KR102021992B1 - Apparatus for controling a trafic signal, method for controling a trafic signal, and recoding medium for controling a tarfic signal - Google Patents

Abstract

The present invention relates to an apparatus for controlling a traffic signal, to a method for controlling a traffic signal, and to a storage medium storing a traffic signal control program. According to one embodiment of the present invention, the apparatus for controlling a traffic signal comprises: a composite data input unit generating composite data for predicting traffic congestion, wherein the composite data is space-time image data based on atmospheric information, area information, and time information; and a traffic flow prediction unit predicting the traffic congestion using the composite data. The traffic flow prediction unit includes: an image labeling control portion labeling the image data of the composite data; and an image CNN learning portion applying a CNN to the labeled image.

Description

Storage device for storing traffic signal control device, traffic signal control method and traffic signal control program {APPARATUS FOR CONTROLING A TRAFIC SIGNAL, METHOD FOR CONTROLING A TRAFIC SIGNAL, AND RECODING MEDIUM FOR CONTROLING A TARFIC SIGNAL}

The present invention relates to a traffic signal control apparatus, a traffic signal control method, and a storage medium for storing a traffic signal control program.

Interest in traffic information is increasing with the use of public transport and own vehicles. Various methods of using traffic information to predict traffic congestion have been proposed. However, these methods cannot be efficiently provided in consideration of various factors that affect traffic. While using the existing traffic signal system, there is a limit in dealing with the traffic congestion in the time zone or region where the traffic is congested.

Deep learning technology has begun to be used to predict traffic congestion, but it has not been able to predict traffic congestion spatially or temporally in urban areas with heavy traffic.

It is an object of the present invention to provide a storage medium for storing a traffic signal control device, a traffic signal control method, and a traffic signal control program, which can spatially predict a traffic congestion area of a downtown area.

Another object of the present invention is to provide a storage medium for storing a traffic signal control device, a traffic signal control method, and a traffic signal control program, which can predict a repetitive traffic congestion time in a downtown area in real time.

Another object of the present invention is to predict traffic congestion in terms of time or space based on deep learning technology, and to remotely control a traffic signal cycle based on artificial intelligence (AI) according to the prediction result. To provide a signal control apparatus, a traffic signal control method and a storage medium for storing a traffic signal control program.

Another object of the present invention is to provide a traffic signal control device, a traffic signal control method, and a traffic signal control program, which can hybridly control a traffic signal by reflecting both traffic congestion and traffic congestion. It is to provide a storage medium for storing.

One embodiment of the present invention, a composite data input unit for generating a composite data for traffic congestion prediction, wherein the composite data is space-time image data based on atmospheric information, area information, and time information; And a traffic flow predictor for predicting traffic congestion using the composite data, wherein the traffic flow predictor is an image labeling controller for labeling the image data of the composite data, and an image applying a CNN to the labeled image. It provides a traffic signal control device including a CNN learning unit.

In one embodiment of the present invention, a composite data generation step of generating composite data for traffic congestion prediction, wherein the composite data is space-time image data based on atmospheric information, region information, and time information; And a traffic flow prediction step of predicting traffic congestion using the composite data, wherein the traffic flow prediction step comprises: an image labeling step of labeling the image data of the composite data; and applying a CNN to the labeled image. It provides a traffic signal control method comprising an image CNN step.

One embodiment of the present invention, a composite data input module for generating composite data for traffic congestion prediction, wherein the composite data is space-time image data based on atmospheric information, area information, and time information; And a traffic flow prediction module for predicting traffic congestion using the composite data, wherein the traffic flow prediction module comprises: an image labeling control module for labeling the image data of the composite data; Provides a storage medium for storing a program for controlling traffic signals, including an image CNN learning module applying Convolution Neutral Network.

According to the present invention, traffic congestion can be predicted in consideration of spatial or temporal factors.

According to the present invention, traffic congestion can be predicted in real time through a pattern in which traffic congestion occurs.

According to the present invention, the traffic signal can be efficiently controlled in consideration of the presence or absence of traffic congestion.

1 is a view showing a configuration of a traffic signal control apparatus according to an embodiment of the present invention.
2 is a diagram illustrating an example of generating input data according to an embodiment of the present invention.
3 is a diagram illustrating an example of generating various channel data with factors that may affect traffic conditions.
4 is a diagram illustrating an example of a deep learning based extended channel image CNN (Convolution Neutral Network) learning unit according to an embodiment of the present invention.
5 is a diagram illustrating a configuration of a traffic signal control apparatus according to an embodiment of the present invention.
FIG. 6 is a diagram illustrating a configuration of a traffic signal control device long-short memory (LSTM) learning unit according to an embodiment of the present invention.
7 illustrates an example of deep learning based real-time traffic congestion prediction and remote traffic signal control according to an embodiment of the present invention.
8 illustrates an example of a deep Q-network (DQN) traffic signal cycle control unit based on Reinforcement Learning according to an embodiment of the present invention.
9 is a view showing an example of a traffic signal control method according to an embodiment of the present invention.

Hereinafter, a detailed embodiment of the present invention will be described with reference to the drawings.

1 is a view showing a configuration of a traffic signal control apparatus according to an embodiment of the present invention.

The traffic signal control apparatus according to an embodiment of the present invention may include a customized composite data input unit 1100, a deep learning based real-time traffic flow prediction unit 1200, and / or a traffic congestion area congestion time signal controller 1300. have.

The customized composite data input unit 1100 may receive traffic congestion information for traffic congestion prediction. The traffic congestion information may include weather information, atmospheric information, local information, time information, etc. according to the type of data. The weather information indicates weather related information such as temperature and wind speed, and the atmospheric information means information related to the atmosphere such as fine dust (for example, PM10, PM2.5, greenhouse gas, carbon dioxide, carbon monoxide, etc.), air pollution, The area information may include information related to the target area such as speed, lane, and road environment, and the time information may include information on days of the week such as holidays and time zones in which the day is divided into time sections (for example, morning, afternoon, Evening, dawn, morning, etc.). According to one embodiment of the present invention, the traffic congestion information may include all information related to traffic conditions.

According to an embodiment, the customized composite data input unit 1100 may convert traffic congestion information into composite data. The complex data conversion here refers to a process of converting traffic congestion information into space-time image data. Through the space-time image data conversion process, data can be converted into an image visually or spatially, and the data can be converted into a two-dimensional space-time image. Referring to Figure 1, the custom composite data input unit 1100 according to an embodiment of the present invention is a two-dimensional space-time image, fine dust, air pollution on the weather, temperature, wind speed in relation to the target urban congestion area Complex data such as two-dimensional space-time images, speed, lanes, and two-dimensional space-time images of road environments can be generated. According to an embodiment of the present invention, the composite data may be referred to as data. According to an embodiment of the present invention, the customized composite data input unit may be referred to as a composite data input unit.

The deep learning based real-time traffic flow prediction unit 1200 may receive the complex data from the customized complex data input unit 1100 and predict the traffic flow based on the deep learning. The deep learning based real-time traffic flow prediction unit 1200 according to an embodiment of the present invention includes a deep learning based extended-channel image labeling device 1210, a deep learning based extended-channel image CNN learning unit 1220, and a congested region CNN. -An LSTM module 1230, and / or a traffic congestion index prediction unit 1240 in a downtown congestion area. Here, CNN refers to the method of Convolution Neutral Network. Specifically, when the input network has various structures, it refers to a data processing method for convolving the input matrix-type data by applying a convolution filter to aggregate and represent it as one data. According to an embodiment of the present invention, the deep learning based real-time traffic flow predictor 1200 may be referred to as a traffic flow predictor.

The deep learning based extended-channel image labeling device 1210 may process extended-channel labeling on the composite data received from the customized composite data input unit 1100 based on deep learning. Here, labeling refers to a process of labeling an image of a spatiotemporal image, which is complex data, before the deep learning extension-channel CNN learning process is performed. The composite data is composed of a plurality of images that change with time. Accordingly, in order to process a plurality of time series images, labeling of images is required, and through image labeling, complex data may be labeled as congestion, displeasure, congestion, etc., which are criteria for determining traffic congestion. The traffic flow prediction unit may predict traffic congestion using the image data and the labeling information of the image. Image data and labeling information are values used for traffic congestion prediction and are always processed in pairs. Therefore, the image data and labeling information may be regarded as one data from the perspective of the prediction unit.

According to an embodiment of the present invention, the deep learning based extension-channel image labeling device 1210 may be referred to as an image labeling device or a labeling device.

The labeling device 1210 may generate a set of channel data with independent characteristics, which will be described in detail later with reference to FIG. 3. The labeling device 1210 includes two-dimensional image data of the traffic state of the target area as illustrated in FIG. 2 and two-dimensional image data of various factors that may affect the traffic state of the target area as illustrated in FIG. 3. Can be processed.

The labeling device 1210 may provide a plurality of channel data for traffic flow prediction to the deep learning based extended-channel image CNN learner 1220.

The labeling device 1210 may be configured to generate a plurality of channels of 2D image data for the deep learning-based extended-channel image CNN learner 1220 to generate more accurate results. It can be provided as a first data set and a second data set for testing. For example, if the labeling device 1210 provides a ratio of 8: 2 for the first data set for training and the second data set for the test, the CNN learner 1220 uses the first data set to provide a neural network. The data may be learned, and the algorithm may be applied to the second data set after the algorithm is optimized for the neural network weight. As such, the labeling apparatus 1210 may provide the CNN learner 1220 with 2D image data including a plurality of channels so as to increase the accuracy of data execution performed by the CNN learner 1220.

As such, the labeling device 1210 may provide a training data set, thereby optimizing a weight value required for a calculation later, and thereby providing data so that neural network learning is stabilized. The deep learning based extended-channel image labeling apparatus 1210 according to an embodiment of the present invention may be referred to as a labeling apparatus or a labeling controller.

The deep learning-based extended-channel image CNN learner 1220 may receive the labeled image data from the labeling device 1210 and perform CNN processing on the composite data which is a two-dimensional space-time image. Specific CNN processing will be described in detail with reference to FIGS. 4 to 5. According to an embodiment of the present invention, the deep learning-based extended-channel image CNN learner 1220 may be referred to as a CNN learner.

The congestion area CNN-LSTM module 1230 receives data from the deep learning based extended-channel image labeling device 1210 and the deep learning based extended-channel image CNN learner 1220. Herein, the data refers to images CNN processed by the deep learning extension-channel image CNN learner after the deep learning based extension-channel labeling process is completed. Detailed operations of the congestion zone CNN-LSTM module 1230 will be described in detail with reference to FIG. 6. The congestion zone CNN-LSTM module 1230 according to an embodiment of the present invention may be referred to as a congestion zone CNN-LSTM controller, or an LSTM controller.

The urban congestion area traffic congestion index prediction unit 1240 may predict a traffic congestion index of a target area (congestion area) by receiving data for traffic congestion prediction from the congestion area CNN-LSTM module 1230. A detailed operation of the traffic congestion index predictor 1240 will be described in detail with reference to FIG. 7.

The traffic congestion area congestion time signal controller 1300 may include a multi-agent enhanced learning DQN based traffic signal controller 1310 and / or a traffic congestion prediction based AI traffic signal interval remote controller 1320.

The multi-agent reinforcement learning DQN-based traffic signal controller 1310 may control the traffic signal system to control urban traffic congestion by using the value predicted by the traffic congestion index predictor 1240. By applying artificial intelligence reinforcement learning, traffic signals in congested areas can be controlled through traffic congestion index prediction values.

The traffic congestion prediction-based AI traffic signal interval remote controller 1320 may remotely control the traffic signal system by reflecting the predicted traffic signal cycle on the actual intersection traffic signal.

The traffic signal control apparatus according to an embodiment of the present invention may predict traffic congestion with respect to the traffic congestion time and space as described above, and control the traffic signal by AI reinforcement learning or remotely according to the predicted result. Furthermore, an embodiment of the present invention can complementarily use the existing traffic signal system for time and space without traffic congestion.

The detailed operation of the traffic congestion area congestion time signal controller 1300 will be described in detail with reference to FIG. 8.

As described above, the traffic signal control apparatus according to an embodiment of the present invention predicts traffic congestion based on deep learning technology, and controls a signal for traffic congestion prediction. According to one embodiment, the traffic signal control device may be referred to as Deep-Traffic Congestion (Deep-TraC).

2 illustrates an example of generating input data according to an embodiment of the present invention. In this example, the departments, malls, and densely populated areas are illustrated as the target areas for modeling the input data as the sections in which the traffic situation information changes or the traffic congestion is used to predict the traffic situation information. FIG. 2 illustrates composite data in the form of a space-time image described with reference to FIG. 1.

The figure shows a 2X2 intersection as an example of a road in this area, with the left side of the figure in the order of intersections, directions, and intersections, and the right side of the figure in the order of the intersections (space) and the speed on the intersection (v) over time. Is a color or contrast diagram.

On the left side of the drawing, N, S, E, and W represent north, south, east, and west directions, respectively, and 1, 2, 3, 4, ~, 7, and 8 represent road numbers along the direction of the round trip road, respectively. .

For example, in the figure, 1 indicates a first road in the west (W) to east (E) direction, 2 indicates a first road in the east (E) to west (W) direction, and 3 indicates a west ( W) may represent a second road in the east (E) direction, and number 4 may represent a second road in the east (E) to west (W) direction. Similarly, number 5 is the third road from south (S) to north (N), 6 is the third road from north (N) to south (S), and number 7 is the fourth from south (S) to north (N) The road 8 may indicate a fourth road from the north (N) to the south (S) direction.

According to an embodiment of the present invention, the intersection may be converted into two-dimensional space-time image data to provide traffic state information. To this end, an embodiment may use the converted value of the traffic state or traffic information to the section speed using the GPS data of the probe vehicle on the corresponding road.

Then, as shown in the right side of the drawing, in the target area, the horizontal axis represents a number according to the direction of the road in which the vehicle moves, and section communication information can be obtained as the speed of the vehicle for eight roads according to each number. For example, if 10 velocity data are obtained for each section (horizontal axis), the size of the horizontal pixel of the 2D image data may be 8 (the number of roads) * 10 (data of the velocity level for each road). have.

The two-dimensional image data according to an embodiment of the present invention has been described with the vertical as the time axis, and the data on the time axis can also be divided according to a predetermined time. For example, you may have collected vehicle speed data on the road in five minutes, and then divide it back into a fixed time unit of the day.

For example, suppose you divide a 24-hour time zone into six units of four-hour intervals. For example, 0 to 04 o'clock in the morning, 04 to 08 o'clock in the morning, 08 to 12 am, 12 to 16 o'clock in the afternoon, 16:00 to 20 o'clock in the evening, and 20 to 24 o'clock in the evening 6 By dividing the data into units, and collecting the traffic state information for each time unit by 5 minutes, the vertical size of the 2D image data may be 4 hours * 12 (5 minutes data per hour).

In this way, the 2D image data may be expressed as a 2D image of 80 * 60 pixels. The number of data for the neural network for predicting and analyzing traffic information is 6 time units (4 hours) * 365 days, 2190 data can be secured. If this data has been obtained for three years, 6570 data can be represented with 3 * 2190.

Of course, the number of data, the data generation interval, and the amount of data of the 2D image data generated at a specific intersection may be set differently according to a user or an embodiment as an embodiment.

As described above, the embodiment of the present invention may generate two-dimensional image data for predicting traffic state information.

FIG. 3 is a diagram illustrating an example of generating, as various channel data, factors that may affect traffic conditions.

According to an embodiment of the present invention, data of various factors that may affect traffic conditions may be generated as channel data to perform deep channel based deep learning using complex data. That is, a data set having a plurality of independent characteristics that may affect traffic conditions may be treated as an expansion channel, and thus traffic state information may be obtained by considering this in combination.

In the embodiment of this figure, the two-dimensional image data of the traffic speed may be one channel data as the traffic state information described in the above drawings. This example is the remaining channel data in addition to the two-dimensional image data (1) of the traffic speed. 5), the temperature of the target intersection (6), the wind speed on the target intersection (7), the first fine dust (PM 10) on the target intersection (8), the second fine dust on the target intersection (PM 2.5) (9), the target Ozone concentration 10 on the intersection, carbon dioxide concentration 11 on the target intersection, carbon monoxide concentration 12 on the target intersection, and sulfur dioxide concentration 13 on the target intersection can be generated as the respective channel data. Here, the target means an object for controlling traffic signals.

In addition, various factors that may affect traffic conditions may be generated as channel data. In this example, an example of generating 13 channel data is described. Accordingly, the two-dimensional image data of the traffic speed illustrated in FIG. 1 is the basis among the channel data illustrated in this example, and the remaining channel data may be changed. And although the same channel data is used as in this example, some channel data different from the channel data of (1) may be channel data having very little change in space and time.

As described above, an embodiment of the present invention can generate data having various independent characteristics regardless of various situations, and can accurately predict traffic state information of a target region by analyzing the correlation of the channel data.

4 is a diagram illustrating an example of a deep learning based extended channel image CNN learning unit according to an embodiment of the present invention. 4 to 5, a detailed operation of the deep learning based extended-channel image CNN learner 1220 illustrated in FIG. 1 will be described. As described above with reference to FIGS. 1 to 3, complex data represented by an image related to fine dust, weather information, holiday information, local environment, time information, and the like may be input to the CNN learning unit 1220. A detailed operation of the CNN learner 1220 of FIG. 4 will be described with reference to FIG. 5.

5 is a diagram illustrating a configuration of a traffic signal control apparatus according to an embodiment of the present invention.

FIG. 5 illustrates one configuration of the CNN learning unit 1220 described with reference to FIG. 4. The CNN learning unit 1220 according to an embodiment of the present invention may be referred to as an extension-channel feature extraction apparatus.

An embodiment of the traffic signal control apparatus according to the present invention may include a labeling input unit 5110, an extended channel characteristic extractor 5000, an extended channel association 5140, and an extended channel output unit 5150. The extended channel feature extractor 5000 may include a convolution module 5120 and a pooling module 5130.

The labeling input unit 5110 corresponds to the labeling apparatus of FIG. 1, and the labeling apparatus and the CNN learning unit of FIG. 1 operate in association with each other. ) Together. As described above, the labeling input unit 5110 may generate a set of channel data having independent characteristics as described with reference to FIG. 3. The labeling input unit 100 is two-dimensional image data of the traffic state of the target region as illustrated in FIG. 2 and two-dimensional image data of various factors that may affect the traffic state of the target region as illustrated in FIG. 3. Can be generated.

The labeling input unit 5110 may provide a plurality of channel data for the traffic signal control and prediction of the target region to the extended channel characteristic extractor 5000. The labeling input unit 5110 may provide a plurality of channels with two-dimensional image data so that the extended channel feature extractor 5000 may generate more accurate results, and a first data set for training the extended channel feature extractor 5000. Can be provided as a second data set for testing. For example, if the labeling input unit 5110 provides the first data set for training and the second data set for testing at a ratio of 8: 2, the extended channel characteristic extractor 5000 may use the first data set. The neural network may be used to learn the corresponding data, and the algorithm may be applied to the second data set after an algorithm for optimizing the weight of the neural network. As such, the labeling input unit 5110 may provide two-dimensional image data including a plurality of channels to the extended channel feature extractor 5000 so as to increase the accuracy of data performed by the extended channel feature extractor 5000. .

As such, the labeling input unit 5110 may provide a training data set, thereby optimizing a weight value required for a calculation later, and may provide data so that neural network learning is stabilized.

The extended channel feature extractor 5000 may perform CNN (Convolution Neutral Network) processing based on the 2D image data. The specific method of CNN processing is as follows.

The extended channel feature extractor 5000 may include a convolution module 5120 for convolving two-dimensional image data of each channel according to the CNN method and a pooling module 5130 for pooling and filtering them. Can be. The extended channel characteristic extractor 5000 may further include a ReLu filter module (not shown) that filters the data calculated by the pooling module 5130 using a ReLu filter.

If the extended channel characteristic extractor 5000 receives the two-dimensional image data for each channel illustrated in FIG. 3, the convolution module 5120, the pooling module 5130, and the ReLu filter module are provided for each of the 13 extended channels. You can proceed independently. The extended channel characteristic extractor 5000 may identify time and space in which the time, space, characteristic data of the time and space affecting traffic conditions, weather, fine dust, and greenhouse gases affect the target area.

The extension channel association unit 5140 performs a lower neural network of the CNN technique to connect the data provided by the extension channel feature extractor 5000 for each fully-connected layer. If data per channel or layer is related, it can be processed by associating it with softmax active module. The extension channel association unit 300 may calculate the associated characteristic information of the data from the N data set through the fully-connected layer in which N data channels or layers are related. The softmax active module can be a vector matrix or the like of associating N data layers.

The extension channel association unit 5140 may output traffic state information according to the associated characteristic information.

According to an embodiment of the present invention, a three-dimensional deep learning-based CNN may be performed using a three-dimensional filter, but as described above, when the factors affecting traffic conditions have characteristics that do not change in a target region, three-dimensional Instead of using filters, two-dimensional filters can be used to speed up processing while maintaining accuracy.

The expansion channel output unit 5150 may visualize and provide traffic state information output from the expansion channel association unit 5140 to the user according to a predetermined principle. This embodiment discloses an example in which the extended channel output unit 5150 provides traffic conditions of a target area in four states of congestion, congestion, normal, and smooth. An embodiment of the present invention is not limited thereto, and may provide a traffic state in more detail by using associated layer data.

Embodiments of the present invention can explain the effects of characteristic factors such as weather on traffic conditions, and explain the effects of fine dust or air pollution on traffic congestion, and also apply to the complex relationships of the factors. Repetitive or non-repetitive traffic status information can be provided in the area.

FIG. 6 is a diagram illustrating a configuration of a traffic signal control device long-short memory (LSTM) learning unit according to an embodiment of the present invention. Referring to FIG. 6, a data processing process of the deep learning based short and long term memory LSTM learning unit will be described. In other words, the deep learning based real-time traffic flow prediction unit 1200 labels the complex data by the deep learning based extended-channel image labeling device 1210, and then the deep learning based extended-channel image CNN learner 12220. The CNN processing of the data labeled by the CNN may be performed by the congestion region CNN-LSTM module 1230 to perform LSTM processing. That is, FIG. 6 illustrates a process of combining data processing with CNN processing and LSTM processing.

The LSTM learning unit according to an embodiment of the present invention enables real-time prediction of GPS-based traffic data on a target road for a traffic congestion index. Here, the LSTM learning unit may convert a target section into a two-dimensional space-time grid and convert the velocity into a two-dimensional space-time image having a function as a function. At this time, there may be one channel. 2 to 3, it has been described that the two-dimensional spatiotemporal image data may be a value such as speed using GPS vehicle data. An embodiment of the present invention may analyze a pattern in the image data using time series image data.

Deep learning based short and long term memory LSTM technology can be used. Specifically, LSTM technology is a method of extracting a pattern in time series (time) data. LSTM technology can be compared with Recurrent Neural Network (RNN). Compared to the RNN method, LSTM technology uses short and long term memory, so more accurate results can be obtained. Short- and long-term memory means not keeping the past past pattern of time series data (text or picture) for a very long time, nor relying on short-term memory like RNN. The time series data may include hidden patterns, and the LSTM technology extracts such patterns into short and long term memory.

According to an embodiment of the present invention, time series-based GPS vehicle data is input to the LSTM learning unit as shown in FIG. 6. Input values composed of time series data can be composite data of multiple images, and these input values are output values composed of transformed data after LSTM technology processing. Here, the output value has a characteristic of a prediction value for real time prediction. 1 and 6, a plurality of images of the deep learning extension-channel image CNN learner for which deep learning based extension-channel labeling is completed are input to the LSTM learner. That is, the plurality of images are labeled as congestion, desired or congestion through the labeling device, and the labeled images are input to the LSTM learning unit in a pair. Thereafter, the LSTM learning unit may analyze the pattern information in relation to the traffic information for predicting the traffic congestion in the image through pair information consisting of the image and the labeling result.

7 illustrates an example of deep learning based real-time traffic congestion prediction and remote traffic signal control according to an embodiment of the present invention. Referring to FIG. 7, a detailed operation of the urban congestion index predictor 1240 shown in FIG. 1 will be described.

The deep learning based traffic congestion prediction unit according to an embodiment of the present invention uses a hybrid prediction model in which a traffic congestion index image and data of a customized composite data input unit are mixed. As shown in FIG. 7, the present invention can predict real-time traffic congestion target section and / or traffic congestion time zone using data processed using the CNN-LSTM neural network of the CNN-LSTM module described above. In other words, both spatial and temporal data can be predicted. The traffic congestion prediction unit according to an embodiment of the present invention may predict the traffic congestion for each fully-connected layer in a local unit (for example, 2x2, 3x3, 4x4 intersection) by receiving the derived local target section. Furthermore, the present invention can control the AI-based remote traffic signal at the predicted traffic congestion time zone and the target section road intersection.

8 illustrates an example of a deep Q-network (DQN) traffic signal cycle control unit based on Reinforcement Learning according to an embodiment of the present invention. The traffic signal period control unit according to an embodiment of the present invention may use a traffic signal control algorithm. Specifically, as shown in FIG. 8, an embodiment of the present invention may apply reinforcement learning based multi-agent deep-Q neural network (DQN) to an intersection. Deep-Q neural network technology (DQN) refers to a method of controlling a multi-agent (eg, the number of intersections) traffic signal based on the image.

Referring to FIG. 8, when there is a 2x2 intersection according to an embodiment of the present invention, the traffic light may represent red or green. In FIG. 8, red is indicated by reference numeral 800 and green is indicated by reference numeral 810. Here, a box marked with two spaces or two pixels represents a vehicle, and vehicles are moved one space per second. The vehicle moves at the intersection, and the change in position of the vehicle is input to the DQN traffic signal cycle controller. The position change of the vehicle is input to the DQN traffic signal period control unit as an input value, and the reward value is set to minimize the waiting time of the vehicles. According to an embodiment of the present invention, the traffic signal may be controlled using the input value and the compensation value. An embodiment of the present invention may acquire a time zone in which traffic congestion is expected based on deep learning, and control traffic signals using artificial intelligence in terms of time and space based on the expected time zone. In addition, one embodiment of the present invention, when there is no traffic congestion can be used by complementary to the existing traffic signal system by using a mixture of methods using the existing traffic signal together.

9 is a view showing an example of a traffic signal control method according to an embodiment of the present invention. Referring to Figure 9 describes a traffic signal control method according to an embodiment of the present invention.

Generate complex data for traffic congestion prediction (S100). A detailed configuration of the composite data and a method of generating the composite data are as described with reference to FIG. 1.

Traffic congestion is predicted using the composite data (S200). Specific steps for estimating traffic congestion refer to steps S300 and S400 below.

Here, an image labeling process of labeling the image data of the composite data is performed (S300). Details of the image labeling are as described with reference to FIG. 1.

Further, an image CNN process applying a CNN to the labeled image (S400). Details of the image CNN are as described with reference to FIG. 4.

A module, unit, or block according to embodiments of the present invention may be a processor / hardware that executes successive procedures stored in a memory (or storage unit). Each step or method described in the above embodiments may be performed by hardware / processors. In addition, the methods proposed by the present invention can be executed as code. This code can be written to a processor readable storage medium and thus read by a processor provided by an apparatus according to embodiments of the present invention.

For convenience of description, each drawing is divided and described, but it is also possible to design a new embodiment by merging the embodiments described in each drawing. And, according to the needs of those skilled in the art, it is also within the scope of the present invention to design a computer-readable recording medium having a program recorded thereon for executing the embodiments described above.

Apparatus and method according to the present invention is not limited to the configuration and method of the embodiments described as described above, the above-described embodiments may be selectively all or part of each embodiment so that various modifications can be made It may be configured in combination.

On the other hand, the image processing method of the present invention can be implemented as a processor-readable code on a processor-readable recording medium provided in the network device. The processor-readable recording medium includes all kinds of recording devices that store data that can be read by the processor. Examples of the processor-readable recording medium include ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, and the like, and the processor-readable recording medium is distributed in a networked computer system. In a distributed fashion, processor-readable code can be stored and executed.

In addition, although the preferred embodiment of the present invention has been shown and described above, the present invention is not limited to the above-described specific embodiment, the technical field to which the invention belongs without departing from the spirit of the invention claimed in the claims. Of course, various modifications can be made by those skilled in the art, and these modifications should not be individually understood from the technical spirit or the prospect of the present invention.

In addition, in this specification, both the object invention and the method invention are described, and description of both invention can be supplementally applied as needed.

Although embodiments of the present invention have been described above with reference to the accompanying drawings, those skilled in the art to which the present invention pertains may implement the present invention in other specific forms without changing the technical spirit or essential features thereof. I can understand that. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive.

Claims (13)

Complex data input unit for generating complex data for traffic congestion prediction,
The composite data is spatiotemporal image data based on atmospheric information, local information, and temporal information; And
A traffic flow prediction unit for predicting traffic congestion using the complex data;
The traffic flow prediction unit
An image labeling controller for labeling the image data of the composite data;
Including an image CNN learning unit applying a CNN (Convolution Neutral Network) to the labeled image,
Traffic signal control device. The method of claim 1,
The air information relates to fine dust or air pollution,
The area information relates to the lane or road environment of the road,
The time information indicates a time zone for dividing the time of day,
Traffic signal control device. The method of claim 1,
The image CNN learning unit applies an image change technology based on two-dimensional space-time,
Traffic signal control device. The method of claim 1,
The traffic flow prediction unit
Further comprising a CNN-LSTM (Long-Short Term Memory) control unit for applying LSTM to the labeling and CNN applied image,
The CNN-LSTM control unit receives GPS-based traffic data and converts the target section for the complex data into a two-dimensional space-time to convert it into a two-dimensional space-time image.
Traffic signal control device. The method of claim 4, wherein
The CNN-LSTM control unit
Processing the converted two-dimensional space-time image by deep learning based long-short memory LSTM (Long-Short Term Memory) method,
Traffic signal control device. The method of claim 1,
Further comprising a signal control unit for controlling the traffic signal using the congestion index data obtained through the labeling and the CNN applied image,
Traffic signal control device. Complex data generation step of generating complex data for traffic congestion prediction,
The composite data is spatiotemporal image data based on atmospheric information, local information, and temporal information; And
A traffic flow prediction step of predicting traffic congestion using the complex data;
The traffic flow prediction step
An image labeling step of labeling the image data of the composite data;
An image CNN step of applying a CNN to the labeled image,
Traffic signal control method. The method of claim 7, wherein
The air information relates to fine dust or air pollution,
The area information relates to the lane or road environment of the road,
The time information indicates a time zone for dividing the time of day,
Traffic signal control method. The method of claim 7, wherein
The image CNN step applies an image change technology based on two-dimensional space-time,
Traffic signal control method. The method of claim 7, wherein
The traffic flow prediction step
The method further includes a CNN-LSTM step of applying LSTM to the labeling and CNN-applied images,
In the CNN-LSTM step, receiving GPS-based traffic data and converting a target section for the complex data into a two-dimensional space-time, converting it into a two-dimensional space-time image,
Traffic signal control method. The method of claim 10,
The CNN-LSTM step
Processing the converted two-dimensional space-time image by deep learning based long-short memory LSTM (Long-Short Term Memory) method,
Traffic signal control method. The method of claim 7, wherein
Signal control step of controlling the traffic signal using the congestion index data obtained through the labeling and the CNN applied image,
Traffic signal control method.



delete

Images