Parameter optimization apparatus, object approach determination system, parameter optimization method, and computer program

Abstract

To provide technique for easily optimizing a parameter in a parameter optimization system.SOLUTION: A parameter optimization apparatus includes: an object time series acquisition unit which acquires an object time series including information on positions of an object at each time; a ground-truth time series acquisition unit which acquires, for the object for which the object time series is acquired, a ground-truth time series including information on an approach time zone in which the object is approaching a specific region, the time zone being measured in parallel with acquisition of the object time series; a determination time series acquisition unit which acquires a determination time series output from an object approach determiner by inputting the object time series to the object approach determiner; a storage unit which stores a time series set obtained by associating the object time series, the ground-truth time series, and the determination time series; an evaluation unit which evaluates a parameter using the time series set; a search unit which searches for a new parameter to be set to the object approach determiner, using the result evaluated by the evaluation unit; and an output unit which outputs the parameter searched by the search unit.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a parameter optimization device, an object approach determination system, a parameter optimization method, and a computer program. [Background technology]

[0002] 2. Description of the Related Art Conventionally, a parameter optimization device is known that optimizes parameters of an object approach determiner that determines the presence or absence of an object approaching a specific area (Non-Patent Document 1). [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Zhaozhong Chen, Harel Biggie, Nisar Ahmed, Simon Julier, Christoffer Heckman, “Kalman Filter Auto-tuning through Enforcing Chi-Squared Normalized Error Distributions with Bayesian Optimization” 34 pages, 9 figures, submitted to IEEE Transactions on Aerospace and Electronic Systems, [Retrieved July 19, 2023], Internet<https: / / arxiv.org / abs / 2306.07225> Summary of the Invention [Problem to be solved by the invention]

[0004] However, even with the prior art such as that disclosed in Non-Patent Document 1, there is still room for improvement in terms of the technique for easily optimizing parameters in a parameter optimization device.

[0005] The present invention has been made to solve the above-mentioned problems, and has an object to provide a technique that enables parameters to be easily optimized in a parameter optimization device. [Means for solving the problem]

[0006] The present invention has been made to solve at least part of the above-mentioned problems, and can be realized in the following forms.

[0007] (1) According to one aspect of the present invention, there is provided a parameter optimization device that optimizes parameters set in an object approach determiner that determines the presence or absence of an object approaching a specific area. This parameter optimization device includes an object time series acquisition unit that acquires an object time series including information about the position of the object at each time; a correct time series acquisition unit that acquires a correct time series for the object from which the object time series is acquired, the correct time series including information about an approach time period when the object is approaching the specific area and an identification number, measured in parallel with the acquisition of the object time series; a judgment time series acquisition unit that acquires a judgment time series output from the object approach determinator by inputting the object time series to the object approach determinator, the judgment time series including information about a result of an approach determination of the object calculated using the parameters and the object time series; a memory unit that stores a time series set in which the object time series, the correct time series, and the judgment time series are linked; an evaluation unit that evaluates parameters set in the object approach determinator using the time series set stored in the memory unit; a search unit that searches for new parameters to be set in the object approach determinator using the evaluation result by the evaluation unit; and an output unit that outputs the parameters searched for by the search unit.

[0008] According to this configuration, the correct time series including information on the approach time zone and the identification number when the object approaches the specific area is measured in parallel with the acquisition of the object time series including information on the object's position at each time, so that the correct time series can be easily acquired. Here, the "information on the position" includes not only information on the object's position itself and information on the distance between the object and the specific area, but also information on the temporal change in the position, such as the change in the object's velocity vector and the apparent size of the object, which can be derived from the object's position information. This makes it easy to compare the judgment time series including information on the result of the object's approach judgment obtained by inputting the parameters and the object time series to the object approach judger with the correct time series, so that the effort required to search for parameters suitable for the object approach judgment can be reduced. Therefore, the parameters can be easily optimized.

[0009] (2) In the parameter optimization device of the above embodiment, the correct time series acquisition unit may include a signal transmitter provided on the object, which emits a signal including information about the approach time period when the object approaches the specific area, and a signal receiver capable of receiving the signal emitted by the signal transmitter, and may acquire the correct time series using the signal received by the signal receiver. According to this configuration, the correct time series is measured based on the signal emitted by the signal transmitter provided on the object. This allows the correct time series acquisition unit to easily acquire the correct time series, making it easy to compare the judgment time series with the correct time series. Therefore, the effort required to search for parameters suitable for judging the approach of an object can be further reduced, and the parameters can be further easily optimized.

[0010] (3) In the parameter optimization device of the above embodiment, the object time series acquisition unit may acquire the object time series including information on the distance between the object and the specific area, and the correct time series acquisition unit may use the object time series to acquire a correct time series including information on an evaluation exclusion time period in which no approach judgment is performed because the distance between the object and the specific area is within a predetermined distance. According to this configuration, if the object is too close to the specific area, there is no need to perform an approach judgment, so the correct time series acquisition unit uses the object time series including information on the distance between the object and the specific area to acquire a correct time series including information on an evaluation exclusion time period in which no approach judgment is performed. This makes it possible to exclude unnecessary judgments from the object approach judgments. Therefore, the effort required to search for parameters suitable for object approach judgment can be further reduced, making it easier to optimize the parameters.

[0011] (4) In the parameter optimization device of the above embodiment, the evaluation unit may use the correct time series including information on the approach time period and information on the evaluation excluded time period to calculate the number of erroneous determinations corresponding to false reports of approach and the number of undetermined determinations corresponding to missed approaches, and evaluate the parameters set in the object approach determiner using the number of erroneous determinations and the number of undetermined determinations. According to this configuration, the evaluation unit evaluates the parameters using the number of erroneous determinations and the number of undetermined determinations calculated using the correct time series. This can reliably improve the determination accuracy of the object approach determiner.

[0012] (5) According to another aspect of the present invention, there is provided an object approach determination system. This object tracking system includes the parameter optimization device described above, and the object approach determiner to which the parameters output by the parameter optimization device are input. With this configuration, the parameters output by the parameter optimization device are set in the object approach determiner, thereby improving the accuracy of the object approach determination in the object approach determiner. This makes it possible to prepare an object approach determiner with a relatively high accuracy in determining the approach of an object.

[0013] (6) According to yet another aspect of the present invention, there is provided a parameter optimization method for optimizing parameters set in an object approach determiner that determines the presence or absence of an object approaching a specific area. The parameter optimization method includes a first step of acquiring an object time series including information about the position of an object at each time, a second step of acquiring a correct answer time series including information about an approach time period and an identification number of an object for which the object time series is acquired, measured in parallel with the acquisition of the object time series, in which the object approaches the specific area, a third step of acquiring a judgment time series output from the object approach determiner by inputting the object time series to the object approach determiner, the judgment time series including information about a result of the approach determination of the object calculated using the parameters and the object time series, a storage step of storing a time series set in which the object time series, the correct answer time series, and the judgment time series are linked, an evaluation step of evaluating parameters set in the object approach determiner using the time series set, a search step of searching for new parameters to be set in the object approach determiner using the evaluation result in the evaluation step, and an output step of outputting the parameters searched in the search step. According to this configuration, the ground truth time series acquired in the second step can be easily acquired because it is measured in parallel with the acquisition of the object time series including information on the object's position at each time. This makes it easy to compare the judgment time series with the ground truth time series in the evaluation step, thereby reducing the effort required to search for parameters suitable for judging the approach of an object. Therefore, the parameters can be easily optimized.

[0014] (7) According to yet another aspect of the present invention, there is provided a computer program for causing a computer to execute optimization of parameters set in an object approach determiner that determines the presence or absence of an object approaching a specific area. This computer program causes a computer to execute a first function of acquiring an object time series including information about the position of an object at each time; a second function of acquiring a correct answer time series for an object from which the object time series is acquired, the correct answer time series including information about an approach time period when the object is approaching the specific area and an identification number, measured in parallel with the acquisition of the object time series; a third function of acquiring a judgment time series output from the object approach judger by inputting the object time series to the object approach judger, the judgment time series including information about a result of an approach judgement of the object calculated using the parameters and the object time series; a memory function of storing a time series set in which the object time series, the correct answer time series, and the judgment time series are linked; an evaluation function of evaluating parameters set in the object approach judger using the time series set; a search function of searching for new parameters to be set in the object approach judger using the evaluation result by the evaluation function; and an output function of outputting the parameters searched for by the search function. According to this configuration, the correct time series acquired by the second function can be easily acquired because it is measured in parallel with the acquisition of the object time series including information on the object's position at each time. This makes it easy to compare the judgment time series by the evaluation function with the correct time series, reducing the effort required to search for parameters suitable for judging the approach of an object. Therefore, the parameters can be easily optimized.

[0015] The present invention can be realized in various forms, for example, in the form of a system including a parameter optimization device, a control method for these devices and systems, a computer program for causing these devices and systems to perform parameter optimization, a server device for distributing the computer program, a non-transitory storage medium on which the computer program is stored, etc. [Brief description of the drawings]

[0016] [Figure 1] FIG. 1 is a diagram illustrating an overview of an object approach determination system according to a first embodiment. [Diagram 2] 1 is a diagram illustrating a schematic configuration of a parameter optimization device according to a first embodiment. [Diagram 3] 4 is a flowchart of an object approach determination method according to the first embodiment. [Figure 4] 13 is a diagram showing an object time series acquired by an object time series acquisition unit; FIG. [Diagram 5] 13 is a diagram showing a determination time series acquired by a determination time series acquisition unit of the comparative example. FIG. [Figure 6] 4 is a diagram showing a determination time series acquired by a determination time series acquisition unit in the first embodiment; FIG. [Figure 7] 1A to 1C are diagrams illustrating an overview of a modified example of the object approach determination system according to the first embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0017] First Embodiment FIG. 1 is a diagram for explaining an outline of an object approach judgment system of a first embodiment. The object approach judgment system 100 of this embodiment is a system for judging the presence or absence of an object approaching a specific area, and for example, a crosswalk where a vehicle and a pedestrian ("object") may cross is regarded as a "specific area" and the presence or absence of a pedestrian approaching the crosswalk is judged. The judgment result by the object approach judgment system 100 is used to warn pedestrians by a separately provided attention calling device. As shown in FIG. 1, the object approach judgment system 100 of this embodiment includes a parameter optimization device 1 and an object approach judger 5. Note that FIG. 1 shows a state at a stage before the object approach judger 5 is operated and at a stage where a plurality of parameters set in the object approach judger 5 are optimized by the parameter optimization device 1.

[0018] First, the object approaching determinator 5, whose parameters are optimized by the parameter optimization device 1 of this embodiment, will be described. The object approaching determinator 5 receives information such as feature quantities such as the position at each time and an identification number for identifying the same object for one or more objects as input, and outputs information regarding the presence or absence of an object approaching a specific area. Note that known object approaching determinators include SORT (Simple Online and Realtime Tracking), DeepSORT, ByteTrack, etc., FairMOT (Fair Multiple Object Tracking) and TransMOT, which simultaneously perform object detection and tracking, etc. The parameters set in the object approaching determinator 5 affect the quality of the determination of the presence or absence of an object approaching a specific area.

[0019] The parameter optimization device 1 optimizes parameters to be set in the object approach determinator 5 so as to obtain an appropriate output from the object approach determinator 5. In this embodiment, the parameter optimization device 1 includes a test camera 21, a mobile terminal 22, and a personal computer (PC) 30.

[0020] The test camera 21 is prepared separately from the object approach determiner 5, and captures an image for adjusting parameters set in the object approach determiner 5. The test camera 21 is electrically connected to the personal computer 30, and transmits the captured image to the personal computer 30. In this embodiment, the test camera 21 captures an image A1 including pedestrians H1 and H2 and a crosswalk S1 as a "specific area", as shown in FIG. 1. FIG. 1 shows an image including a pedestrian H1 moving in a direction D1 so as to approach the crosswalk S1, and a pedestrian H2 moving in a direction D2 so as to move away from the crosswalk. Note that the image captured by the test camera 21 does not need to include the "specific area", and it is possible to determine the presence or absence of an object approaching the "specific area" as long as it is known that the "specific area" exists in any direction relative to the imaging range.

[0021] Each of pedestrians H1 and H2 included in image A1 carries a mobile terminal 22. The mobile terminal 22 can transmit various signals by being operated by pedestrians H1 and H2. The signals transmitted by the mobile terminal 22 are received by a personal computer 30.

[0022] The personal computer 30 searches for optimal values ​​of parameters set in the object approach determiner 5 using the image captured by the test camera 21, the signal transmitted by the mobile terminal 22, and the information output by the object approach determiner 5 as a test output, and outputs the searched parameters. The personal computer 30 includes an input / output terminal 31, a storage medium 32, and a CPU (Central Processing Unit) 33. The input / output terminal 31 is an interface with an external recording device such as a keyboard, a mouse, a display, a USB memory, or a flash memory, and a transceiver capable of transmitting and receiving signals, and receives various information and outputs information related to the searched parameters to the object approach determiner 5. The storage medium 32 is a general term for memory devices, and can use various types of storage devices such as a ROM (Read Only Memory), a RAM (Random Access Memory), a solid state drive (SSD), a hard disk (HDD), and a flash memory card. The storage medium 32 stores information input by the input / output terminal 31, and contains a computer program for optimizing the parameters of the object approach determiner 5. The function of the storage medium 32 will be described in detail later. The CPU 33 executes the functions of various programs by expanding the programs stored in the ROM of the storage medium 32 into the RAM. Details of the functions executed by the CPU 33 will be described later. Note that, although the parameter optimization device 1 includes one personal computer 30 in this embodiment, it may be configured with a plurality of arithmetic devices.

[0023] Fig. 2 is a diagram showing a schematic configuration of the parameter optimization device 1 of the first embodiment. Fig. 2 is a diagram showing the relationship between functions of the parameter optimization device 1. The parameter optimization device 1 functions as an object time series acquisition unit 11, a correct answer time series acquisition unit 12, a judgment time series acquisition unit 13, a storage unit 14, an evaluation unit 15, and a parameter search unit 16.

[0024] The object time series acquisition unit 11 includes the test camera 21 and the object detection unit 33a. As described above, the test camera 21 continuously captures images including the pedestrians H1 and H2 and the crosswalk S1.

[0025] The object detection unit 33a obtains the type of each object, the confidence level that the object is of that type, and the circumscribing rectangle on the image of the object at each time point using the imaging results from the test camera 21. As a result, the object detection unit 33a creates an object time series including information on the position of the object at each time. Here, the "information on the position" included in the object time series includes not only information on the object's position itself and information on the distance between the object and a specific area, but also information on the temporal change in position, such as the change in the object's velocity vector and the apparent size of the object, which can be derived from the object's position information. The object detection unit 33a of this embodiment corresponds to the object detection function of the CPU 33. In this embodiment, the object detection unit 33a limits the type of object to be detected to only people in order to suppress erroneous detection of people including pedestrians. In addition, the object detection unit 33a may discard the detection result when the confidence level that the detected object is a person of the specified type is low.

[0026] The correct time series acquisition unit 12 acquires a correct time series including information on an approach time period when an object approaches a specific area, measured in parallel with the acquisition of the object time series, for an object whose object time series is acquired. The correct time series acquisition unit 12 of this embodiment has a mobile terminal 22 carried by the pedestrians H1 and H2. Each of the pedestrians H1 and H2 notifies the personal computer 30 of an approach start time, which is the time when the approach starts, and an approach end time, which is the time when the approach ends, by operating the mobile terminal 22 when the pedestrian starts approaching the crosswalk S1 and when the pedestrian ends approaching the crosswalk S1. These operations by the pedestrians H1 and H2 are performed simultaneously with the imaging by the test camera 21. The function of the correct time series acquisition unit 12 will be described in detail later.

[0027] The judgment time series acquisition unit 13 inputs the object time series to the object approach judger 5 to acquire the judgment time series output from the object approach judger 5. The judgment time series acquisition unit 13 in this embodiment corresponds to a control function of the CPU 33 that controls the drive of the object approach judger 5. The object approach judger 5 is started when an initial value of a parameter is set or when a parameter searched by a parameter search unit 16 described later is input. When the object approach judger 5 is started, the judgment time series acquisition unit 13 inputs the object time series acquired by the object time series acquisition unit 11 to the object approach judger 5. As a result, the object approach judger 5 outputs a judgment time series including information on the pedestrian approach judgement calculated using the parameter and the object time series. The judgment time series acquisition unit 13 acquires the judgment time series output by the object approach judger 5.

[0028] The storage unit 14 stores a time series set in which the object time series acquired by the object time series acquisition unit 11, the correct time series acquired by the correct time series acquisition unit 12, and the judgment time series acquired by the judgment time series acquisition unit 13 are linked. The storage unit 14 in this embodiment corresponds to a storage medium 32 provided in the personal computer 30. In this embodiment, one time series set corresponds to one pattern (movement pattern) related to the movement of either the pedestrian H1 or the pedestrian H2 with respect to the crosswalk S1. Therefore, in this embodiment, in order to optimize the parameters of the object approach determiner 5 for each of the multiple movement patterns, multiple time series sets are stored separately. Therefore, one correct time series and one judgment time series are linked and stored in correspondence with the object time series of one movement pattern. As a method for storing such a time series set, in addition to a hard disk, a file system that can group individual electronic files by folders or directories, and a combination of file formats that can express a time series can be used. In this case, the time series set can be configured by storing files corresponding to the object time series, the correct answer time series, and the judgment time series in a group, for example. Also, a relational database system or the like can be used instead of the file system and electronic files. The file system can be a known NTFS or the like, the file format can be JSON Lines, Apache Parquet, or the like, and the relational database system can be SQLite, or the like.

[0029] The evaluation unit 15 evaluates the parameters set in the object approach determinator 5 by using the time series set stored in the storage unit 14. The evaluation unit 15 in this embodiment corresponds to the evaluation function of the CPU 33. The evaluation unit 15 compares the determined time series with the correct time series for each of the multiple time series sets stored in the storage medium 32, and evaluates the parameters set in the object approach determinator 5. The function of the evaluation unit 15 will be described in detail later.

[0030] The parameter search unit 16 uses the evaluation result by the evaluation unit 15 to search for optimal values ​​of parameters set in the object approach determinator 5. The parameter search unit 16 in this embodiment corresponds to a parameter search function of the CPU 33. The parameter search unit 16 can use a computer program that performs known black-box optimization such as Bayesian optimization and genetic algorithms.

[0031] The output unit 17 outputs the parameters searched for by the parameter search unit 16 to the object approach determiner 5. The output unit 17 in this embodiment corresponds to an input / output terminal 31 that electrically connects the personal computer 30 and the object approach determiner 5.

[0032] Fig. 3 is a flowchart of the object approach determination method of this embodiment. Next, the object approach determination method by the object approach determination system 100 will be described. Here, a method for optimizing parameters set in the object approach determiner 5 by the parameter optimization device 1 before actually operating the object approach determiner 5 will be described in detail. When executing the object approach determination method shown in Fig. 3, it is assumed that any parameter, for example, an initial value of the parameter, has already been set in the object approach determiner 5.

[0033] In the object approach determination method of this embodiment, first, as a preparation step for parameter optimization, an object time series and a correct answer time series are input (step S11). In step S11, the object time series acquisition unit 11 uses the test camera 21 to capture an image including pedestrians H1, H2 and the crosswalk S1. The captured image is output to the object detection unit 33a. The object detection unit 33a uses the image input from the test camera 21 to create an object time series including information on the position at each time for each of the pedestrians H1, H2. The object time series created by the object detection unit 33a is stored in the storage medium 32.

[0034] In step S11, in parallel with the image capture by the test camera 21, the portable terminals 22 carried by the pedestrians H1 and H2 included in the image are used to record the approach start time and approach end time of each of the pedestrians H1 and H2 to the crosswalk S1. Specifically, each of the pedestrians H1 and H2 operates the portable terminals 22 by voice, buttons, etc., and when starting to approach the crosswalk S1, the portable terminals 22 transmit a signal that allows the approach start time and their own identification number to be recognized, and when ending the approach to the crosswalk S1, the portable terminals 22 transmit a signal that allows the approach end time and their own identification number to be recognized. When the personal computer 30 receives the signal transmitted from the portable terminal 22, the personal computer 30 links the signal to the object time series stored in the storage medium 32. That is, in step S11, the object time series and the correct answer time series are input to the personal computer 30 at the same time. In this embodiment, the storage medium 32 stores a plurality of combinations of object time series and correct answer time series that are linked to each other.

[0035] In step S11 of this embodiment, the correct time series acquisition unit 12 includes an evaluation exclusion time period in the correct time series for each of the pedestrians H1 and H2 so that the evaluation of the approach judgment is not performed in a situation where the pedestrians are too close to the crosswalk S1 and the approach judgment is ineffective. The evaluation exclusion time period is linked to the object time series stored in the storage medium 32 by transmitting a signal in the same way as the approach start time and the approach end time, in response to an instruction from the mobile terminal 22 owned by each of the pedestrians H1 and H2.

[0036] Next, the object approach determiner 5 is driven to acquire a judgment time series (step S12). In step S12, the judgment time series acquisition unit 13 inputs the object time series stored in the storage medium 32 to the object approach determiner 5 activated by the parameter search unit 16. As a result, the object approach determiner 5 outputs the judgment result for the pedestrian H1 to the judgment time series acquisition unit 13 based on the parameters already set. The judgment result for the pedestrian H1 or H2 output to the judgment time series acquisition unit 13 is further linked to the combination of the object time series and the correct answer time series stored in the storage medium 32 as a judgment time series, and a time series set corresponding to one operation pattern is created. In this embodiment, the storage medium 32 stores a plurality of combinations of the object time series and the correct answer time series, and in step S12, each of the plurality of object time series is input to the object approach determiner 5, which outputs a judgment time series separately. Therefore, the storage medium 32 stores a plurality of time series sets in which an object time series and a determination time series of a correct answer time series are linked together.

[0037] Next, the time series sets are evaluated (step S13). In step S13, the evaluation unit 15 obtains an evaluation index value for each of the multiple time series sets. The evaluation unit 15 evaluates the parameters set in the object approach determiner 5 by aggregating the obtained multiple evaluation index values.

[0038] In this embodiment, the evaluation unit 15 evaluates the parameters by calculating the number of erroneous judgments corresponding to false reports of approach and the number of undetermined judgments corresponding to missed approaches for each of the multiple time series sets. The number of erroneous judgments and the number of undetermined judgments are calculated using the approach time zone and the evaluation excluded time zone included in the correct time series. Specifically, the correct time series acquisition unit 12 takes into consideration that overlapping of time zones occurs when identification numbers are assigned to each time zone. Therefore, among judgments at each time that are not included in any evaluation excluded time zone in the judgment time series, a judgment that there is approach even though it is not included in any approach time zone is defined as an erroneous judgment. In addition, a judgment that there is no approach even though it is included in any approach time zone is defined as an undetermined judgment. The number of erroneous judgments and undetermined judgments defined in this way are counted for each time series set, and are aggregated into the number of erroneous judgments and the number of undetermined judgments as a whole. In step S16 described later, when multi-objective optimization is possible, the parameter search unit 16 sets the number of erroneous judgments and the number of undetermined judgments as evaluation index values. On the other hand, when multi-objective optimization is not possible, the parameter searching unit 16 sets the overall evaluation based on the weighted average of the number of erroneous judgments and the number of unjudged judgments as the evaluation index value.

[0039] When aggregating the evaluation index values ​​for each time series set, they may be aggregated as they are so that the actual length of time is reflected. Alternatively, they may be normalized and then aggregated so as to represent the ratio to the number of judgments of the judgment time series. In this case, the number of judgments when normalizing may be limited to the number of judgments that are not included in any evaluation exclusion time period. As the representative value used when aggregating the evaluation index values, the median may be used, taking into consideration that the distribution is unpredictable, or the maximum value representing the worst case, or the third quartile may be used as a representative of cases that are not as extreme as the worst case. Also, if one is in a position where the influence of outliers is actively accepted, the average value, or the sum of the average value closer to the worst case and the standard deviation while adopting the concept of the average value, may be used.

[0040] Next, it is determined whether or not the termination condition is satisfied (step S14). In step S14, the evaluation unit 15 determines whether or not the evaluation result in step S13 satisfies a predefined termination condition. Here, the termination condition may be that the time required to optimize the parameters exceeds a predefined time limit, that a desired evaluation index value is obtained, or that either of these two conditions is satisfied. In step S14, if the evaluation unit 15 determines that the evaluation result satisfies the predefined termination condition (step S14: Yes), the process proceeds to step S15. In step S14, if the evaluation unit 15 determines that the evaluation result does not satisfy the predefined termination condition (step S14: No), the process proceeds to step S16.

[0041] If it is determined in step S14 that the evaluation result in step S13 satisfies a predetermined end condition, the operation of the object approach determiner 5 is started (step S15). In step S15, the object approach determiner 5 uses the parameters input in the previous step S13 to determine whether or not a pedestrian is approaching the crosswalk S1.

[0042] In step S14, if it is determined that the evaluation result in step S13 does not satisfy the predetermined termination condition, new parameters are generated (step S16). In step S16, the parameter search unit 16 searches for and generates new parameters as a black-box optimization problem such as Bayesian optimization or a genetic algorithm. For example, when a trade-off between the number of misjudgments and the number of non-judgments in the evaluation unit 15 is intentionally considered, by using a program capable of multi-objective optimization such as Optuna, a Pareto-optimal parameter candidate can be obtained, eliminating the need for wasteful trial and error. In addition, parameter candidates may be enumerated based on an evaluation index value obtained by having the evaluation unit 15 perform evaluation using a part of the time-series set, and the evaluation unit 15 may perform evaluation using the parameter candidates and the remainder of the time-series set. In this case, candidates whose evaluation index value calculated using the parameter candidates and the remainder of the time-series set is worse than the evaluation index value when the parameter candidates were enumerated, or candidates whose evaluation can be said to be poor in comparison with a threshold value may be excluded. For example, for machine learning, a set ratio such as 75% may be used to split the data into training and testing data, as in the well-known sklearn.model_selection.train_test_split. This makes it possible to narrow down parameter candidates and prevent overfitting (overlearning in machine learning).

[0043] After step S16, the process returns to step S12, and the object approach determiner 5 is driven using the newly generated parameters and the object time series stored in the storage medium 32 to obtain a new determination time series. In this way, in the parameter optimization method of this embodiment, if the termination condition in step S14 is not satisfied, the optimization process consisting of steps S12, S13, S14, and S16 is repeated to optimize the parameters.

[0044] Fig. 4 is a diagram showing a specific example of an object time series acquired by the object time series acquisition unit. Fig. 4(a) shows a state in which an object moves away from a specific area, that is, a state in which the distance (vertical axis) between the object and the specific area increases. Fig. 4(b) shows a state in which an object approaches a specific area, that is, a state in which the distance (vertical axis) between the object and the specific area decreases. In the above-mentioned object approach determination method, in step S12, the determination time series acquisition unit 13 inputs an object time series as shown in Fig. 4 to the object approach determiner 5 to obtain a determination time series.

[0045] Fig. 5 is a diagram showing a determination time series acquired by a determination time series acquisition unit of a comparative example. Fig. 5(a) shows a determination time series when the object time series shown in Fig. 4(a) is input to an object approach determiner, and Fig. 5(b) shows a determination time series when the object time series shown in Fig. 4(b) is input to an object approach determiner. Each of Fig. 5(a) and Fig. 5(b) shows a straight line J1 shown approximately parallel to the vertical axis representing distance for each time when it is determined that there is approach (an object is approaching a specific area) as information included in the determination time series.

[0046] In the determination time series acquired by the determination time series acquisition unit of the comparative example shown in FIG. 5, the parameters input to the object approach determiner are manually adjusted so as to prevent erroneous determination that an object is approaching even though it is not included in any approach time period when the object is moving away from a specific area. For this reason, as shown in FIG. 5(a), even if there is a time period when the object approaches the specific area, such as between time t1 and time t2, when the object is moving away from the specific area, a determination that an object is approaching is not output. However, as shown in FIG. 5(b), when the object approaches the specific area, the frequency of the determination that an object is approaching becomes relatively low, and there is a risk of overlooking the approach of the object (for example, between time t3 and time t4 in FIG. 5(b)).

[0047] Fig. 6 is a diagram showing a determination time series acquired by the determination time series acquisition unit of this embodiment. Fig. 6(a) shows a determination time series when the object time series shown in Fig. 4(a) is input to the object approach determiner, and Fig. 6(b) shows a determination time series when the object time series shown in Fig. 4(b) is input to the object approach determiner. As with Fig. 5, Fig. 6(a) and Fig. 6(b) each show a straight line J1 that is drawn substantially parallel to the vertical axis representing distance for each time as information included in the determination time series when it is determined that there is approach.

[0048] In the judgment time series acquired by the judgment time series acquisition unit of this embodiment shown in FIG. 6, the parameters input to the object approach judger are set so as to exclude the evaluation of the beginning of receding. As a result, as shown in FIG. 6(a), when an object recedes from a specific region, even if there is a time period in which the object approaches the specific region, such as between time t01 and time t02, a judgment of approach is output. Also, as shown in FIG. 5(b), after time t5, judgments of approach are continuously output. In this way, the object approach judger of this embodiment is less likely to miss the approach of an object to a specific region.

[0049] According to the parameter optimization device 1 of the present embodiment described above, the correct time series including information on the approach time period when the pedestrian approaches the crosswalk and the identification number is measured in parallel with the acquisition of the object time series including information on the pedestrian's position at each time, so that the correct time series can be easily acquired. This makes it easy to compare the determination time series including information on the result of the pedestrian approach determination obtained by inputting the parameters and the object time series to the object approach determiner 5 with the correct time series, so that the effort required to search for parameters suitable for the pedestrian approach determination can be reduced. Therefore, the parameters can be easily optimized.

[0050] Moreover, according to the parameter optimization device 1 of this embodiment, the correct time series is measured based on the signal transmitted by the mobile terminal 22 carried by the pedestrian. This allows the personal computer 30 to easily acquire the correct time series, and thus makes it easy to compare the judgment time series with the correct time series. This further reduces the effort required to search for parameters suitable for pedestrian approach judgment, and thus makes it easier to optimize the parameters.

[0051] Furthermore, according to the parameter optimization device 1 of this embodiment, since there is no need to perform approach judgment when a pedestrian is too close to the crosswalk, the correct time series acquisition unit 12 uses an object time series including information about the distance between the pedestrian and the crosswalk to acquire a correct time series including information about the evaluation exclusion time period during which approach judgment is not performed. This makes it possible to exclude unnecessary judgments among pedestrian approach judgments. Therefore, the effort required to search for parameters suitable for pedestrian approach judgment can be further reduced, making it easier to optimize parameters.

[0052] Moreover, according to the parameter optimization device 1 of this embodiment, the evaluation unit 15 evaluates the parameters using the number of erroneous determinations and the number of undetermined determinations calculated using the correct time series. This can reliably improve the determination accuracy of the object approach determiner 5.

[0053] Furthermore, according to the object approach determination system 100 of this embodiment, the parameters output by the parameter optimization device 1 are set in the object approach determiner 5, thereby improving the accuracy of pedestrian approach determination in the object approach determiner 5. This makes it possible to prepare an object approach determiner 5 with relatively high accuracy in pedestrian approach determination.

[0054] Furthermore, according to the parameter optimization method of this embodiment, the correct time series acquired in step S11 can be easily acquired because it is measured in parallel with the acquisition of the object time series including information on the pedestrian's position at each time. This makes it easy to compare the judgment time series with the correct time series in step S13, reducing the effort required to search for parameters suitable for pedestrian approach judgment. Therefore, the parameters can be easily optimized.

[0055] Furthermore, according to the computer program of this embodiment, the correct time series acquired by the correct time series acquisition unit 12 can be easily acquired because it is measured in parallel with the acquisition of the object time series including information on the pedestrian's position at each time. This makes it easy for the evaluation unit 15 to compare the judgment time series with the correct time series, thereby reducing the effort required to search for parameters suitable for pedestrian approach judgment. Therefore, the parameters can be easily optimized.

[0056] <Modifications of this embodiment> The present invention is not limited to the above-described embodiment, and can be embodied in various forms without departing from the spirit and scope of the invention. For example, the following modifications are also possible.

[0057] [Variation 1] In the above embodiment, the object time series acquisition unit 11 captures images using the test camera 21, and creates an object time series using the images captured by the object detection unit 33a. However, the configuration of the object time series acquisition unit 11 is not limited to this. Instead of the test camera 21, an object approach determiner 5 in which parameters output by the parameter optimization device 1 are set may be used, and it does not have to be a single test camera, but may be one using multiple sensors. In addition, when optimizing parameters for an object tracker that can simultaneously perform object detection and tracking, the object detection unit may not be required.

[0058] [Variation 2] In the above embodiment, the correct time series acquisition unit 12 is the mobile terminal 22 carried by the pedestrian. However, the configuration of the correct time series acquisition unit 12 is not limited to this. A correct time series may be obtained by associating an identification number with the approach start or approach end each time a specific signal associated with a time is issued.

[0059] FIG. 7 is a schematic diagram for explaining an outline of a modified example of the object approach judgment system 100 of the first embodiment. The parameter optimization device 1 included in the object approach judgment system 100 shown in FIG. 7 includes a test camera 21 and a personal computer 30. In the parameter optimization device 1 of FIG. 7, the correct time series acquisition unit 12 acquires a correct time series by passing the headlights L1 of the automobile Am1 included in the image captured by the test camera 21. Specifically, in the image A1 captured by the test camera 21, the automobile Am1 moves from the back side to the front side. In this case, when the distance between the automobile Am1 and the crosswalk S1, which is a specific area, becomes equal to or less than a predetermined distance, the automobile Am1 can record the approach start time of the automobile Am1 in the image A1 by passing the headlights L1. In addition, when the automobile Am1 passes through the crosswalk S1 and leaves the crosswalk S1, the automobile Am1 can record the approach end time of the automobile Am1 in the image A1 by passing the headlights L1. In this way, the parameter optimization device 1 can acquire a ground truth time series measured in parallel with acquisition of an object time series.

[0060] In addition, when acquiring a correct time series, an order may be set for the timing of the start and end of the approach for each of the two pedestrians H1 and H2. For example, an order may be set such that pedestrian H2 starts approaching after pedestrian H1 starts approaching, pedestrian H2 stops moving after pedestrian H2 starts approaching, and pedestrian H1 stops approaching after pedestrian H2 stops approaching. In this case, too, the order of the start and end of the evaluation exclusion can be controlled to include the evaluation exclusion time period in the correct time series.

[0061] [Variation 3] In the above embodiment, the judgment time series acquisition unit 13 acquires the judgment time series by inputting the object time series and parameters to the actual object approach judger 5. When the actual object approach judgement is performed by an electronic circuit or a computer for edge computing, the object approach judger may be an actual device or may be simulated on the same computer as the evaluation unit 15 and the parameter search unit 16 or on another computer. If the object approach judger is realized as a computer program and the parameters are not precisely optimized, the computer program may be normally operated on the same computer as the evaluation unit 15 and the parameter search unit 16 or on another computer.

[0062] [Variation 4] In the above embodiment, the objects that the object approach determiner 5 determines to be approaching are two pedestrians approaching a crosswalk. However, the objects that the object approach determiner 5 determines to be approaching are not limited to pedestrians. Any object that may be approaching the specific area, such as a bicycle approaching a crosswalk or a car approaching an intersection, may be used. Also, the number of objects approaching the specific area may be one.

[0063] [Variation 5] In the above embodiment, the object approach determination system 100 regards a crosswalk where a vehicle and a pedestrian may cross as a specific area and determines whether or not a pedestrian is approaching the crosswalk. However, the specific area is not limited to this.

[0064] Although the present aspect has been described above based on the embodiment and modified examples, the above-mentioned embodiment of the aspect is intended to facilitate understanding of the present aspect and does not limit the present aspect. The present aspect may be modified or improved without departing from the spirit and scope of the claims, and equivalents are included in the present aspect. Furthermore, if a technical feature is not described as essential in this specification, it may be deleted as appropriate.

[0065] <Application example 1> A parameter optimization device that optimizes parameters set in an object approach determinator that determines whether or not an object is approaching a specific area, comprising: an object time series acquisition unit that acquires an object time series including information about the position of the object at each time; a correct time series acquisition unit that acquires a correct time series including information about an approach time period and an identification number of an object that is approaching the specific area, the information being measured in parallel with the acquisition of the object time series, for the object whose object time series is acquired; a determination time series acquisition unit that acquires a determination time series output from the object approach determiner by inputting the object time series to the object approach determiner, the determination time series including information about a result of the object approach determination calculated using the parameters and the object time series; a storage unit that stores a time series set in which the object time series, the correct answer time series, and the determination time series are linked; an evaluation unit that evaluates parameters set in the object approach determiner by using the time series set stored in the storage unit; a search unit that searches for new parameters to be set in the object approach determiner using the evaluation result by the evaluation unit; An output unit that outputs the parameters searched by the search unit. Parameter optimizer. <Application example 2> The parameter optimization device according to Application Example 1, The correct time series acquisition unit is a signal transmitter provided on the object, the signal transmitter emitting a signal including information regarding the approach time period when the object approaches the specific area; A signal receiver capable of receiving a signal transmitted by the signal transmitter, Obtaining the correct time series using a signal received by the signal receiver. Parameter optimizer. <Application example 3> The parameter optimization device according to Application Example 1 or 2, the object time series acquisition unit acquires the object time series including information regarding a distance between the object and the specific area; the correct time series acquisition unit acquires, using the object time series, a correct time series including information on an evaluation exclusion time period during which a proximity determination is not performed because a distance between the object and the specific area is within a predetermined distance; Parameter optimizer. <Application Example 4> The parameter optimization device according to any one of Application Examples 1 to 3, The evaluation unit is Calculating the number of erroneous determinations corresponding to false reports of approach and the number of undetermined determinations corresponding to missed approaches using the correct answer time series including information on the approach time period and information on the evaluation excluded time period; evaluating parameters set in the object approach determiner using the number of erroneous determinations and the number of undetermined determinations; Parameter optimizer. <Application example 5> An object approach determination system, comprising: A parameter optimization device according to any one of Application Examples 1 to 4; the object approach determiner to which the parameters output by the parameter optimization device are input, Object approach judgment system. <Application Example 6> A parameter optimization method for optimizing parameters set in an object approach determiner that determines whether or not an object is approaching a specific area, comprising the steps of: A first step of obtaining an object time series including information about the object's location over time; a second step of acquiring a ground truth time series for the object from which the object time series is acquired, the ground truth time series including information on an approach time period during which the object is approaching the specific area and an identification number, the information being measured in parallel with the acquisition of the object time series; a third step of acquiring a judgment time series output from the object approach determiner by inputting the object time series to the object approach determiner, the judgment time series including information on a result of the object approach determination calculated using the parameters and the object time series; a storage step of storing a time series set in which the object time series, the correct answer time series, and the determination time series are linked; an evaluation step of evaluating parameters set in the object approach determiner by using the time series set; a search step of searching for new parameters to be set in the object approach determiner using an evaluation result in the evaluation step; and an output step of outputting the parameters searched for in the search step. Parameter optimization methods. <Application Example 7> A computer program for causing a computer to execute optimization of parameters set in an object approach determiner that determines the presence or absence of an object approaching a specific area, comprising: A first function for obtaining an object time series including information about the object's location over time; A second function of acquiring a ground truth time series for an object from which the object time series is acquired, the ground truth time series including information on an approach time period during which the object is approaching the specific area and an identification number, the information being measured in parallel with the acquisition of the object time series; a third function of acquiring a judgment time series output from the object approach determiner by inputting the object time series to the object approach determiner, the judgment time series including information on a result of the object approach determination calculated using the parameters and the object time series; a storage function for storing a time series set in which the object time series, the correct answer time series, and the determination time series are linked; an evaluation function for evaluating parameters set in the object approach determiner by using the time series set; a search function that searches for new parameters to be set in the object approach determiner using an evaluation result by the evaluation function; and an output function for outputting the parameters searched for by the search function. Computer program. [Explanation of symbols]

[0066] 1,2...Parameter optimization device H1, H2: Pedestrians (objects) 5…Object proximity determiner 11...Object time series acquisition section 12... Correct time series acquisition section 13...Judgment time series acquisition section 14...Storage section 15…Evaluation section 16…Parameter search section 17...Output section 21…Test camera 22. Mobile terminal 31…Input / output terminal 32…Storage medium 33a...Object detection unit 100…Object approach judgment system

Claims

1. A parameter optimization device that optimizes parameters set in an object approach determinator that determines whether or not an object is approaching a specific area, comprising: an object time series acquisition unit that acquires an object time series including information about the position of the object at each time; a correct time series acquisition unit that acquires a correct time series including information about an approach time period and an identification number of an object that is approaching the specific area, the information being measured in parallel with the acquisition of the object time series, for the object whose object time series is acquired; a determination time series acquisition unit that acquires a determination time series output from the object approach determiner by inputting the object time series to the object approach determiner, the determination time series including information about a result of the object approach determination calculated using the parameters and the object time series; a storage unit that stores a time series set in which the object time series, the correct answer time series, and the determination time series are linked; an evaluation unit that evaluates parameters set in the object approach determiner by using the time series set stored in the storage unit; a search unit that searches for new parameters to be set in the object approach determiner using the evaluation result by the evaluation unit; An output unit that outputs the parameters searched by the search unit. Parameter optimizer.

2. The parameter optimization device according to claim 1, The correct time series acquisition unit is a signal transmitter provided on the object, the signal transmitter emitting a signal including information regarding the approach time period when the object approaches the specific area; A signal receiver capable of receiving a signal transmitted by the signal transmitter, Obtaining the correct time series using a signal received by the signal receiver. Parameter optimizer.

3. The parameter optimization device according to claim 1 or 2, the object time series acquisition unit acquires the object time series including information regarding a distance between the object and the specific area; the correct time series acquisition unit acquires, using the object time series, a correct time series including information on an evaluation exclusion time period during which a proximity determination is not performed because a distance between the object and the specific area is within a predetermined distance; Parameter optimizer.

4. The parameter optimization device according to claim 3, The evaluation unit is Calculating the number of erroneous determinations corresponding to false reports of approach and the number of undetermined determinations corresponding to missed approaches using the correct answer time series including information on the approach time period and information on the evaluation excluded time period; evaluating parameters set in the object approach determiner using the number of erroneous determinations and the number of undetermined determinations; Parameter optimizer.

5. An object approach determination system, comprising: A parameter optimization device according to claim 1 or 2; the object approach determiner to which the parameters output by the parameter optimization device are input, Object approach judgment system.

6. A parameter optimization method for optimizing parameters set in an object approach determiner that determines whether or not an object is approaching a specific area, comprising the steps of: a first step of obtaining an object time series comprising information about the object's location over time; a second step of acquiring a ground truth time series for the object from which the object time series is acquired, the ground truth time series including information on an approach time period during which the object is approaching the specific area and an identification number, the information being measured in parallel with the acquisition of the object time series; a third step of acquiring a determination time series output from the object approach determiner by inputting the object time series to the object approach determiner, the determination time series including information on a result of the object approach determination calculated using the parameters and the object time series; a storage step of storing a time series set in which the object time series, the correct answer time series, and the determination time series are linked; an evaluation step of evaluating parameters set in the object approach determiner by using the time series set; a search step of searching for new parameters to be set in the object approach determiner using an evaluation result in the evaluation step; and an output step of outputting the parameters searched for in the search step. Parameter optimization methods.

7. A computer program for causing a computer to execute optimization of parameters set in an object approach determiner that determines the presence or absence of an object approaching a specific area, comprising: A first function for obtaining an object time series including information about the object's location over time; A second function of acquiring a ground truth time series for an object from which the object time series is acquired, the ground truth time series including information on an approach time period during which the object is approaching the specific area and an identification number, the information being measured in parallel with the acquisition of the object time series; a third function of acquiring a judgment time series output from the object approach determiner by inputting the object time series to the object approach determiner, the judgment time series including information on a result of the object approach determination calculated using the parameters and the object time series; a storage function for storing a time series set in which the object time series, the correct answer time series, and the determination time series are linked; an evaluation function for evaluating parameters set in the object approach determiner by using the time series set; a search function that searches for new parameters to be set in the object approach determiner using an evaluation result by the evaluation function; and an output function for outputting the parameters searched for by the search function. Computer program.

Images