Vehicle speed prediction device and prediction method using the device

Abstract

This invention relates to a vehicle speed prediction device and a prediction method using the same device. The vehicle speed prediction device includes a data preprocessor and a deep learning modeling device. The data preprocessor performs preprocessing on correct answer data and input data for deep learning. The deep learning modeling device is configured to learn the relationship between lane speed for each exit lane based on the representative speed of each road segment, the number of detectors collected, and the collection time.

Description

[0001] Cross Reference to Related Applications

[0002] This application claims the benefit of Korean Patent Application No. 10-2020-0170853, filed December 8, 2020, in the Korean Intellectual Property Office, the disclosure of which is incorporated herein in its entirety by reference. TECHNICAL FIELD

[0003] The present application relates to a vehicle speed prediction device and a prediction method using the same, and more particularly, to a vehicle speed prediction device for each lane based on deep learning and a vehicle speed prediction method using the same. BACKGROUND

[0004] Generally, traffic speed can be predicted in a manner of predicting a speed value after 'b' minutes in the future by using speed data of a corresponding point (a spatial range) for 'a' minutes in the past (a time range). In the present specification, traffic speed refers to a speed at which a vehicle travels on a road, and can be collected by using an average speed at which a probe (a vehicle) passes a general link (a road section, which is a minimum unit of a road). That is, traffic speed can be calculated by a travel time and a travel distance of all probes passing a link.

[0005] On the other hand, when a continuous general link in which congestion is sustained is referred to as a link for each lane, since the length of the link is long, probes that pass only a part of the link can frequently occur, and it is difficult to collect speed for each lane in units of the link for each lane. In other words, the speed for each lane in which congestion is sustained has a lower probe collection rate than the general link speed, and thus reliability is reduced.

[0006] When the link speed for each lane is not generated as such, the speed is predicted by using a pattern speed collected in the same time period in the past. However, since the past pattern speed cannot reflect a current traffic situation, accuracy is low due to low real-time performance. This can result in inaccurate arrival time prediction and reduce user satisfaction.

[0007] Accordingly, there is a need to improve the accuracy of predicting a speed for each lane and predicting an arrival time that reflects an actual traffic situation when a speed for each lane is not collected. SUMMARY

[0008] One aspect of the present application provides a vehicle speed prediction device for each lane and a vehicle speed prediction method, which can effectively predict an accurate vehicle speed for each lane that reflects real-time traffic situation.

[0009] An aspect of the present application provides a vehicle speed prediction device and a vehicle speed prediction method for each lane, which applies a deep learning model configured to learn a speed relationship for each exit lane by using a real-time speed of a lane speed that is not classified as an exit lane.

[0010] An aspect of the present application provides a vehicle speed prediction device and a vehicle speed prediction method for each lane, which performs preprocessing of converting different types of data representing traffic conditions into data suitable for a deep learning model.

[0011] The technical problems to be solved by the present application are not limited to the above-mentioned problems, and any other technical problems not mentioned herein will be clearly understood by those skilled in the art through the following description.

[0012] According to an aspect of the present application, a vehicle speed prediction device includes a data preprocessor that performs preprocessing of correct answer data and input data for deep learning, and a deep learning modeling device configured to learn a relationship for a lane speed of each exit lane based on a representative speed of each link, a number of collected probes, and a collection time.

[0013] In one embodiment, the deep learning modeling device can be configured to learn using a fully connected model that can include one or more input layers, one or more hidden layers, and one or more output layers, a vector associated with a representative speed of each link, a number of collected probes, and a collection time can be input to the one or more input layers, and the one or more output layers can output a speed of a probe traveling in each lane of a link at the same point in time.

[0014] In one example, the one or more hidden layers can consist of predetermined units obtained through an optimization process, and the deep learning modeling device can apply a normalization process for adjusting a scale of values constituting the units to each layer.

[0015] In one example, the deep learning modeling device can apply a dropout technique for reducing some information of the units to "0" to each layer at a predetermined rate.

[0016] In one example, the link of each lane representing an interval of each lane to be learned can include a queue link of no direction, a start link before an exit lane, and a plurality of exit links (termination link 1, termination link 2,...), the representative speed of each link can include a speed of the queue link and a speed of the start link, and the lane speed of each exit lane can include a speed of each exit link.

[0017] In one example, the data preprocessor can include a correct answer data preprocessor that converts probe trajectory data collected at an exit time, obtained by traveling through all road segments of each lane, into data based on an entry time, and performs harmonic mean processing on the converted data in a preset reference time unit, and an input data preprocessor that performs harmonic mean processing on representative speeds of each road segment in a preset reference time unit.

[0018] In one example, the correct answer data preprocessor can obtain representative speeds of each lane based on a sum of travel distances of all probes traveling in a time period and a sum of travel times of the probes in the time period, according to a time period of a day divided by a preset reference time unit.

[0019] In one example, the input data preprocessor can perform harmonic mean processing of representative speeds of each road segment in a time period, and can obtain representative speeds of each road segment, which are harmonic mean-processed, in a time period.

[0020] In one example, the data preprocessor can perform minimum-maximum normalization processing on representative speeds of each road segment, a number of collected probes, a collection time, and a proportion of lane speeds of each exit lane.

[0021] In one example, when data about representative speeds of each road segment, a number of collected probes, and a collection time is missing, the data preprocessor can use a preset value as the missing data, and the missing data about representative speeds of each road segment can include a first specific value, the missing data about a number of collected probes can include a second specific value, and the missing data about a collection time can include index information of a preset reference time.

[0022] According to an aspect of the present disclosure, a vehicle speed prediction method includes a data pre-processing step of pre-processing correct answer data and input data for deep learning, and a deep learning modeling step of learning a relationship of lane speeds for each exit lane based on representative speeds of each road segment, a number of collected probes, and a collection time. BRIEF DESCRIPTION OF DRAWINGS

[0023] The above and other objects, features and advantages of the present disclosure will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings, in which:

[0024] Figure 1 is a control block diagram of a vehicle speed prediction apparatus according to an embodiment of the present disclosure;

[0025] Figure 2A and Figure 2Bis a schematic diagram showing deep learning modeling according to embodiments of the present application, wherein, Figure 2A shows a deep learning structure, Figure 2B shows each lane prediction problem for deep learning;

[0026] Figure 3 is a schematic diagram showing a road segment for each lane according to embodiments of the present application;

[0027] Figure 4 is a schematic diagram showing a target area for verifying deep learning modeling according to embodiments of the present application;

[0028] Figure 5 is a table graph comparing errors between deep learning modeling according to embodiments of the present application and an existing pattern speed replacement method, wherein, ① Gwangbok-dong intersection (turning in three directions), ② Sinchon station intersection (extreme congestion section), ③ Dongbu Steel entrance intersection (many queue sections), ④ Sapaeng-Daero Pampu IC entrance section (highway entrance);

[0029] Figure 6 is a table graph comparing collection rates between deep learning modeling according to embodiments of the present application and an existing pattern speed replacement method, wherein, ① Gwangbok-dong intersection (turning in three directions), ② Sinchon station intersection (extreme congestion section), ③ Dongbu Steel entrance intersection (many queue sections), ④ Sapaeng-Daero Pampu IC entrance section (highway entrance);

[0030] Figure 7A and Figure 7B is a table graph comparing travel prediction times between deep learning modeling according to embodiments of the present application and an existing pattern speed replacement method; and

[0031] Figure 8 is a control flow diagram of a vehicle speed prediction method according to embodiments of the present application. DETAILED DESCRIPTION

[0032] It should be understood that the term "vehicle" or "vehicle's" or other similar terms used herein generally include motor vehicles, such as passenger vehicles including sport utility vehicles (SUVs), buses, trucks, various commercial vehicles, boats including various boats and ships, aircraft, etc., and include hybrid vehicles, electric vehicles, plug-in hybrid electric vehicles, hydrogen-powered vehicles, and other alternative fuel vehicles (e.g., fuels derived from non-petroleum sources). As referred to herein, a hybrid vehicle is a vehicle having two or more power sources, such as a vehicle having both gasoline power and electric power.

[0033] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. As used herein the term "and / or" includes any and all combinations of one or more of the associated listed items. Throughout this specification, unless explicitly described to the contrary, the word "comprise" and variations such as "comprises" or "comprising" will be understood to imply the inclusion of stated elements but not the exclusion of any other elements. In addition, the terms "unit", "device", "component", and "module" described in the specification mean a unit for processing at least one function and operation, and can be implemented by hardware components or software components and combinations thereof.

[0034] Further, the control logic of the present application can be implemented as a non-volatile computer readable medium on a computer readable medium, which contains executable program instructions executed by a processor, a controller, or the like. Examples of the computer readable medium include, but are not limited to, ROM, RAM, compact disc (CD)-ROM, magnetic tapes, floppy disks, flash drives, smart cards, and optical data storage devices. The computer readable medium can also be distributed over network-coupled computer systems so that the computer readable medium is stored and executed in a distributed fashion, e.g., by a remote information processing server or a controller area network (CAN).

[0035] Hereinafter, some embodiments of the present application will be described in detail through the accompanying drawings. When adding reference numerals to the elements of each drawing, it should be noted that the same elements are allocated the same reference numerals as far as possible even if they are indicated on different drawings. Further, in describing embodiments of the present application, if it is determined that the detailed description of related known configurations or functions will interfere with the understanding of embodiments of the present application, a detailed description thereof will be omitted.

[0036] In describing components of embodiments of the present application, terms such as first, second, A, B, (a), (b), etc. can be used. These terms are used only to distinguish one component from another, and the terms do not limit the nature, sequence, or order of the components. Also, unless otherwise defined, all terms used herein, including technical or scientific terms or terminology, have the same meaning as those understood by one of ordinary skill in the art to which the present application pertains. Terms such as those defined in a generally used dictionary should be interpreted as having a meaning that is consistent with its meaning in the context of the relevant art, and should not be interpreted in an idealized or overly formal sense unless expressly so defined in the present application.

[0037] Hereinafter, a vehicle speed prediction device according to an embodiment of the present application will be described with reference to the accompanying drawings. Figures 1 to 8 Embodiments of the present application are described in detail.

[0038] The present application relates to a deep learning model that, when a speed is missing in a process of providing a speed of a section of a specified lane, predicts an accurate speed of each lane by learning a relationship of a real-time speed, a number of collections, and a time provided by a general section having a high collection rate. Hereinafter, a speed prediction device for performing such deep learning modeling and a method of using the same will be described with reference to the accompanying drawings.

[0039] Figure 1 A control block diagram of a vehicle speed prediction device according to an embodiment of the present application.

[0040] As Figure 1 indicated, the vehicle speed prediction device can include a data pre-processor 100 that pre-processes correct answer data and input data for deep learning, and a deep learning modeling device 200 that learns a relationship of a lane speed of each exit lane based on a representative speed of each section, a number of collected probes, and a collection time.

[0041] The vehicle speed prediction device can include a deep neural network having one or more hidden layers, i.e., an algorithm capable of performing deep learning modeling, as Figures 2A to 2B indicated, and the vehicle speed prediction device can be implemented with a device capable of performing the algorithm.

[0042] Specifically, the vehicle speed prediction device according to the present embodiment attempts to solve a prediction problem for each lane by utilizing deep learning that has recently shown excellent performance in various fields (see Figure 2A ). Deep learning can learn a non-linear relationship between input data and output data using a deep neural network, and various types of data can be easily input at the same time using the deep neural network. Because the model can learn a relationship of each data by itself, deep learning can effectively obtain a prediction value desired by a user.

[0043] According to one example, the deep learning modeling device 200 can learn using a fully connected model, and the model can include one or more input layers, one or more hidden layers (e.g., 6 hidden layers), and one or more output layers.

[0044] In another aspect, Figure 3 is a schematic diagram showing a road segment of each lane according to an embodiment of the present application.

[0045] As shown, the road segment of each lane indicates an interval of each lane as a learning object, and the road segment of each lane can include a queue link having no directionality, a from link before an exit lane, and a plurality of exit links ("to link 1", "to link 2",...).

[0046] In one embodiment, a vehicle speed is predicted for each exit link ("to link 1", "to link 2",...) through deep learning modeling, and the vehicle speed including each road segment speed can be defined as a speed of each lane or a speed of each exit lane.

[0047] As Figure 2B shown, a representative speed (real-time representative speed) of each road segment, the number of collected probes (the number of collected samples), and the collection time can be input to the input layer, and the speed at which the probe travels on the road segment at the same time for each lane (i.e., the real-time speed of each lane) can be obtained from the output layer.

[0048] Specifically, the input layer is designed to have a dimension of (the number of queue links + the number of from links) x 3, and a vector regarding the representative speed, the number of collected probes, and the collection time of each road segment, for example, a vector [road segment speed, probe collection number, entry time] of each general road segment can be added to a row to form an input vector.

[0049] The road segment speed is based on the speed of the queue link and the from link, and the collection number can mean the number of probes entering the corresponding time slot.

[0050] In addition, according to one example, the hidden layer can consist of [64, 128, 128, 128, 64, 32] through an experimental optimization process.

[0051] The deep learning modeling device 200 can apply batch normalization (Batch Normalization) to each layer, adjust the scale of the values of each hidden layer so as not to lean to either side, and can apply a drop out technique to each layer, in which some information of a unit is reduced to "0" at a predetermined ratio during learning, so that learning does not concentrate on learning data. The predetermined ratio can be set to about 15%.

[0052] In addition, the deep learning modeling device 200 can design modeling to apply a widely used Relu function as an activation function, and can perform model optimization using an Adam optimizer as an optimization algorithm.

[0053] Accordingly, the deep learning modeling device 200 can simultaneously receive information of a representative speed of each of the queue link and the start link, the number of collected probes, and the collected time, and can predict the speed of each lane.

[0054] Because the speed of each lane is generated from the exit lane as described above, by utilizing the representative speed of each queue link which does not have directionality, it can be designed to learn the relationship between straight (main lane), right turn (exit lane 1), and left turn (exit lane 2). The deep learning modeling device 200 can learn the relationship between the speed of each lane, the number of collections, the entry time, and the speed of each lane.

[0055] On the other hand, the data preprocessor 100 according to the present embodiment can include a correct answer data preprocessor 120 that converts probe trajectory data collected at an exit time, obtained by traveling all links of each lane, into data based on an entry time, and performs harmonic mean processing (e.g., aggregation) on the converted data in a preset reference time unit, and an input data preprocessor 110 that performs harmonic mean processing on a representative speed of each link in a preset reference time unit.

[0056] The correct answer data preprocessor 120 can obtain a representative speed of each lane based on the sum of the travel distance of all probes traveling for a time period and the sum of the travel time according to the time period of one day divided by a preset reference time unit.

[0057] The correct answer data can be obtained using probe trajectory data collected by traveling all links of each lane, and such probe trajectory data can be collected as an actual travel distance and a required time corresponding to a link. Because all collected data is recorded based on an exit time, the correct answer data preprocessor 120 can convert the collected data into data corresponding to an entry time by subtracting the required time from the exit time.

[0058] For example, because 288 time periods are set by dividing 1440 minutes of a day by 5-minute increments, and based on the entry time, a representative speed of each exit link, i.e., a representative speed of each lane (for each turning direction), can be calculated by dividing the sum of travel distances of all probes traveling in one time period by the sum of travel times to obtain a harmonic mean. The calculated representative speed is used as correct answer data for the corresponding time period.

[0059] The input data pre-processor 110 can perform a harmonic mean process of a representative speed of each link in units of time periods, and can obtain a representative speed of each link after the harmonic mean process in units of time periods.

[0060] In one embodiment, input data can be obtained using real-time speed links of general links constituting a link of each lane. The real-time speed of a general link is a speed collected after correction by calculating in units of one minute with a median of probes traveling in a corresponding interval. The input data pre-processor 110 performs a harmonic mean process on the collected information in units of time periods corresponding to 5 minutes, and obtains a representative speed of each time period.

[0061] The data pre-processor 100 can perform minimum-maximum normalization on a representative speed of each link, a collected number of probes, a collection time, and a proportion of a lane speed of each exit lane.

[0062] Because proportions of three types of data, i.e., speed, number of samples, and entry time, input as input data to a learning layer are different, normalization processing is required. The data pre-processor 100 normalizes to a minimum-maximum with the same proportion using a maximum value and a minimum value of all speed data for input data / correct answer data, and can perform minimum-maximum normalization on a collected number by using a maximum value and a minimum value of all numbers of probes that have entered a corresponding time period. In addition, entry time information can be converted to an index of 288 time periods of a corresponding time, and minimum-maximum normalization processing can be performed.

[0063] Learning data can be configured based on input data and output data thus pre-processed, and the present embodiment optimizes a batch size to be 32 through an experiment.

[0064] All data sets are randomly mixed and verified as 5-fold cross-validation sets, and a maximum iteration (Epoch) value is selected to be 1000.

[0065] On the other hand, the data preprocessor 100 according to the present embodiment can utilize a preset value as missing data when data on the representative speed of each link, the number of collected probes, and the collection time is missing.

[0066] Learning data can be configured only when correct answer data (probe trajectory data) is collected, and learning data can be configured by processing a value of missing input data when correct answer data is collected, except for any input data (real-time speed of a general link). By compensating for missing data, more learning data can be secured, and the accuracy of learning can be improved.

[0067] The missing data associated with the representative speed of each link can be set to a first specific value.

[0068] As a specific example, when input data is missing, -1, -9999, and 0 are selected as missing speed values as a result of an experiment, and since the learning result of -1 has the highest learning accuracy, the missing data associated with the representative speed of each link can be set to -1 according to one example.

[0069] The data preprocessor 100 can configure input data by utilizing a general link speed value of -1, the number of probes collected of 0, and corresponding entry time period index information associated with missing input data.

[0070] Hereinafter, an experimental example for verifying the effectiveness of deep learning modeling according to the present embodiment will be described.

[0071] The collected speed data associated with the link probe of each lane and the real-time representative speed data of a general link are set to utilize data, four spatial restrictions of Seoul are set as a spatial range, and a time restriction from January 10, 2020 to March 31, 2020 is set as a time range. The number of samples that can be learned by 5-minute aggregation is 23616.

[0072] In order to verify modeling, all data is divided into five, into a training set and a test set, and the average of the accuracy rate of each test set is evaluated by utilizing five-fold cross-validation.

[0073] In addition, the learning time can be shortened by applying an early stop function in Epoch, in which the accuracy of a validation set is secured by utilizing a part of the training set as the validation set.

[0074] Figure 4 FIG. 1 is a schematic diagram illustrating a target area for verifying deep learning modeling according to an embodiment of the present invention.

[0075] As Figure 4As shown, in order to verify the prediction model, all 4 sections were selected: 1. Newshasan intersection (Gangnam-gu, Newshadong), 2. Gwangnamsan intersection (Gangdong-gu, Gwangsandong), 3. East Steel Inlet intersection (Guro-gu, Wollyudong), and 4. Shapin-Dalupodong IC Inlet section (Seocho-gu, Dalupodong).

[0076] Each section can include a special nature of the road, for example, the Newshasan intersection has a characteristic of a region where extreme congestion frequently occurs, the Gwangnamsan intersection has a characteristic of a queue section including three turns of straight, left, and right, the East Steel Inlet intersection has a characteristic of a queue section of many general road segments (11), and the Shapin-Dalupodong IC Inlet section has a characteristic that a highway inlet section can be set.

[0077] Figure 5 is a table comparing errors between the deep learning modeling according to the embodiment of the present application and the existing pattern speed replacement method, and when using deep learning modeling instead of the existing pattern speed replacement method (baseline), Figure 5 indicates the degree of improvement in the relative error (mean absolute percentage error; MAPE).

[0078] The existing pattern speed replacement method is a method of replacing the predicted speed with the pattern speed collected at the same time in the past, and has a limitation of low accuracy due to low real-time because it cannot reflect the current traffic situation.

[0079] As Figure 5 shown, in the case of deep learning modeling, it can be seen that the relative error (MAPE) is improved compared to the existing pattern replacement method. That is, as the relative error value decreases, the accuracy of the vehicle speed prediction increases.

[0080] In the case of ② Newshasan intersection, where the arrival time error frequently occurs due to a relatively low collection rate, it can be seen that both the straight and right turns have a relative error of about 20%. Also, in the case of ④ Shapin-Dalupodong IC Inlet section, which is an entrance lane of a highway, although the collection rate of the straight lane is high, it can be seen that the accuracy is improved by about 12%.

[0081] In addition, in the case of the right turn exit lane of ④ Shapin-Dalupodong IC, there is an error of 50.48% when using the existing pattern replacement method, but when using deep learning prediction, an error of 26.88% occurs, improving the error rate by about 46% or more.

[0082] By Figure 5 , it can be seen that the accuracy is improved by at least 8% and at most 50% when applying deep learning modeling compared to the existing pattern replacement method.

[0083] Figure 6 A table showing the collection rate between the deep learning modeling according to the embodiment of the present application and the existing pattern speed replacement method.

[0084] As Figure 6 shown, according to the existing (as-is (when data for each lane is utilized)) method, in the case of the ② Sinchon intersection, it shows a collection rate of 10% or less. In this case, for about 90% or more for which speed prediction data is not collected, the existing pattern speed with an error of 25% (24.08% and 23.87% in Figure 5 ) can be utilized for prediction of speed.

[0085] In contrast, when the prediction method utilizing deep learning is used, in about 75% or more of the total, a more accurate value can be predicted with an error of about 20% (19.95% and 20.01% in Figure 5 ).

[0086] In addition, in the case of the ④ Shihyeon-Daero Damo IC section and the ③ East Steel Inlet intersection having many queue sections, since the collection rate is about 99%, it can be seen that there is almost no utilization of the pattern speed.

[0087] In this way, when the prediction method utilizing deep learning is used, it is possible to improve the accuracy of speed prediction by supplementing the missing data, as well as to improve the data collection rate.

[0088] Figure 7A and Figure 7B is a table showing the travel prediction time between the deep learning modeling according to the embodiment of the present application and the existing pattern speed replacement method, and shows the results of experiments performed by a probe traveling at the Sinchon intersection at a specific time (about 18:00 on February 14, 2020).

[0089] Figure 7A shows the routes searched to the arrival point at 18:05 and 18:13, and Figure 7B the predicted speeds are shown in the table based on pattern and deep learning prediction.

[0090] When the speed is predicted based on the existing pattern, since the speed for each lane is missing when the route search is attempted at 18:05, the existing pattern speed of 7.12 km / h is utilized. Thereafter, at 18:13, the actual vehicle speed is 2.02 km / h, but when the existing pattern speed of 7.12 km / h is utilized for prediction, the prediction error is as high as 252.48%. In other words, the expected arrival time increases, resulting in a large arrival time error.

[0091] In contrast, in the case of using deep learning, when searched at 18:05, the speed of each lane was predicted to be 2.08 km / h, and the vehicle speed at 18:13 was 2.02 km / h, so it can be confirmed that the error rate was significantly improved (252.48% → 2.97%).

[0092] That is, because the existing pattern speed does not reflect real-time traffic conditions, the arrival time error mainly occurs in a single case. In contrast, when deep learning using real-time speed prediction is used, the arrival time can be more accurately predicted.

[0093] Figure 8 is a control flowchart of a vehicle speed prediction method according to an embodiment of the present invention.

[0094] The vehicle speed prediction method according to the present embodiment can refer to Figure 8 The summary is as follows.

[0095] First, the data preprocessor 100 of the vehicle speed prediction device can perform pre-processing on the correct answer data and the input data for deep learning (S810).

[0096] The data preprocessor 100 performs data preprocessing by scaling and normalizing the data for learning.

[0097] Specifically, the data preprocessor 100 can perform harmonic mean processing on the representative speed of each link in units of reference time, can convert the probe trajectory data collected at the exit time, which is obtained by traveling through all links of each lane, into data based on the entry time, and can perform harmonic mean processing on the converted data in a preset reference time unit.

[0098] The speed of each lane for vehicle speed prediction should be output based on the entry time of the link of each lane, but because the probe trajectory data used as the correct answer data is collected based on the exit time, it is necessary to convert the probe trajectory data based on the entry time. The data preprocessor 100 can generate the correct answer data by performing harmonic mean processing (aggregation) on the entry time in a preset reference time unit (for example, a reference 5-minute unit).

[0099] In addition, the data preprocessor 100 can generate the input data by performing harmonic mean processing (aggregation) on the representative speed of the general link constituting the link of each lane, which is collected in units of 1 minute, in units of 5 minutes.

[0100] According to one example, the data preprocessor 100 obtains the representative speed of each lane (i.e., the correct answer data) based on a time period divided by a reference time unit, the total distance traveled by all detectors traveling within a time period, and the total travel time. It can perform harmonic averaging on the representative speed of each road segment on a time period basis, and can obtain the harmonic averaged representative speed of each road segment (i.e., the input data) on a time period basis.

[0101] In addition, the data preprocessor 100 can perform preprocessing (S820) to normalize the correct answer data and input data and insert missing data.

[0102] Because the proportions of the three types of data (i.e., speed, sample size, and entry time) are different, normalization is required, and the data preprocessor 100 can perform minimum-maximum normalization on the representative speed of each road segment, the number of detectors collected, the collection time, and the lane speed of each exit lane (i.e., the ratio of input data to correct answer data).

[0103] The data preprocessor 100 can apply minimum-maximum normalization adjusted to the same proportion by utilizing the maximum and minimum values ​​of velocity data about input data and correct answer data, and can apply minimum-maximum normalization to the number of detectors representing the number of detectors entering the corresponding time period by utilizing the maximum and minimum values ​​collected by all detectors, and can similarly apply minimum-maximum normalization to the entry time information after being converted to an index of 288 time periods corresponding to the time.

[0104] On the other hand, in the data preprocessing step, when data associated with the representative speed of each road segment, the number of detectors collected, and the collection time are missing, preset values ​​can be used as the missing data.

[0105] For example, missing data regarding the representative speed of each road segment can be provided using a first specific value (e.g., -1), missing data regarding the number of detectors collected can be provided using a second specific value (e.g., 0), and missing data regarding the collection time can be provided using index information of a preset reference time. Because the learning data includes such missing data, learning accuracy can be improved.

[0106] When such data is preprocessed and learning is performed, deep learning modeling can be performed to learn the relationship between lane speeds for each exit lane based on the representative speed of each road segment, the number of detectors collected, and the collection time (S830).

[0107] Deep learning modeling steps can utilize fully connected models for learning. In this case, the model used for learning can include one or more input layers, one or more hidden layers (e.g., six hidden layers), and one or more output layers. As an example, a vector representing the speed of each road segment, the number of detectors collected, and the collection time can be input into the input layer, and the output layer can output the speed of the detectors traveling on each lane of the road segment at the same time.

[0108] Furthermore, according to the example, the hidden layer can be composed of predetermined units obtained through an optimization process, for example, units of [64, 128, 128, 128, 64, 32].

[0109] Furthermore, in the deep learning modeling process, the normalization process of adjusting the proportions of the constituent unit values ​​can be applied to each layer.

[0110] Furthermore, in the deep learning modeling step, the modeling algorithm can be corrected by applying a discarding technique, which reduces some information of the unit to "0", to each layer at a predetermined rate, so that the learning is not focused on the learning data.

[0111] Through this learning modeling, the speed of each lane or the arrival time can be predicted, and compared with existing methods, both accuracy and collection rate can be improved.

[0112] As described above, this invention relates to a deep learning model for predicting the speed of each lane when the speed of a segment of road is missing for each lane by utilizing real-time traffic information forming the missing general road segment; when an appropriate speed is not provided for each lane due to low existing collection rates, the real-time speed of a general road segment with a high collection rate can be used to predict the speed of each lane; and more complex speed predictions for each lane can be made by providing information about the number of collections, entry times, and speed data. This invention provides higher prediction accuracy than existing alternative methods by reflecting real-time traffic conditions rather than using patterned speeds when the speed of each lane is missing in existing methods.

[0113] According to embodiments of the present invention, the vehicle speed prediction device and vehicle speed prediction method for each lane can effectively predict the accurate speed for each lane, reflecting real-time traffic conditions.

[0114] According to an embodiment of the present invention, the vehicle speed prediction device and method for each lane can apply a deep learning model by utilizing the real-time speed of lanes not classified as exit lanes, the deep learning model being used to learn the speed relationship for each exit lane.

[0115] According to embodiments of the present invention, the vehicle speed prediction device and vehicle speed prediction method for each lane can perform preprocessing to convert different types of data representing traffic conditions into data suitable for deep learning models.

[0116] Furthermore, according to embodiments of the present invention, the vehicle speed prediction device and vehicle speed prediction method for each lane can apply a deep learning model when no real-time traffic information is collected and no speed for each lane is generated, wherein the output of the deep learning model reflects the predicted speed of each lane in real time rather than the pattern speed.

[0117] Therefore, unlike conventional speed, speed prediction, which utilizes real-time traffic information for each lane, can simultaneously improve accuracy and collection rate, and can significantly supplement the arrival time provided to users.

[0118] In addition, various effects that can be directly or indirectly identified through this document can be provided.

[0119] The above description is merely an illustration of the technical concept of the present invention. Those skilled in the art will be able to make various modifications and changes without departing from the essential characteristics of the present invention.

[0120] Therefore, the embodiments disclosed in this invention are not intended to limit the technical concept of the invention, but rather to explain it, and the scope of the technical concept of the invention is not limited by these embodiments. The scope of protection of this invention should be interpreted as defined in the appended claims, and all technical concepts within their equivalent scope should be interpreted as being included within the scope of this invention.

Claims

1. A vehicle speed prediction apparatus comprising: a data pre-processor configured to perform pre-processing on correct answer data and input data for deep learning; and a deep learning modeling apparatus configured to learn a relationship of a lane speed for each exit lane of a road based on a representative speed of each link, a number of probes collected, and a collection time, wherein the data pre-processor comprises: a correct answer data pre-processor configured to convert probe trajectory data collected at an exit time by traveling all links of each lane into data based on an entry time, and to perform harmonic mean processing on the converted data in a preset reference time unit; and an input data pre-processor configured to perform harmonic mean processing on a representative speed of each link in a preset reference time unit; the correct answer data is obtained using probe trajectory data collected by traveling all links of each lane; the road comprises at least one lane and a plurality of links indicating one section of the at least one lane as a learning object of the deep learning modeling apparatus. 2.The vehicle speed prediction apparatus according to claim 1, wherein the deep learning modeling apparatus is configured to learn using a fully connected model, the fully connected model comprises one or more input layers, one or more hidden layers, and one or more output layers, a vector associated with a representative speed of each link, a number of probes collected, and a collection time is input to the one or more input layers, and the one or more output layers output a speed of a probe traveling in a link of each lane at the same time point. 3.The vehicle speed prediction apparatus according to claim 2, wherein one or more hidden layers are composed of predetermined units obtained through an optimization process, the deep learning modeling apparatus applies a normalization process for adjusting a ratio of values constituting units to each layer.

4. The vehicle speed prediction device according to claim 3, wherein the deep learning modeling apparatus applies a dropout technique for reducing some information of units to "0" to each layer at a predetermined rate. 5.The vehicle speed prediction apparatus according to claim 2, wherein a link of each lane representing a section of each lane to be learned comprises a queue link, a start link before an exit lane, and a plurality of exit links, a representative speed of each link comprises a speed of a queue link and a speed of a start link, and a lane speed of each exit lane comprises a speed of each exit link.

6. The vehicle speed prediction apparatus according to claim 1, wherein the correct answer data pre-processor obtains a representative speed of each lane based on a sum of travel distances of all probes traveling in a time period and a sum of travel times in a time period divided by a preset reference time unit for one day.

7. The vehicle speed prediction device according to claim 6, wherein the input data pre-processor performs harmonic mean processing of a representative speed of each link in a time period unit, and obtains a representative speed of each link after the harmonic mean processing in a time period unit.

8. The vehicle speed prediction apparatus according to claim 1, wherein the data pre-processor performs minimum-maximum normalization processing on a representative speed of each link, a number of probes collected, a collection time, and a ratio of a lane speed of each exit lane. 9.The vehicle speed prediction apparatus of claim 8, wherein, when data regarding the representative speed of each link, the number of collected probes, and the collection time is missing, the data preprocessor uses a preset value as the missing data, the missing data regarding the representative speed of each link includes a first specific value, the missing data regarding the number of collected probes includes a second specific value, and the missing data regarding the collection time includes index information of a preset reference time. 10.A vehicle speed prediction method comprising: a data pre-processing step of performing pre-processing on correct answer data and input data for deep learning; a deep learning modeling step of learning a relationship of a lane speed for each exit lane based on a representative speed of each link, a number of collected probes, and a collection time; wherein the data pre-processing step comprises: a correct answer data pre-processing step of converting probe trajectory data collected at an exit time by traveling all links of each lane into data based on an entry time, and harmonically averaging the converted data in a preset reference time unit; an input data pre-processing step of performing harmonic averaging on the representative speed of each link in a preset reference time unit; the correct answer data is obtained using probe trajectory data collected by traveling all links of each lane. 11.The vehicle speed prediction method of claim 10, wherein, the deep learning modeling step is configured to learn using a fully connected model, the fully connected model comprises one or more input layers, one or more hidden layers, and one or more output layers, a vector associated with the representative speed of each link, the number of collected probes, and the collection time is input to the one or more input layers, and the one or more output layers output a speed of a probe traveling a link of each lane at the same time point. 12.The vehicle speed prediction method of claim 11, wherein, one or more hidden layers are composed of predetermined cells obtained through an optimization process, the deep learning modeling step applies a normalization process for adjusting a ratio of values constituting a cell to each layer.

13. The vehicle speed prediction method according to claim 12, wherein the deep learning modeling step applies a dropout technique for reducing some information of a cell to "0" to each layer at a predetermined rate.

14. The vehicle speed prediction method according to claim 10, wherein the correct answer data pre-processing step obtains a representative speed of each lane based on a sum of travel distances of all probes traveling in a time period and a sum of travel times according to a time period of a day divided by a preset reference time unit.

15. The vehicle speed prediction method according to claim 14, wherein the input data pre-processing step performs harmonic averaging on the representative speed of each link in a time period unit to obtain the representative speed of each link harmonically averaged in a time period unit.

16. The vehicle speed prediction method according to claim 10, wherein the data pre-processing step performs minimum-maximum normalization on the representative speed of each link, the number of collected probes, the collection time, and a ratio of a lane speed of each exit lane. 17.The vehicle speed prediction method of claim 16, wherein, When data on a representative speed of each link, a number of collected probes, and a collection time is missing, the data preprocessing step uses a preset value as the missing data, The missing data on the representative speed of each link includes a first specific value, the missing data on the number of collected probes includes a second specific value, and the missing data on the collection time includes index information of a preset reference time.

Images