Computing device and method for inspecting learning data

Abstract

The invention discloses a computing device and method for checking learning data. The computing device includes a parking space recognition device that recognizes a parking space using at least one parking space recognition model for recognizing the parking space based on training data of the parking space, and a data checking device, the data checking device checks information about the parking line and the parking space based on the parking space recognition data recognized by the parking space recognition device.

Description

[0001] Related Applications

[0002] This application claims the benefit of priority from Korean Patent Application No. 10-2023-0138859 filed on October 17, 2023, in the Korean Intellectual Property Office, the entire contents of which are incorporated herein by reference. Technical Field

[0003] The present disclosure relates to a computing device and method for examining training data. Background Art

[0004] Various systems for assisting a driver in driving are loaded into the vehicle.

[0005] As an example, the parking space recognition device may be a system that assists parking of a vehicle, which may recognize a parking space from a surrounding image to guide a driver to the parking space, and may support parking assistance when the driver attempts to park in the selected parking space.

[0006] Such a parking space recognition device may learn data using a deep learning model and may recognize a parking space from an input image based on the training data.

[0007] However, when constructing training data and validation data, the deep learning model for parking space recognition should perform data checks to ensure data quality.

[0008] People directly check the training data and validation data of deep learning models used for parking space recognition in the past. However, as the amount of training data and validation data for deep learning models gradually increases, the time and manpower required for data checking increases, resulting in an increase in the cost of checking data. Summary of the invention

[0009] The present disclosure has been made to solve the above-mentioned problems occurring in the prior art while keeping the advantages achieved by the prior art intact.

[0010] An aspect of the present disclosure provides a computing device for automatically performing an operation for checking parking space recognition data recognized based on a learning model for parking space recognition without human intervention to reduce time spent on checking data and reduce costs, and a method for checking training data.

[0011] Another aspect of the present disclosure provides a computing device for additionally learning a parking space by means of a learning model based on a result of checking data to improve performance of the learning model and a method for checking training data.

[0012] The technical problems to be solved by the present disclosure 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 to which the present disclosure belongs from the following description.

[0013] According to one aspect of the present disclosure, a computing device may include: a parking space recognition device that recognizes a parking space using at least one parking space recognition model that recognizes a parking space based on training data for a parking space; and a data checking device that checks information about parking lines and parking slots based on parking space recognition data recognized by the parking space recognition device.

[0014] In an embodiment, the parking space recognition device may include: at least one parking space recognition model that divides an input image of a parking space into units of a predetermined size and number and recognizes the position and type of a parking line and a parking space of each divided unit; and a recognition device that combines results recognized from at least one parking space recognition model to recognize the parking space and output parking space recognition data.

[0015] In an embodiment, the parking space recognition data may include parking line information regarding the position of the parking line and the type of the parking line and parking space information regarding the position of the parking space and the type of the parking space.

[0016] In an embodiment, the parking space recognition apparatus may be formed as an integrated model in which a plurality of parking space recognition models and recognition devices are coupled to each other.

[0017] In an embodiment, the data checking device may include: a position accuracy determining device, which determines the position accuracy of the parking space identification data based on the position information of the parking lines and parking spaces; a type accuracy determining device, which determines the type accuracy of the parking space identification data based on the type information of the parking lines and parking spaces; a position consistency analyzing device, which analyzes the position consistency of the parking space identification data based on the position information of the parking lines and parking spaces; a type accuracy analyzing device, which analyzes the type consistency of the parking space identification data based on the type information of the parking lines and parking spaces; and a controller, which outputs the result of checking the parking space identification data based on the results of the analysis accuracy and consistency of the positions and types of the parking lines and parking spaces.

[0018] In an embodiment, the position accuracy determining device may calculate a probability value of the probability that a parking line and a parking space will exist for each cell obtained by dividing the parking space, and may determine the position accuracy of the parking space recognition data based on the number of cells whose probability values ​​are within a reference range.

[0019] In an embodiment, the position accuracy determining device may complete the check of the position accuracy of the parking space identification data when the position accuracy of the parking space identification data is greater than or equal to a predetermined first reference value, and may perform a precise check on the position information of the parking space identification data when the position accuracy of the parking space identification data is less than the predetermined first reference value.

[0020] In an embodiment, the type accuracy determination device may determine the type accuracy of the parking space recognition data based on the number of cells in which the probability value of the probability of the presence of parking lines and parking spaces for each cell obtained by dividing the parking spaces is greater than a reference value and the entropy of the cells in which the probability value is greater than the reference value.

[0021] In an embodiment, when the type accuracy of the parking space identification data is greater than or equal to a predetermined second reference value, the type accuracy determination device may complete the check of the type accuracy of the parking space identification data, and when the type accuracy of the parking space identification data is less than the predetermined second reference value, the type accuracy determination device may perform a precise check of the type information of the parking space identification data.

[0022] In one embodiment, the computing device may further include a trigger generator that generates a trigger signal for additional learning of at least one parking space recognition model when the location accuracy of the parking space recognition data is less than a first reference value or when the type accuracy of the parking space recognition data is less than a second reference value.

[0023] In an embodiment, the position consistency analysis device may compare the per-unit position information of the parking space identification data with a reference true value (GT, ground truth) of the per-unit position information obtained from an image of the actual parking space, count the number of units in which the per-unit information of the parking space identification data is different from the GT of the per-unit position information, and may analyze the position consistency of the parking space identification data based on the number of counted units.

[0024] In an embodiment, the position consistency analysis device may complete the check of the position consistency of the parking space identification data when the number of counted units is less than a predetermined third reference value, and may add an indicator for checking the unit in which the per-unit information of the parking space identification data is different from the GT for the per-unit position information when the number of counted units is greater than or equal to the predetermined third reference value.

[0025] In an embodiment, the type consistency analysis device can calculate the per-type probability value of each unit in the parking space identification data, can compare the type information with the highest probability value for each unit with the reference true value (GT) of the per-unit type information obtained from the image of the actual parking space, can count the number of units whose type information with the highest probability value for each unit is different from the GT for the per-unit type information obtained from the image of the actual parking space, and can analyze the type consistency of the parking space identification data based on the number of counted units.

[0026] In an embodiment, when the number of counted units is less than a predetermined fourth reference value, the type consistency analysis device can complete the type consistency check of the parking space identification data, and when the number of counted units is greater than or equal to the determined fourth reference value, check the type information of each unit.

[0027] According to another aspect of the present disclosure, a method for checking training data may include: identifying a parking space using at least one parking space recognition model through a parking space recognition device, wherein the at least one parking space recognition model recognizes the parking space based on training data for the parking space; and checking information about parking lines and parking spaces based on the recognized parking space recognition data.

[0028] In an embodiment, identifying a parking space may include: dividing an input image for a parking space into units having a predetermined size and number by at least one parking space recognition model, and identifying a position and a type of a parking line and a parking space for each divided unit by at least one parking space recognition model, and combining results of recognition from at least one parking space recognition model to identify the parking space and output parking space recognition data.

[0029] In an embodiment, the parking space recognition data may include parking line information regarding the position of the parking line and the type of the parking line and parking space information regarding the position of the parking space and the type of the parking space.

[0030] In an embodiment, the inspection may include: determining the position accuracy of parking space identification data based on the position information of parking lines and parking spaces; determining the type accuracy of parking space identification data based on the type information of parking lines and parking spaces; analyzing the position consistency of parking space identification data based on the position information of parking lines and parking spaces, analyzing the type consistency of parking space identification data based on the type information of parking lines and parking spaces, and outputting the inspection result of parking space identification data based on the analysis result of the accuracy and consistency of the positions and types of parking lines and parking spaces.

[0031] In an embodiment, determining the position accuracy may include: calculating a probability value of the probability that a parking line and a parking space will exist for each cell obtained by dividing the parking space, and determining the position accuracy of the parking space identification data based on the number of cells whose probability values ​​are within a reference range, and when the position accuracy of the parking space identification data is greater than or equal to a predetermined first reference value, completing a check on the position accuracy of the parking space identification data; when the position accuracy of the parking space identification data is less than the predetermined first reference value, performing a precise check on the position information of the parking space identification data.

[0032] In an embodiment, determining the type accuracy may include: determining the type accuracy of the parking space identification data based on the number of cells whose probability values ​​of the probability that there will be parking lines and parking spaces for each cell obtained by dividing the parking spaces are greater than a reference value and the entropy of the cells whose probability values ​​are greater than the reference value, and when the type accuracy of the parking space identification data is greater than or equal to a predetermined second reference value, completing the check on the type accuracy of the parking space identification data, and when the type accuracy of the parking space identification data is less than the predetermined second reference value, accurately checking the type information of the parking space identification data.

[0033] In an embodiment, the method may further include generating a trigger signal for additional learning of at least one parking space recognition model when the position accuracy of the parking space recognition data is less than a first reference value or when the type accuracy of the parking space recognition data is less than a second reference value.

[0034] In an embodiment, the analysis of the position consistency may include: comparing the per-unit position information of the parking space identification data with a reference true value (GT) of the per-unit position information obtained from an image of the actual parking space, counting the number of cells in which the per-unit information of the parking space identification data is different from the GT for each unit position information, and analyzing the position consistency of the parking space identification data based on the number of counted cells, and when the number of counted cells is less than a predetermined third reference value, completing the check on the position consistency of the parking space identification data, and when the number of counted cells is greater than or equal to the predetermined third reference value, checking the position information of each unit.

[0035] In an embodiment, analyzing type consistency may include: calculating a per-type probability value for each cell in the parking space identification data; comparing the type information having the highest probability value for each cell with a reference true value (GT) of the per-cell type information obtained from an image of the actual parking space; counting the number of cells for which the type information having the highest probability value for each cell is different from the GT for the per-cell type information obtained from the image of the actual parking space, and analyzing the type consistency of the parking space identification data based on the counted number of cells, and when the counted number of cells is less than a predetermined fourth reference value, completing the check on the type consistency of the parking space identification data, and when the counted number of cells is greater than or equal to the predetermined fourth reference value, checking the type information of each cell. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and other objects, features and advantages of the present disclosure will become more apparent from the following detailed description in conjunction with the accompanying drawings:

[0037] Figure 1 is a diagram showing a configuration of a computing device according to an embodiment of the present disclosure;

[0038] Figure 2 is a diagram illustrating an operation of identifying a parking space cited to describe an embodiment of the present disclosure;

[0039] Figure 3 and Figure 4 is a diagram showing details of parking space recognition data according to an embodiment of the present disclosure; and

[0040] Figure 5 , Fig. 6A , Figure 6B , Fig. 7A and Figure 7B is a diagram illustrating an operational flow of a method for checking training data according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0041] Hereinafter, some embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In addition, when a reference numeral is attached to a constituent element of each drawing, when the same constituent element is shown in other drawings, the constituent element is also represented by the same reference numeral. In addition, when describing the embodiments of the present disclosure, detailed descriptions of well-known features or functions will be excluded so as not to unnecessarily obscure the subject matter of the present disclosure.

[0042] When describing components according to embodiments of the present disclosure, terms such as "first", "second", "A", "B", "(a)", "(b)", etc. may be used. These terms are intended only to distinguish one component from another, and these terms do not limit the nature, order, or sequence of the corresponding components. In addition, unless otherwise defined, all terms used herein, including technical and scientific terms, have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs. Those terms as defined in commonly used dictionaries should be interpreted as having a meaning equal to the contextual meaning in the relevant art, and should not be interpreted as having an ideal or overly formal meaning unless expressly defined in this application as having an ideal or overly formal meaning.

[0043] Figure 1 is a diagram showing a configuration of a computing device according to an embodiment of the present disclosure. Figure 2 is a diagram referred to to describe an operation of recognizing a parking space according to an embodiment of the present disclosure. Figure 3 and Figure 4 is a diagram showing details of parking space recognition data according to an embodiment of the present disclosure.

[0044] See also Figure 1 , a computing device according to an embodiment of the present disclosure may include a parking space recognition device 10 and a data checking device 100 .

[0045] When receiving parking space recognition data recognized from the parking space recognition device 10, the data checking device 100 may determine the accuracy of the position and type based on information such as parking lines and parking spaces recognized from the received data, and may analyze the consistency between the position and type to check the parking space recognition data.

[0046] Here, the parking space recognition device 10 may be configured to include a plurality of parking space recognition models PSRNet# 1 to PSRNet#N 11 and a recognition device 15 .

[0047] The plurality of parking space recognition models 11 may recognize a parking space based on training data for the parking space. Each of the plurality of parking space recognition models 11 may estimate and output the position and type of a parking line and a parking space from an input image based on the training data.

[0048] At this time, if Figure 2 As shown, the multiple parking space recognition models 11 can divide the input image into units corresponding to the set size and number, and can estimate and output the position and type of parking lines and parking spaces unit by unit based on the training data. In this article, the size of the unit and the number of units can be flexibly set according to the size of the input image. For example, if the size of the input image is 1023×1024, the size of the unit can be set to 32×32, and the number of units can be set to 1024.

[0049] The recognition device 15 may combine the output values ​​of the plurality of parking space recognition models 11 to finally recognize the parking space. Here, the recognition device 15 may sort the output values ​​of the plurality of parking space recognition models 11 unit by unit, and may select the middle value of the output value for each unit to finally recognize the parking space. As another embodiment, the recognition device 15 may select the average value of the output values ​​of the plurality of parking space recognition models 11 to finally recognize the parking space.

[0050] At this time, the recognition device 15 may provide the finally recognized parking space recognition data to the data checking device 100. Herein, the parking space recognition data may include position and type information of parking lines and parking spaces for each unit.

[0051] The data checking device 100 may include a controller 110 , a position accuracy determining device 120 , a type accuracy determining device 130 , a position consistency analyzing device 140 , a type consistency analyzing device 150 , and a trigger generator 160 .

[0052] The controller 110 may be connected with corresponding components of the data check apparatus 100 to perform the overall functions of the data check apparatus 100. Here, the controller 110 may be a hardware device such as a processor or a central processing unit (CPU), or may be a program executed by a processor.

[0053] When receiving the parking space identification data finally identified from the parking space identification device 10, the controller 110 may control the operations of the position accuracy determination device 120, the type accuracy determination device 130, the position consistency analysis device 140, the type consistency analysis device 150 and the trigger generator 160 to check the parking space identification data finally identified against the output value from the parking space identification model 11.

[0054] In this context, the parking space recognition data finally recognized by the parking space recognition device 10 can be represented as follows: Figure 3 As shown, the attribute value of the parking space identification data can be expressed as follows Figure 4 shown.

[0055] See also Figure 3 and Figure 4 , the parking space recognition data may generally include parking line information and parking space information.

[0056] Herein, the stop line information may include position information of the stop line, type information of the stop line, and information on whether a start point of the stop line is visible / invisible.

[0057] The location information of the parking line may include information such as the starting point, midpoint, and end point of the parking line. For example, the location information of the parking line may include attribute information such as the starting point coordinates, midpoint coordinates, and end point coordinates of the parking line, and a parking line ID (ParkingLineid) configured with a parking line main ID, a parking line small ID, a constant, etc.

[0058] In addition, the type information of the stop line may include attribute information, for example, single line (SingleLine) or double line (DoubleLine).

[0059] Furthermore, the information on whether the start point of the stop line is visible / invisible may include attribute information, for example, visible (the start point is visible) or invisible (the start point is invisible).

[0060] Meanwhile, the parking space information may include the location and type information of the parking space. In addition, the parking space information may further include the usage type information of the parking space, the occupancy status type information of the parking space, and the like.

[0061] Here, the position information of the parking space may be represented as information of a parking line pair constituting the parking space. The information of the parking line pair constituting the parking space may include attribute information, for example, two parking line IDs (ParkingLineid) forming a pair in the parking space.

[0062] The type information of the parking space may include attribute information, for example, “vertical”, “parallel”, “diagonal”, or “staircase”.

[0063] The usage type information of the parking space may include attribute information, for example, "normal", "disabled sign", "women-only sign", "EV sign", "RPP area sign", "others" or "unknown".

[0064] The occupancy status type information of the parking space may include attribute information, for example, “occupied”, “empty”, or “unknown”.

[0065] The controller 110 may provide the parking space identification data to the position accuracy determination device 120 and the type accuracy determination device 130 to determine the accuracy of the parking lines and parking spaces for the parking spaces.

[0066] Therefore, the position accuracy determining device 120 may determine the position accuracy of the parking line and the parking space in the recognized parking space based on the parking line and the parking space position information for each unit included in the parking space recognition data input from the controller 110 .

[0067] In detail, the position accuracy determination device 120 can determine the position accuracy of the parking line and the parking space based on the starting point coordinates, midpoint coordinates and end point coordinates of the parking line, a parking line ID (ParkingLineid) configured with a parking line main ID, a parking line small ID and a constant, and two parking line IDs of two parking lines forming a pair in the parking lines constituting the parking space.

[0068] Herein, the position accuracy determining device 120 may calculate a probability value of the probability that a parking line and a parking space will exist for each cell based on the parking line and parking space position information of each cell included in the parking space identification data, and may calculate the position accuracy of the parking space based on the probability value calculated for each cell.

[0069] The location accuracy of the parking space can be calculated using Equation 1 below.

[0070] Equation 1:

[0071]

[0072] In the above equation 1, conf.1 refers to the position accuracy, Th min Refers to the minimum reference value, Th max refers to a maximum reference value, p(i) refers to a probability value of the probability that a parking line and a parking space will exist in the i-th cell, and count() refers to a function for counting the number of cells that meet the conditions.

[0073] The probability value of the existence of a parking line and a parking space for each unit refers to the starting point coordinates, midpoint coordinates and end point coordinates of the existence of the parking line of the i-th unit, the parking line IDParkingLineid configured with parking line large number, parking line small number, constant, etc., and the probability value of the probability of the two parking line IDs forming a pair in the parking lines that constitute the parking space.

[0074] Therefore, the position accuracy determination device 120 can count the number of cells in which the probability value of the probability that a parking line and a parking space will exist is within a reference range (i.e., the number of cells in which the probability value has a value between the minimum reference value and the maximum reference value), and can calculate the position accuracy of the parking line and the parking space of the parking space based on the inverse of the number of counted cells.

[0075] When the position accuracy recognized from the position accuracy determining device 120 is less than the first reference value, the controller 110 may determine that the position of the parking space is inaccurate to perform a precise check, and may transfer the result to the trigger generator 160 .

[0076] Therefore, when the type accuracy is less than the first reference value, the trigger generator 160 may generate a trigger signal for additionally learning the position of the parking space. The controller 110 may deliver the trigger signal generated by the trigger generator 160 to the parking space recognition device 10.

[0077] Therefore, each parking space recognition model 11 of the parking space recognition device 10 can additionally learn the position of the parking space based on the trigger signal.

[0078] At the same time, when the position accuracy identified from the position accuracy determination device 120 is greater than or equal to the first reference value, the controller 110 can determine that the inspection of the position accuracy for the parking space is completed, and can request the position consistency analysis device 140 to analyze the position consistency of the parking line and the parking space for the parking space.

[0079] Therefore, the type accuracy determining device 130 may determine the type accuracy of the parking line and the parking space in the recognized parking space based on the parking line and parking space type information for each unit included in the parking space recognition data input from the controller 110 .

[0080] In detail, the type accuracy determining device 130 can determine the type accuracy of the type of the parking line, whether the starting point of the parking line is visible / invisible, the type of the parking space, the usage type of the parking space, the occupancy status type of the parking space, etc.

[0081] Here, the type accuracy determination device 130 can calculate the probability value and entropy of the probability of the existence of parking lines and parking spaces for each unit based on the parking line and parking space type information for each unit included in the parking space identification data, and can calculate the type accuracy for the parking lines and parking spaces of the parking spaces based on the probability values ​​and entropy calculated for each unit.

[0082] The parking space type accuracy may be calculated using Equation 2 below.

[0083] Equation 2:

[0084]

[0085] In the above equation 2, conf.2 refers to type accuracy, p(i) refers to the probability value of the probability that there will be a parking line and a parking space in the i-th unit, and B{p(i)} refers to a function that outputs "1" when the probability value p(i) of the probability that there will be a parking line and a parking space in the i-th unit is greater than a reference value and outputs "0" when the probability value p(i) is less than or equal to the reference value. As an embodiment, the reference value may be set to 0.8, but is not limited thereto. Obviously, it can be adjusted at any time according to the implementation form.

[0086] In addition, Entropy(i) refers to the entropy of the i-th unit, and count() refers to a function for counting the number of units satisfying a condition.

[0087] Here, p(i) refers to the starting point coordinates, midpoint coordinates and end point coordinates of the parking line in the i-th unit, the parking line ID ParkingLineid configured with the parking line main ID, parking line small ID, constants, etc., and the probability value of the probability of the two parking line IDs forming a pair among the parking lines that constitute the parking space.

[0088] The type accuracy determination device 130 can count the number of cells in which the B(p(i)) value is "1" for each cell based on the probability value p(i) of the probability that there are parking lines and parking spaces in the i-th cell, and can calculate the type accuracy of parking lines and parking spaces based on the sum of the inverse of the number of counted cells and the inverse of the entropy of the cells in which the B(p(i)) value is "1" for each cell.

[0089] When the type accuracy identified from the type accuracy determination device 130 is less than the second reference value, the controller 110 may determine that the type is inaccurate with respect to the parking line and the parking space to perform a precise check, and may deliver the result to the trigger generator 160 .

[0090] Therefore, when the type accuracy is less than the second reference value, the trigger generator 160 may generate a trigger signal for additionally learning the type of the parking space. When the trigger generator 160 generates a trigger signal for additional learning, the controller 110 may transmit the generated trigger signal to the parking space recognition device 10.

[0091] Therefore, each parking space recognition model 11 of the parking space recognition device 10 may additionally learn the type of parking space based on the trigger signal.

[0092] At the same time, when the type accuracy identified from the type accuracy determination device 130 is greater than or equal to the second reference value, the controller 110 can determine that the check on the type accuracy of the parking space is completed, and can request the type consistency analysis device 150 to analyze the parking lines and type consistency of the parking spaces.

[0093] When the accuracy of the parking space position (ie, the position accuracy) determined by the position accuracy determining device 120 is greater than or equal to the first reference value, the position consistency analyzing device 140 may analyze the position consistency.

[0094] Here, in order to identify the consistency between the key point position information of the parking space identification data, for example, the starting point coordinates, midpoint coordinates and end point coordinates of the parking line and the parking line IDParkingLineid configured with the parking line main ID, parking line small ID, constant, etc., and the two parking IDs of the two parking lines forming a pair in the parking lines constituting the parking space, the position consistency analysis device 140 can compare the key point position information of the parking space identification data with the benchmark true value (GT) of the parking line and parking space position obtained from the image of the actual parking space.

[0095] At this time, the position consistency analyzing device 140 may compare the key point positions of the parking space recognition data with the GT on a cell-by-cell basis, and may count the number of cells in which the two values ​​are different from each other.

[0096] The number of cells counted in conjunction with the position of the parking space may be expressed as in Equation 3 below.

[0097] Equation 3:

[0098] n1=count(q(i)≠B{p(i)})

[0099] In the above equation 3, n1 refers to the number of cells in which the key point position of the parking space recognition data is different from that of the GT, q(i) refers to a function for outputting "1" when a key point exists in the i-th cell in the GT and outputting "0" when the key point does not exist, p(i) refers to a probability value of the probability that a parking line and a parking space exist in the i-th cell in the parking space recognition data, B{p(i)} refers to a function that outputs "1" when p(i) is greater than a reference value and outputs "0" when p(1) is less than or equal to the reference value, and count() refers to a function that counts the number of cells that meet the conditions.

[0100] As shown in the above Equation 3, the position consistency analysis device 140 can compare the GT-based q(i) value with B{p(i)} based on the parking space recognition data for each cell, and can count the number of cells in which the two values ​​are different from each other.

[0101] At this time, when the number of counted cells is greater than or equal to the third reference value, the position consistency analysis device 140 may determine that it has failed to check the position consistency of the parking space identification data. In this case, the position consistency analysis device 140 may add an indicator to the cell having a value different from GT for quickly checking the cell. The position consistency analysis device 140 may provide the controller 110 with information of the cell to which the indicator is added.

[0102] Meanwhile, when the number of counted cells is less than the third reference value, the position consistency analysis device 140 may complete the check for the position consistency of the parking space recognition data and may provide the result to the controller 110 .

[0103] Therefore, the controller 110 may output a result of checking the position accuracy and the position consistency of the parking space.

[0104] When the accuracy of the type of the parking space determined by the type accuracy determining device 130 (ie, the type accuracy) is greater than or equal to the second reference value, the type consistency analyzing device 150 may analyze the type consistency.

[0105] Here, in order to identify the consistency of the type output value of the parking space identification data, for example, the type of the parking line, whether the starting point of the parking line is visible / invisible, the type of the parking space, the usage type of the parking space, the occupancy status type of the parking space, etc., the type consistency analysis device 150 may compare the type output value of the parking space identification data with the type GT obtained from the actual parking space image.

[0106] At this time, the type consistency analyzing device 150 may compare the type output value of the parking space recognition data with the GT on a cell-by-cell basis, and may count the number of cells in which the two values ​​are different from each other.

[0107] The number of units counted in conjunction with the type of parking space may be expressed as in Equation 4 below.

[0108] Equation 4:

[0109] n2=count(k(i)≠T{t(i)})

[0110] In the above equation 4, n2 represents the number of cells in which the type output value of the parking space identification data is different from GT, k(i) represents the type value of the i-th cell in GT, t(i) represents the probability value of each type of the i-th cell in the parking space identification data, T{t(i)} represents a function for outputting the type value with the highest t(i), and count() represents a function for counting the number of cells that meet the conditions.

[0111] As shown in the above Equation 4, the type consistency analysis device 150 may compare the k(i) value based on the GT with T{t(i)} based on the parking space recognition data for each cell, and may count the number of cells in which the two values ​​are different from each other.

[0112] At this time, when the number of cells combined with the parking space count is greater than or equal to the fourth reference value, the type consistency analysis device 150 may determine that it has failed to check the type consistency of the parking space identification data. In this case, the type consistency analysis device 150 may add an indicator to the cell having a value different from GT for quickly checking the cell. The type consistency analysis device 150 may provide the controller 110 with information of the cell to which the indicator is added.

[0113] Meanwhile, when the number of cells counted in conjunction with the type of the parking space is less than the fourth reference value, the type consistency analysis device 150 may complete the check for type consistency of the parking space identification data and may provide the result to the controller 110 .

[0114] Therefore, the controller 110 may output a result of checking the type accuracy and type consistency of the parking space.

[0115] When the checking of the position and type of the parking space recognition data based on the parking space recognition model 11 is completed, the controller 110 may end the checking operation.

[0116] Each device of the computing device according to the present embodiment and each component of the device for performing the above-mentioned operation can be implemented in the form of an independent hardware device including a memory and at least one processor for processing each operation, or can be driven in the form of another hardware device such as a microprocessor or a general-purpose computer system. At least one processor can be implemented as an application-specific integrated circuit (ASIC), a central processing unit (CPU), a microcontroller, a microprocessor, etc. In addition, each device and / or component of the device can further include a non-temporary storage medium storing instructions executed by at least one processor. The non-temporary storage medium may include a flash memory, a hard disk, a solid state disk (SSD), a random access memory (RAM), a static RAM (SRAM), a read-only memory (ROM), a programmable ROM (PROM), an electrically erasable programmable ROM (EEPROM), an erasable programmable ROM (EPROM), etc.

[0117] An operation flow of the apparatus for checking parking space recognition data having the above-described configuration according to an embodiment of the present disclosure will be described in detail.

[0118] Figures 5 to 7B is a diagram illustrating an operational flow of a method for checking training data according to an embodiment of the present disclosure.

[0119] First, refer to Figure 5 When receiving the parking space recognition data recognized based on the parking space recognition model 11 from the parking space recognition device 10 in S110, the data checking device 100 may determine the position accuracy and type accuracy of the parking space recognition data received in S110 in S120. Fig. 6A and Figure 6B implementation method.

[0120] Fig. 6A A flow diagram of operations for determining the positional accuracy of parking space identification data is shown. Figure 6B A flow chart of operations for determining the type accuracy of parking space identification data is shown.

[0121] See also Fig. 6A In S210, the data inspection device 100 can Figure 5 The data checking device 100 may calculate the position accuracy based on the probability value of each unit based on the parking space recognition data received in S110. In S210, the data checking device 100 may calculate the position accuracy with reference to the above equation 1.

[0122] When the position accuracy calculated in S210 is less than the predetermined first reference value TH1 in S220, the data checking device 100 may perform a precision check on the position information of the parking space recognition data in S230. In S240, the data checking device 100 may generate a trigger signal for additionally learning the position information.

[0123] Therefore, the parking space recognition device 10 may additionally learn the position of the parking space through the parking space recognition model 11 based on the trigger signal generated in S240. In this case, the data checking device 100 may re-determine the position accuracy of the parking space recognition data based on the result of the additional learning.

[0124] Meanwhile, when the position accuracy is greater than or equal to the first reference value TH1 in S220, the data checking apparatus 100 may perform Figure 5 S130.

[0125] See also Figure 6B In S310, the data inspection device 100 can Figure 5 The data checking device 100 may calculate the entropy-based type accuracy for each unit of the parking space recognition data received in S110. In S310, the data checking device 100 may calculate the type accuracy with reference to the above-mentioned equation 2.

[0126] When the type accuracy calculated in S310 is less than the predetermined second reference value TH2 in S320, the data check device 100 may perform a precision check of the type information of the parking space recognition data in S330. In S340, the data check device 100 may generate a trigger signal for additionally learning type information.

[0127] Therefore, the parking space recognition device 10 may additionally learn the type of the parking space through the parking space recognition model 11 based on the trigger signal generated in S340. In this case, the data check device 100 may re-determine the type accuracy of the parking space recognition data based on the result of the additional learning.

[0128] Meanwhile, when it is determined in S320 that the type accuracy is greater than or equal to the second reference value TH2, the data checking apparatus 100 may perform Figure 5 S130.

[0129] See again Figure 5 When the position accuracy and type accuracy check for the parking space identification data are completed, in S130, the data checking device 100 may analyze the position consistency and type consistency. In this article, the detailed operation flow of S130 is referred to Fig. 7A and Figure 7B implementation method.

[0130] Fig. 7AA flow chart of operations for analyzing parking space identification data for position consistency is shown. Figure 7B A flow chart of operations for analyzing parking space identification data for type consistency is shown.

[0131] See also Fig. 7A Before analyzing the position consistency for the parking space recognition data, the data checking device 100 may receive a reference true value (GT) of a parking line and a parking space position obtained from an image of an actual parking space.

[0132] Therefore, in S410, the data inspection device 100 may Figure 5 The key point position of the parking space recognition data received in S110 is compared with the GT. In S420, the data checking device 100 may count the number of cells in which the key point position of the parking space recognition data and the GT are different from each other. In S420, the data checking device 100 may calculate the number of cells in which the two values ​​are different from each other with reference to the above-mentioned Equation 3.

[0133] When the number of cells counted in S420 is less than the predetermined third reference value TH3 in S430, the data check device 100 may complete the check of the position consistency of the parking space identification data in S440. When the number of cells is greater than or equal to the predetermined third reference value TH3, the data check device 100 may determine that it has failed to check the position consistency of the parking space identification data in S450, and may add an indicator to the cell having a value different from GT for checking the cell.

[0134] When the position consistency check is completed in S440 or S450, Figure 5 In S140 , the data checking device 100 may output a result of checking the accuracy and consistency of the position of the parking space identification data.

[0135] See also Figure 7B Before analyzing the type consistency for the parking space recognition data, the data checking device 100 may receive a reference true value (GT) of a parking line and a type of a parking space obtained from an image of an actual parking space.

[0136] Therefore, in S510, the data inspection device 100 may Figure 5 The type output value of the parking space identification data received in S110 is compared with the GT. In S520, the data checking device 100 may count the number of cells in which the type output value of the parking space identification data and the GT are different from each other. In S520, the data checking device 100 refers to the above-mentioned Formula 4 and calculates the number of cells in which the two type values ​​are different from each other.

[0137] When the number of cells counted in S520 is less than the predetermined fourth reference value TH4 in S530, the data check device 100 may complete the check of the type consistency of the parking space identification data in S540. When the number of cells is greater than or equal to the predetermined fourth reference value TH4, the data check device 100 may determine that it has failed to check the type consistency of the parking space identification data in S550, and may add an indicator to the cell having a different value from GT for checking the cell.

[0138] When the type consistency check is completed in S540 or S550, Figure 5 In S140 , the data checking device 100 may output a result of checking the accuracy and consistency of the type of the parking space identification data.

[0139] As described above, the computing device and method for checking training data according to an embodiment of the present disclosure can automatically perform operations of checking the location and type of parking space identification data identified based on a learning model for parking space identification without manual intervention, so as to reduce the time spent on checking data based on the learning model and reduce costs, and can additionally learn parking spaces through the learning model based on the results of the checking data to improve the performance of the learning model.

[0140] According to an embodiment of the present disclosure, a computing device may automatically perform an operation for checking parking space recognition data identified based on a learning model for parking space recognition without manual intervention, thereby reducing the time spent on checking data and reducing costs.

[0141] Furthermore, according to an embodiment of the present disclosure, the computing device may additionally learn a parking space through a learning model based on a result of the inspection data, thereby improving the performance of the learning model.

[0142] Although the present disclosure has been described above with reference to exemplary embodiments and the accompanying drawings, the present disclosure is not limited thereto, but various modifications and changes may be made by those skilled in the art without departing from the spirit and scope of the present disclosure as claimed in the appended claims.

[0143] Therefore, the exemplary embodiments of the present disclosure are provided to explain the spirit and scope of the present disclosure, but not to limit them, so that the spirit and scope of the present disclosure are not limited by the embodiments. The scope of the present disclosure should be interpreted based on the attached claims, and all technical concepts within the scope equivalent to the claims should be included in the scope of the present disclosure.

Claims

1. A computing device, comprising: a parking space recognition device configured to recognize a parking space using at least one parking space recognition model, the parking space recognition model being configured to recognize the parking space based on training data for the parking space; as well as a data checking device configured to check information on parking lines and parking spaces based on the parking space recognition data recognized by the parking space recognition device; Wherein, the parking space recognition device comprises: the at least one parking space recognition model configured to divide the input image for the parking space into units having a predetermined size and number, and recognize the position and type of the parking line and the parking space for each of the divided units; and an identification device configured to combine the results of identification from the at least one parking space identification model to identify the parking space and output the parking space identification data; and The parking space identification data includes parking line information about the position of the parking line and the type of the parking line, and parking space information about the position of the parking space and the type of the parking space.

2. The computing device of claim 1, wherein: The parking space recognition device is formed as an integrated model in which a plurality of parking space recognition models and the recognition device are coupled to each other.

3. The computing device of claim 1, wherein: The data checking device comprises: a position accuracy determining device configured to determine the position accuracy of the parking space identification data based on the position information of the parking line and the parking space; a type accuracy determining device configured to determine the type accuracy of the parking space identification data based on the parking line and the type information of the parking space; a position consistency analysis device configured to analyze the position consistency of the parking space identification data based on the position information of the parking line and the parking space; a type accuracy analyzing device configured to analyze type consistency of the parking space identification data based on type information of the parking line and the parking space; and A controller is configured to output a result of checking the parking space recognition data based on an analysis result of the accuracy and consistency of the parking line and the position and type of the parking space.

4. The computing device of claim 3, wherein: The position accuracy determining means calculates a probability value of probability of existence of the parking line and the parking space for each cell obtained by dividing the parking space, and determines the position accuracy of the parking space recognition data based on the number of cells whose probability values ​​are within a reference range.

5. The computing device of claim 4, wherein: When the position accuracy of the parking space identification data is greater than or equal to a predetermined first reference value, the position accuracy determining device completes the check on the position accuracy of the parking space identification data, and when the position accuracy of the parking space identification data is less than the predetermined first reference value, the position accuracy determining device performs a precise check on the position information of the parking space identification data.

6. The computing device of claim 3, wherein: The type accuracy determination device determines the type accuracy of the parking space recognition data based on the number of cells whose probability values ​​of the probability of the existence of the parking line and the parking space for each cell obtained by dividing the parking space are greater than a reference value and the entropy of the cells whose probability values ​​are greater than the reference value.

7. The computing device of claim 6, wherein: When the type accuracy of the parking space identification data is greater than or equal to a predetermined second reference value, the type accuracy determination device completes the check on the type accuracy of the parking space identification data; when the type accuracy of the parking space identification data is less than the predetermined second reference value, the type accuracy determination device performs an accurate check on the type information of the parking space identification data.

8. The computing device of claim 3, further comprising: A trigger generator is configured to generate a trigger signal for additional learning of the at least one parking space recognition model when the position accuracy of the parking space recognition data is less than a first reference value or when the type accuracy of the parking space recognition data is less than a second reference value.

9. The computing device of claim 3, wherein: The position consistency analysis device compares the per-unit position information of the parking space identification data with a reference true value of the per-unit position information obtained from an image of an actual parking space, counts the number of units whose per-unit information of the parking space identification data is different from the reference true value of the per-unit position information, and analyzes the position consistency of the parking space identification data based on the counted number of units.

10. The computing device of claim 9, wherein: When the number of counted units is less than a predetermined third reference value, the position consistency analysis device completes the position consistency check of the parking space identification data, and when the number of counted units is greater than or equal to the predetermined third reference value, the position consistency analysis device adds an indicator, which is used to check the unit whose per-unit information of the parking space identification data is different from the benchmark true value for the per-unit position information.

11. The computing device of claim 3, wherein: The type consistency analyzing device calculates a per-type probability value for each unit in the parking space identification data, compares the type information having the highest probability value for each unit with a reference true value for the per-unit type information obtained from an image of an actual parking space, counts the number of units whose type information having the highest probability value for each unit is different from the reference true value for the per-unit type information obtained from the image of the actual parking space, and analyzes the type consistency of the parking space identification data based on the number of counted units.

12. The computing device of claim 11, wherein: When the number of counted units is less than a predetermined fourth reference value, the type consistency analysis device completes the type consistency check of the parking space identification data, and when the number of counted units is greater than or equal to the predetermined fourth reference value, the type consistency analysis device checks the type information of each unit.

13. A method for checking training data, the method comprising: recognizing, by a parking space recognition device, a parking space using at least one parking space recognition model, the parking space recognition model being configured to recognize the parking space based on training data for the parking space; as well as checking information about parking lines and parking spaces based on the identified parking space identification data; Among them, identifying parking spaces includes: dividing the input image for the parking space into units having a predetermined size and number by the at least one parking space recognition model, and recognizing the position and type of the parking line and the parking space for each of the divided units by the at least one parking space recognition model; and combining the results of recognition from the at least one parking space recognition model to recognize the parking space, and outputting the parking space recognition data; and The parking space identification data includes parking line information about the position of the parking line and the type of the parking line, and parking space information about the position of the parking space and the type of the parking space.

14. The method according to claim 13, wherein: The inspection includes: determining the position accuracy of the parking space identification data based on the position information of the parking line and the parking space; determining type accuracy of the parking space identification data based on the parking line and type information of the parking space; analyzing the position consistency of the parking space identification data based on the position information of the parking line and the parking space; analyzing the type consistency of the parking space identification data based on the type information of the parking line and the parking space; and Based on the analysis result of the accuracy and consistency of the positions and types of the parking lines and the parking spaces, a result of checking the parking space recognition data is output.

15. The method according to claim 14, wherein: Determining location accuracy includes: calculating a probability value of a probability of existence of the parking line and the parking space for each cell obtained by dividing the parking space, and determining the position accuracy of the parking space recognition data based on the number of cells having the probability value within a reference range; and When the position accuracy of the parking space identification data is greater than or equal to a predetermined first reference value, the position accuracy of the parking space identification data is checked, and when the position accuracy of the parking space identification data is less than the predetermined first reference value, a precise check is performed on the position information of the parking space identification data.

16. The method according to claim 14, wherein: Determining type accuracy includes: determining the type accuracy of the parking space recognition data based on the number of cells having a probability value of the probability of the parking line and the parking space existing for each cell obtained by dividing the parking space greater than a reference value and entropy of the cells having the probability value greater than the reference value; and When the type accuracy of the parking space identification data is greater than or equal to a predetermined second reference value, the check on the type accuracy of the parking space identification data is completed; when the type accuracy of the parking space identification data is less than the predetermined second reference value, the type information of the parking space identification data is accurately checked.

17. The method according to claim 14, further comprising: When the position accuracy of the parking space recognition data is less than a first reference value or when the type accuracy of the parking space recognition data is less than a second reference value, a trigger signal for additional learning of the at least one parking space recognition model is generated.

18. The method according to claim 14, wherein: Analyzing position consistency includes: comparing the per-unit position information of the parking space identification data with a reference true value of the per-unit position information obtained from an image of an actual parking space, counting the number of units in which the per-unit information of the parking space identification data differs from the reference true value of the per-unit position information, and analyzing the position consistency of the parking space identification data based on the number of counted units; and When the number of counted cells is less than a predetermined third reference value, checking the position consistency of the parking space identification data is completed, and when the number of counted cells is greater than or equal to the predetermined third reference value, checking the position information of each cell.

19. The method according to claim 14, wherein: Analysis type consistency includes: calculating a per-type probability value for each cell in the parking space recognition data, comparing type information having a highest probability value for each cell with a reference true value for the per-cell type information obtained from an image of an actual parking space, counting the number of cells for which the type information having the highest probability value for each cell is different from the reference true value for the per-cell type information obtained from the image of the actual parking space, and analyzing the type consistency of the parking space recognition data based on the number of counted cells; and When the number of counted cells is less than a predetermined fourth reference value, checking of type consistency of the parking space identification data is completed, and when the number of counted cells is greater than or equal to the predetermined fourth reference value, type information of each cell is checked.

Images