stage non-leaf node comprises the setting condition of the left-turning waiting area and the combination of the number corresponding to the type of the non-motor vehicle signal lamp group. The fifth level non-leaf node also includes a combination of the same direction straight left turn traffic (i.e., DR5) and a number corresponding to the non-motor vehicle signal light group type. And the sixth-level non-leaf node comprises the combination of the right-turn tissue type and the setting condition of the left-turn waiting area. The seventh level non-leaf node comprises a combination of the same-direction straight-going left-turn traffic and a number corresponding to the right-turn organization type. In the leaf nodes of fig. 31, the numbers in "[ ]", integers represent the phase, and percentages represent the degree of recommendation.

Table 6 test sample set intersection phase setting scheme recommendation results

In the above table, the numbers, integers, represent phases and percentages represent recommended degrees in "[ ] ]".

Generally, an optimization method for an intersection phase setting scheme is mostly based on an objective of improving traffic operation benefit of intersections to establish a mathematical model, and a traditional optimization algorithm or an intelligent search algorithm is used for solving the model, so that the optimization of the phase scheme is realized, only traffic requirements or a few specific intersection road conditions are considered, and a relatively complex mathematical model needs to be established and solved in practical application, so that the method has great limitation. According to the model training method and the traffic information processing method provided by the embodiment of the disclosure, static conditions and traffic flow dynamic characteristics of various traffic organization facilities at the intersection are comprehensively considered, and a decision tree is constructed to realize quick recommendation of an optimal phase setting scheme of the intersection and scientifically and reasonably improve the efficiency of design work of an intersection signal control scheme.

An embodiment of the present disclosure further provides a model training apparatus, as shown in fig. 24, including:

a sample collecting module 241, configured to obtain a collected sample collected at a road intersection; the collected samples comprise phase information of traffic lights at the road intersections, traffic flow information at the intersections and facility information of the roads at the intersections;

a conversion module 242, configured to convert the collected sample into collected sample input data according to a preset information data conversion manner;

a training sample input data module 243, configured to expand the collected sample input data to generate training sample input data;

and the training module 244 is used for training a prediction model to be trained according to the input data of the training samples to obtain the trained prediction model, and the prediction model is used for outputting signal lamp phase recommendation information according to the intersection traffic information and the intersection road facility information.

In one embodiment, as shown in fig. 25, the preset information data conversion mode includes a preset number mode and a preset calculation mode; the conversion module includes:

a numbering unit 251 for converting the phase information of the traffic signal and the information of the intersection road facility into numbering data according to a preset numbering mode;

the index unit 252 is configured to calculate traffic index data according to the intersection traffic information in a preset calculation manner;

and the data processing unit 253 is used for obtaining the collected sample input data according to the serial number data and the traffic flow index data.

In one embodiment, the indicator unit is further configured to:

In one embodiment, the intersection asset information comprises: the system comprises intersection type information, left-turn lane widening information, right-turn lane channelizing information, left-turn waiting area information, pedestrian street crossing facility information, left-turn motor vehicle signal lamp group type, right-turn motor vehicle signal lamp group type information and non-motor vehicle signal lamp group type; the numbering unit is also used for:

converting intersection road facility information into a digital number and converting traffic signal lamp phase information into a phase vector according to a preset numbering mode;

forming a facility vector by the number;

and splicing the phase vector and the facility vector, and taking the obtained spliced vector as the serial number data.

In one embodiment, as shown in FIG. 26, the training sample input data module comprises:

a first vector unit 261, configured to perform dimension expansion processing on the number input to obtain a first vector;

a second vector unit 262, configured to perform linear rectification calculation on the first vector to obtain a second vector;

and a second vector processing unit 263, configured to generate training sample input data according to the second vector.

In one embodiment, the second vector processing unit is further configured to:

performing linear activation processing on the second vector to obtain a third vector;

sampling the content corresponding to the phase information of the traffic signal lamp in the third vector to obtain a fourth vector;

and determining training sample input data according to the fourth vector and the collected sample input data.

In one embodiment, the predictive model is a decision tree model; as shown in fig. 27, the model training apparatus further includes:

a sample subset module 271, configured to obtain a sample subset according to the training sample input data, where the sample subset includes at least one of the training sample input data;

a node module 272, configured to construct nodes of the decision tree model according to the sample subsets;

and the model module 273 is configured to obtain the prediction model to be trained according to the nodes of the decision tree model.

In one embodiment, as shown in fig. 28, the sample subset module includes:

a feature unit 281 configured to obtain a plurality of features according to training sample input data; each feature corresponds to traffic signal lamp phase information or intersection road facility information;

an empirical entropy unit 283 for determining an information gain value for each of a plurality of features on the training sample input data; the information gain value is determined according to the characteristic and the conditional empirical entropy of the training sample input data;

and a segmenting unit 283, configured to segment the training sample input data according to the information gain value to obtain a sample subset.

In one embodiment, the segmentation unit is further configured to:

determining target characteristics according to the information gain value;

determining a segmentation point of input data of a training sample according to the target characteristics;

and according to the segmentation points, segmenting the input data of the training sample to obtain a sample subset.

In one embodiment, as shown in fig. 29, the training module comprises:

a loss value unit 291, configured to calculate a loss value corresponding to a leaf node of the decision tree according to training sample input data;

a pruning unit 292 configured to delete the leaf node if the loss value meets a preset pruning condition.

An embodiment of the present disclosure further provides a traffic information processing apparatus, as shown in fig. 30, including:

the information acquisition module 301 is configured to acquire intersection traffic information and intersection road facility information of a road intersection;

the phase prediction module 302 is configured to input intersection traffic information and intersection road facility information into a prediction model to obtain phase recommendation information, where the prediction model is a trained prediction model provided in any one of the embodiments of the present disclosure.

In the technical scheme of the disclosure, the acquisition, storage, application and the like of the personal information of the related user all accord with the regulations of related laws and regulations, and do not violate the good customs of the public order.

The present disclosure also provides an electronic device, a readable storage medium, and a computer program product according to embodiments of the present disclosure.

FIG. 32 shows a schematic block diagram of an example electronic device 320 that may be used to implement embodiments of the present disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the disclosure described and/or claimed herein.

As shown in fig. 32, the apparatus 320 includes a computing unit 321, which can perform various appropriate actions and processes in accordance with a computer program stored in a Read Only Memory (ROM)322 or a computer program loaded from a storage unit 328 into a Random Access Memory (RAM) 323. In the RAM 323, various programs and data required for the operation of the device 320 can also be stored. The calculation unit 321, the ROM 322, and the RAM 323 are connected to each other via a bus 324. An input/output (I/O) interface 325 is also connected to bus 324.

A number of components in device 320 are connected to I/O interface 325, including: an input unit 326 such as a keyboard, a mouse, or the like; an output unit 327 such as various types of displays, speakers, and the like; a storage unit 328 such as a magnetic disk, optical disk, or the like; and a communication unit 329 such as a network card, a modem, a wireless communication transceiver, or the like. The communication unit 329 allows the device 320 to exchange information/data with other devices through a computer network such as the internet and/or various telecommunication networks.

The computing unit 321 may be a variety of general-purpose and/or special-purpose processing components having processing and computing capabilities. Some examples of the computing unit 321 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, and so forth. The computing unit 321 performs the various methods and processes described above, such as the model training method. For example, in some embodiments, the model training method may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as storage unit 328. In some embodiments, some or all of the computer program may be loaded and/or installed onto device 320 via ROM 322 and/or communication unit 329. When the computer program is loaded into RAM 323 and executed by computing unit 321, one or more steps of the model training method described above may be performed. Alternatively, in other embodiments, the computing unit 321 may be configured to perform the model training method by any other suitable means (e.g., by means of firmware).

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuitry, Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), system on a chip (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

Program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program codes, when executed by the processor or controller, cause the functions/operations specified in the flowchart and/or block diagram to be performed. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.

In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), Wide Area Networks (WANs), and the Internet.

The computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server may be a cloud server, a server of a distributed system, or a server with a combined blockchain.

It should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present disclosure may be executed in parallel, sequentially, or in different orders, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and the present disclosure is not limited herein.

The above detailed description should not be construed as limiting the scope of the disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made, depending on design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present disclosure should be included in the scope of protection of the present disclosure.

Claims

1. A model training method, comprising:

2. The method according to claim 1, wherein the preset information data conversion mode comprises a preset numbering mode and a preset calculation mode; the converting the collected sample into collected sample input data according to a preset information data conversion mode comprises:

converting the traffic signal lamp phase information and the intersection road facility information into serial number data according to the preset serial number mode;

according to the preset calculation mode, calculating traffic flow index data according to the intersection traffic flow information;

and acquiring the collected sample input data according to the serial number data and the traffic flow index data.

3. The method according to claim 2, wherein the calculating traffic flow index data according to the intersection traffic flow information in the preset calculation manner includes:

and calculating traffic flow index data according to the intersection type information, the intersection roads corresponding to the intersection traffic flow information and the intersection traffic flow information.

4. The method according to claim 2 or 3, wherein the intersection asset information comprises: the system comprises intersection type information, left-turn lane widening information, right-turn lane channelizing information, left-turn waiting area information, pedestrian street crossing facility information, left-turn motor vehicle signal lamp group type, right-turn motor vehicle signal lamp group type information and non-motor vehicle signal lamp group type; according to the preset numbering mode, the phase information of the traffic signal lamps and the intersection road facility information are converted into numbering data, and the method comprises the following steps:

converting the intersection road facility information into a digital number and converting the traffic signal lamp phase information into a phase vector according to the preset numbering mode;

composing the number into a facility vector;

5. The method of claim 4, wherein said augmenting said collected sample input data to generate training sample input data comprises:

carrying out dimension expansion processing on the serial number data to obtain a first vector;

performing linear rectification calculation on the first vector to obtain a second vector;

and generating training sample input data according to the second vector.

6. The method of claim 5, wherein the generating training sample input data from the second vector comprises:

sampling the content corresponding to the traffic signal lamp phase information in the third vector to obtain a fourth vector;

determining the training sample input data according to the fourth vector and the collected sample input data.

7. The method of any of claims 1-6, wherein the predictive model is a decision tree model; the method further comprises the following steps:

obtaining a sample subset according to the training sample input data, wherein the sample subset comprises at least one of the training sample input data;

constructing nodes of a decision tree model according to the sample subset;

and obtaining the prediction model to be trained according to the nodes of the decision tree model.

8. The method of claim 7, wherein said deriving a subset of samples from said training sample input data comprises:

obtaining a plurality of features according to the training sample input data; each feature corresponds to traffic signal lamp phase information or intersection road facility information;

determining an information gain value for each of the plurality of features on the training sample input data; the information gain value is determined according to the characteristic and the conditional empirical entropy of the training sample input data;

and segmenting the training sample input data according to the information gain value to obtain the sample subset.

9. The method of claim 8, wherein said segmenting the training sample input data according to the information gain values to obtain the sample subset comprises:

determining target characteristics according to the information gain value;

determining a segmentation point of the training sample input data according to the target feature;

and segmenting the training sample input data according to the segmentation points to obtain the sample subset.

10. The method according to any one of claims 7 to 9, wherein the training a prediction model to be trained according to the training sample input data to obtain a trained prediction model comprises:

calculating loss values corresponding to leaf nodes of the decision tree according to the training sample input data;

11. The method of claim 10, wherein the loss value is calculated according to the number of leaf nodes of the decision tree, the number of training sample input data corresponding to each traffic signal phase information of the leaf nodes, and the number of training sample input data corresponding to the leaf nodes.

12. A traffic information processing method, comprising:

inputting the intersection traffic flow information and the intersection road facility information into a prediction model to obtain phase recommendation information, wherein the prediction model is the trained prediction model according to any one of claims 1 to 11.

13. A model training apparatus comprising:

the acquisition sample module is used for acquiring an acquisition sample acquired at a road intersection; the collected samples comprise traffic signal lamp phase information, intersection traffic flow information and intersection road facility information of the road intersection;

14. The apparatus according to claim 13, wherein the predetermined information data conversion manner includes a predetermined number manner and a predetermined calculation manner; the conversion module includes:

the numbering unit is used for converting the traffic signal lamp phase information and the intersection road facility information into numbering data according to the preset numbering mode;

the index unit is used for calculating traffic flow index data according to the preset calculation mode and the intersection traffic flow information;

and the data processing unit is used for obtaining the acquired sample input data according to the serial number data and the traffic flow index data.

15. The apparatus of claim 14, wherein the metric unit is further to:

16. The apparatus of claim 14 or 15, wherein the intersection asset information comprises: the system comprises intersection type information, left-turn lane widening information, right-turn lane channelizing information, left-turn waiting area information, pedestrian street crossing facility information, left-turn motor vehicle signal lamp group type, right-turn motor vehicle signal lamp group type information and non-motor vehicle signal lamp group type; the numbering unit is further configured to:

composing the number into a facility vector;

17. The apparatus of claim 16, wherein the training sample input data module comprises:

the first vector unit is used for carrying out dimension expansion processing on the serial number data to obtain a first vector;

the second vector unit is used for performing linear rectification calculation on the first vector to obtain a second vector;

and the second vector processing unit is used for generating training sample input data according to the second vector.

18. The apparatus of claim 17, wherein the second vector processing unit is further to:

19. The apparatus of any of claims 13-18, wherein the predictive model is a decision tree model; the device further comprises:

a sample subset module, configured to obtain a sample subset according to the training sample input data, where the sample subset includes at least one of the training sample input data;

the node module is used for constructing nodes of a decision tree model according to the sample subset;

and the model module is used for obtaining the prediction model to be trained according to the nodes of the decision tree model.

20. The apparatus of claim 19, wherein the sample subset module comprises:

the characteristic unit is used for obtaining a plurality of characteristics according to the training sample input data; each feature corresponds to traffic signal lamp phase information or intersection road facility information;

an empirical entropy unit to determine an information gain value for each of the plurality of features to the training sample input data; the information gain value is determined according to the characteristic and the conditional empirical entropy of the training sample input data;

and the segmentation unit is used for segmenting the training sample input data according to the information gain value to obtain the sample subset.

21. The apparatus of claim 20, wherein the segmentation unit is further configured to:

determining target characteristics according to the information gain value;

22. The apparatus of any of claims 19-21, wherein the training module comprises:

the loss value unit is used for calculating a loss value corresponding to a leaf node of the decision tree according to the training sample input data;

and the pruning unit is used for deleting the leaf nodes under the condition that the loss value meets the preset pruning condition.

23. The apparatus of claim 22, wherein the loss value is calculated according to the number of leaf nodes of the decision tree, the number of training sample input data corresponding to each traffic signal phase information of the leaf nodes, and the number of training sample input data corresponding to the leaf nodes.

24. A traffic information processing apparatus comprising:

the phase prediction module is configured to input the intersection traffic information and the intersection road facility information into a prediction model to obtain phase recommendation information, where the prediction model is the trained prediction model according to any one of claims 12 to 23.

25. An electronic device, comprising:

at least one processor; and

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-12.

26. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-12.

27. A computer program product comprising computer programs/instructions, characterized in that the computer programs/instructions, when executed by a processor, implement the steps of the method of any of claims 1 to 12.