JP6569388B2 - Control device, control system, control method, and program - Google Patents

Description

  The present invention relates to a control device, a control system, a control method, and a program, and more particularly to a control device, a control system, a control method, and a program that control switching of a processing device that executes analysis processing.

  In recent years, with the development of information processing technology, analysis engines that analyze various data have been developed. For example, an analysis engine that generates position information for tracing a human flow line from video data, an analysis engine that identifies a person from still video data, and an analysis engine that generates text data from audio data have been developed.

  Furthermore, an analysis system capable of obtaining various analysis processing results from input data by combining a plurality of analysis engines of the same type or different types has been developed. For example, analysis processing such as processing video data input from a camera in parallel or in series using a person extraction engine, a flow line extraction engine, a face extraction engine, and a face matching engine to determine a person with a predetermined behavior. A system to do it has been developed.

  An analysis system using the plurality of analysis engines can include a plurality of processing devices. In such an analysis system, it is desirable that the accuracy of the analysis result is ensured to a certain level or more and that the analysis result is output in real time so that the processing speed is high.

  However, in an analysis system that combines a plurality of analysis engines as described above, the load applied to each of the analysis engines can vary greatly depending on the content of the data to be analyzed. In addition, a certain kind of analysis engine keeps a state because it continuously processes while referring to the preceding and succeeding data frames, and the processing apparatus cannot be distributed and executed for each data frame.

  For this reason, the load applied to the analysis processing may be concentrated locally on a specific analysis engine. In such a case, analysis processing is delayed or data loss occurs due to buffer overflow. For this reason, there arises a problem that the processing performance and accuracy of the entire analysis system are not guaranteed.

  In other words, depending on the load applied to the analysis engine, an overload occurs in the processing capability of the processing apparatus executing the processing of the analysis engine, which is caused by an unintended analysis processing delay or unintended data loss. Decrease in analysis accuracy occurs.

  To cope with such a problem, there is a case in which a processing device that executes a certain analysis engine is changed to another processing device that has sufficient processing capacity.

  Patent Document 1 discloses an example of a method for migrating an application to another data center in the analysis system as described above. In the method described in Patent Document 1, first, an integrated management system in a data center that is a migration source of an application generates a migration request document including network requirements, device requirements, and storage requirements, and sends it to another data center. The integrated management system at the requested data center checks each local environment to determine if there are sufficient resources to run the application. Thereby, the data center of the migration source can find a data center capable of executing the application according to the load situation.

  Patent Document 2 discloses a mobile terminal that measures a communication state and predicts a throughput from the measured communication state. The mobile terminal executes processing related to the application on its own device or causes the arithmetic processing device to execute it according to the predicted throughput.

Japanese Patent No. 5137709 JP 2014-164568 A

  An apparatus using the method described in Patent Document 1 described above can suppress a decrease in performance of an application after migration when the application is migrated to another processing apparatus. However, this method does not take into account the performance or accuracy degradation that temporarily occurs during the transition. For example, when performing an operation of temporarily stopping the analysis engine and restarting it on another processing device at the time of transition, the analysis engine discards the temporary data such as background learning data that has been held so far, The background may be learned again. This method does not consider the influence that the analysis accuracy of the analysis engine deteriorates during the period from the completion of the transition process including such relearning to the normal output of the analysis result.

  That is, according to the method described in the above-mentioned prior art document, when the analysis engine is a stateful analysis engine and the load is controlled by switching the processing devices, the plurality of processing devices are efficiently executed by the plurality of processing devices. I can't. Here, the stateful analysis engine keeps the state of the temporary data such as the attribute information of the object being tracked in order to perform continuous processing while referring to the previous and next data frames, and distributes the processing device for each data frame It is an analysis engine that cannot be executed. Here, the switching of the processing device is an operation that starts after the analysis engine being executed by a certain processing device is stopped and moved to another processing device.

  The device disclosed in Patent Document 2 also has the same problem as described above.

  An object of the present invention is to provide a control device, a control system, a control method, and a program for solving the above-described problems that occur when a processing device that executes an analysis engine or the like is switched.

  The control device according to one embodiment of the present invention is configured so that one of a plurality of analysis processes (hereinafter referred to as a switching target analysis process) being executed by the first processing apparatus is switched to the second processing apparatus and executed. A switching accuracy score indicating the accuracy of the switching target analysis processing is calculated based on the accuracy prediction value during the switching processing and the accuracy prediction value after the switching is completed, and based on the calculated switching accuracy score, the switching target A switching determination unit that determines whether or not to switch the analysis process is provided.

  The method according to an embodiment of the present invention provides the switching in the case where one of a plurality of analysis processes (hereinafter referred to as a switching target analysis process) being executed in the first processing apparatus is switched to the second processing apparatus and executed. A switching accuracy score indicating the accuracy of the target analysis processing is calculated based on the accuracy predicted value during the switching processing and the accuracy predicted value after the switching is completed, and the switching target analysis is calculated based on the calculated switching accuracy score. It is determined whether or not to switch processing.

  The program according to one embodiment of the present invention provides the switching in a case where one of a plurality of analysis processes (hereinafter referred to as a switching target analysis process) being executed in the first processing apparatus is switched to the second processing apparatus and executed. A switching accuracy score indicating the accuracy of the target analysis processing is calculated based on the accuracy predicted value during the switching processing and the accuracy predicted value after the switching is completed, and the switching target analysis is calculated based on the calculated switching accuracy score. The computer is caused to execute a switching determination process for determining whether or not to execute the process switching.

  There is an analysis process that temporarily reduces the processing accuracy when the processing apparatus is switched. The control device according to the present invention can efficiently execute the processing device switching of such analysis processing.

FIG. 1 is a configuration diagram of a control system 50 according to the first embodiment. FIG. 2 shows a state of the control system 50 after the analysis process b is switched from the processing apparatus 20-a to the processing apparatus 20-b. FIG. 3 is a diagram for explaining a period related to switching of the analysis process and accuracy prediction of each period. FIG. 4 shows an example of the switching accuracy prediction information stored in the switching accuracy prediction information storage unit 33. FIG. 5 shows an example of the load accuracy prediction information stored in the load accuracy prediction information storage unit 34. FIG. 6 is a graph showing the load prediction accuracy information for the analysis process a in the example of FIG. FIG. 7 shows an example of processing resource information stored in the processing device resource information storage unit 35. FIG. 8 is a flowchart showing the operation of the control device 30. FIG. 9 is a flowchart showing in detail an operation (step S2) in which the switching determination unit 32 determines a switching pattern representing a combination of an analysis process to be switched and a switching destination processing apparatus. FIG. 10 is a configuration diagram of a control system 50 according to the second embodiment. FIG. 11 shows an example of the environment switching accuracy prediction information stored in the environment switching accuracy prediction information storage unit 36. FIG. 12 is a flowchart illustrating the operation of the control device 30 according to the second embodiment. FIG. 13 is a configuration diagram of the control device 30 according to the third embodiment. FIG. 14 shows an example of a user accuracy request stored in the accuracy request information storage unit 37. FIG. 15 is a configuration diagram of the computer device 40. FIG. 16 is a diagram illustrating a specific example of the configuration of the control system 50 according to the first embodiment. FIG. 17 illustrates the value of the switching accuracy prediction information in the specific example of the configuration of the control system 50 according to the first embodiment. FIG. 18 illustrates the value of the accuracy prediction information for each load in the specific example of the configuration of the control system 50 according to the first embodiment. FIG. 19 illustrates the value of the processing device resource information in the specific example of the configuration of the control system 50 according to the first embodiment. FIG. 20 is a table illustrating the load of each analysis program output by the load measurement program 21e. FIG. 21 is a table showing specific calculated values of the pre-switching accuracy prediction value and the pre-switching accuracy score based on the values exemplified in FIGS. 17 to 20. FIG. 22 is a table showing specific calculated values of the switching accuracy predicted value, the post-switching accuracy predicted value, and the switching accuracy score for each switching pattern based on the values illustrated in FIGS. 19 to 20. FIG. 23 is a diagram illustrating a specific example of the configuration of the control system 50 according to the second embodiment. FIG. 24 shows switching accuracy prediction information for each environment stored in the switching accuracy prediction information storage unit 36 for each environment. FIG. 25 shows the accuracy requirement values for each analysis program included in the accuracy requirement information in the specific example of the third embodiment. FIG. 26 is a table showing specific calculated values of the pre-switching accuracy requirement satisfaction rate and the pre-switching accuracy score based on the accuracy requirements shown in FIG. FIG. 27 is a table showing specific calculated values of the switching accuracy requirement satisfaction rate, the post-switching accuracy requirement satisfaction rate, and the switching accuracy score. FIG. 28 is a configuration diagram of the control device 30 according to the fourth embodiment.

[First embodiment]
[Constitution]
FIG. 1 is a configuration diagram of a control system 50 according to the present embodiment. The control system 50 includes a plurality of processing devices 20 and a control device 30 connected thereto. The control system 50 of FIG. 1 shows an example including two processing devices 20 (processing device 20-a and processing device 20-b).

  Each processing device 20 includes an execution unit 22 that executes analysis processing and a load measurement unit 21. The execution unit 22 and the load measurement unit 21 are connected to each other. The execution unit 22 executes one or more analysis processes. The processing device 20-a in FIG. 1 shows an example in which two analysis processes (analysis process a and analysis process b) are being executed. The processing device 20-b in FIG. 1 shows an example in which one analysis process (analysis process c) is performed.

  The analysis process is, for example, a process of receiving video data from a surveillance camera and detecting a person who is in a no entry area from the image. The analysis process may be a process of receiving sound data of the sound collecting microphone and detecting sound having a specific characteristic.

  Each analysis process receives data from the associated data distribution apparatus 10, analyzes the data, and outputs a result. In each analysis process, the analysis result may be output to a screen or a file provided in the processing device 20, or may be transferred to another display device (not shown) and output.

  In the example of FIG. 1, the analysis process a receives data from the data distribution apparatus 10-a and analyzes the data. The analysis process b receives data from the data distribution apparatus 10-b and analyzes the data. The analysis process a receives data from the data distribution apparatus 10-c and analyzes the data.

  The load measurement unit 21 measures the load of the analysis process being executed by the execution unit 22 and transmits the result to the control device 30.

  The control device 30 receives load information of each analysis process from the load measurement unit 21 of each processing device 20 in the control system 50. If the control device 30 determines, based on the load information, that the analysis processing being executed in any of the processing devices 20 should be switched to another processing device 20 from the viewpoint of maintaining and improving the accuracy of the analysis processing. The switching is executed.

  Here, the accuracy of the analysis process is an index indicating the correctness of the analysis process result. For example, when recognizing that a person has entered the prohibited area from the video, it can be said that the accuracy is high in the case where the analysis processing result indicates an alert for the video in which the person has actually entered the prohibited area. . On the other hand, in the case where the analysis process result does not show an alert even though a person has entered the prohibited area, or the analysis process result shows an alert when the person has not entered the prohibited area, the accuracy is low. I can say that.

  Further, for example, the accuracy indicates the correctness of the matching result in face matching. The accuracy of the person specified by the collation system using the human face image as input is high if the rate of matching with the actual person is high, and the accuracy is low if the ratio is low.

  When the analysis processing load is high with respect to the capability of the processing device 20, the accuracy of the analysis result may be reduced. For example, in the video stream analysis processing, continuous frames are generally buffered and analyzed for each frame. When the processing load is high and the number of frames to be buffered exceeds the upper limit, processing for thinning out the frames in the buffer may be performed. For this reason, there may be a phenomenon in which sufficient data cannot be obtained for performing the analysis process correctly and the accuracy of the analysis result is lowered.

  The control device 30 determines the necessity of switching the processing device 20 that executes the analysis processing to another processing device 20 in order to prevent or eliminate the above-described decrease in accuracy of the analysis result, and performs switching when necessary. To do.

  The control device 30 includes a switching execution unit 31, a switching determination unit 32, a switching accuracy prediction information storage unit 33, a load accuracy prediction information storage unit 34, and a resource information storage unit 35.

  The switch execution unit 31 and the switch determination unit 32 are connected to each other, and are further connected to the execution unit 22 and the load measurement unit 21 of the processing device 20, respectively. The switching determination unit 32 is also connected to a switching accuracy prediction information storage unit 33, a load accuracy prediction information storage unit 34, and a resource information storage unit 35.

  Based on the load information obtained from each processing device 20, the switching determination unit 32 determines a switching pattern that represents a combination of the analysis processing to be switched and the switching destination processing device 20. The switching execution unit 31 executes the switching determined by the switching determination unit 32.

  Here, the switching operation will be described with reference to FIGS. 1 and 2. In FIG. 1, when switching from the processing device 20-a to the processing device 20-b for the analysis processing b, the switching execution unit 31 stops the analysis processing b in the execution unit 22 of the processing device 20-a. Instruct. The execution unit 22 of the processing apparatus 20-a stops the execution of the analysis process b and instructs the load measurement unit 21 in the apparatus to stop the measurement of the analysis process b.

  Next, the switching execution unit 31 moves data necessary for continuing the analysis process b from the processing device 20-a to the processing device 20-b. For example, the switching execution unit 31 receives necessary data from the execution unit 22 of the processing device 20-a, and transfers the data to the execution unit 22 of the processing device 20-b. Here, the data necessary for the continuation of the analysis process is, for example, setting information such as an identifier of the associated data distribution apparatus 10 and other process progress information. Data necessary for continuing the analysis process may include a program, a script, and the like necessary for the analysis process b.

  Finally, the switching execution unit 31 instructs the execution unit 22 of the processing device 20-b to start the analysis process b. The execution unit 22 of the processing device 20-b starts execution of the analysis process b, and instructs the load measurement unit 21 in the apparatus to start measurement of the analysis process b.

  It should be noted that information and the like necessary for the execution of the analysis process b may be preliminarily arranged in both the processing device 20-a and the processing device 20-b and set up in an executable state. In this case, the switching execution unit 31 omits the data movement operation necessary for continuing the analysis process b.

  FIG. 2 shows a state of the control system 50 after the analysis process b is switched from the processing apparatus 20-a to the processing apparatus 20-b. The switching of other analysis processes is performed in the same manner.

  The switching determination unit 32 predicts the accuracy of each analysis process in the process of determining a switching pattern based on the load information obtained from each processing device 20. Here, the accuracy prediction will be described. FIG. 3 is a diagram for explaining a period related to switching of the analysis process and accuracy prediction of each period. When determining whether or not to switch any analysis process, the switching determination unit 32 predicts the accuracy of each analysis process for a prediction period having a certain length given by the parameter. The switching determination unit 32 performs accuracy prediction in two cases, when the analysis process is switched (b in FIG. 3) and when it is not performed (a) in FIG.

  In the case where the analysis process is switched (b in FIG. 3), the switching determination unit 32 performs accuracy prediction by dividing the prediction period into a switching period and a post-switching prediction period. The switching period is a period from when the analysis processing is stopped by the original processing device 20 and restarted by another processing device 20 until the analysis processing finishes the preliminary processing such as background learning and outputs the analysis result in a steady state. Represents. The switching accuracy prediction is an accuracy prediction value for this period.

  The post-switching prediction period is a period for outputting the analysis result in a steady state after the switching period has passed and the analysis process in which the processing device 20 has been switched finishes the learning process or the like. The post-switching accuracy prediction is an accuracy prediction value for this period.

  In the analysis process, when the processing device 20 is switched, the temporary data such as the background learning data held so far may be discarded and the preprocessing such as background learning may be performed again in the separate processing device 20. is there. Therefore, the analysis accuracy decreases during the period until learning is completed and the analysis result is output in a steady state, that is, during the switching period. Therefore, the accuracy prediction at switching is often lower than the accuracy prediction after switching. In order to reflect this temporary decrease in accuracy associated with the switching of the analysis process in the prediction, the switching determination unit 32 performs the accuracy prediction by dividing the prediction period into a switching period and a post-switching prediction period.

  On the other hand, when the analysis processing is not switched (a in FIG. 3)), the switching determination unit 32 performs uniform accuracy prediction for the entire prediction period. The accuracy prediction before switching is an accuracy prediction value in this case.

  The switching accuracy prediction information storage unit 33 stores switching accuracy prediction information for each analysis process. The switching accuracy prediction information includes switching accuracy prediction and data indicating a switching period. FIG. 4 shows an example of the switching accuracy prediction information stored in the switching accuracy prediction information storage unit 33. The example of FIG. 4 represents the accuracy prediction when switching is performed as a relative value when the maximum accuracy value is 1.0. For example, when a person in the video data takes an abnormal action, the analysis process detects the abnormal action and issues an alert. In the case where the correct number of alerts in the switching period is reduced by half due to the switching, the example of FIG. 4 shows the switching accuracy prediction as 0.5.

  Note that the switching accuracy prediction information is not limited to the example illustrated in FIG. 4, and may be expressed using other functions or table formats. Further, the coefficients and tabular values of these functions may be values calculated from accuracy values measured by executing switching, or values calculated from theoretical values or the like.

  The load accuracy prediction information storage unit 34 stores load accuracy prediction information for each analysis process. The accuracy prediction information for each load represents accuracy prediction according to the load at the time of prediction for each analysis process.

  FIG. 5 shows an example of the load accuracy prediction information stored in the load accuracy prediction information storage unit 34. FIG. 6 is a graph showing the load prediction accuracy information for the analysis process a in the example of FIG.

  Here, in the example of FIGS. 5 and 6, the load factor L is the total load of the analysis processing for the processing resources of the processing device 20. For example, when there are four processors of the processing device 20 and loads for six processors occur in total, the load factor L is 6/4 = 1.5.

  In general, when the load factor L is less than 1.0, the processing resources of the processing device 20 have sufficient capacity, and the accuracy of analysis does not decrease. When the load factor L is 1.0 or more, the processing device 20 is in an overload state, and the accuracy of the analysis processing decreases as the load factor L increases. In the examples of the accuracy prediction information for each load in FIGS. 5 and 6, the accuracy when the load factor L is less than 1.0 is 1.0, and the load factor L increases when the load factor L is 1.0 or more. Accordingly, the function linearly decreases until the accuracy reaches 0.0. Here, in this example, the accuracy prediction value is constant regardless of the length of the prediction period.

  Note that the per-load accuracy prediction information is not limited to the examples illustrated in FIGS. 5 and 6, and may be expressed using other functions or table formats such as a function that varies depending on the length of the prediction period. . In addition, the coefficients and tabular values of these functions may be values calculated from load values measured by performing analysis, or values calculated from theoretical values, etc. Good.

  The processing device resource information storage unit 35 stores the processing resources of each processing device 20. FIG. 7 shows an example of processing resource information stored in the processing device resource information storage unit 35.

  Here, in the example of FIG. 7, the processing resource of the processing device 20 is expressed as time (seconds) in which processing can be executed by the entire processing device 20 per unit time. For example, when the number of processors in the processing device 20 is four, the time during which processing can be executed by the entire processing device per second is 4 seconds. In addition, when the number of processors of the processing device 20 is one and the processing performance of the processor is half that of the reference processor, the time that can be executed by the entire processing device 20 per second is 0.5 seconds. It becomes.

  These values may be values calculated from the load values measured and executed, or may be values calculated from theoretical values or the like.

[Operation]
FIG. 8 is a flowchart showing the operation of the control device 30. First, the switching determination unit 32 acquires the load of each analysis process from the load measurement unit 21 of each processing device 20 (step S1). Based on the load, the switching determination unit 32 determines a switching pattern representing a combination of the analysis processing to be switched and the processing device 20 that is the switching destination, and transmits the switching pattern to the switching execution unit 31 (step S2). Based on the transmitted switching pattern, the switching execution unit 31 switches the processing device 20 for analysis processing to be switched (step S3). The control device 30 repeatedly executes steps S1, S2, and S3 at predetermined intervals.

  FIG. 9 is a flowchart showing in detail an operation (step S2) in which the switching determination unit 32 determines a switching pattern representing a combination of an analysis process to be switched and a switching destination processing apparatus. The switching determination unit 32 calculates a pre-switching accuracy score based on the analysis processing load information (step S11). The accuracy score before switching is obtained by quantifying the degree of accuracy of analysis in the prediction period when switching is not performed (a in FIG. 3).

  In addition, this score and the score that will be described later are values that increase as accuracy increases. Each score may be smaller as accuracy increases. The switching determination unit 32 calculates the accuracy score before switching, for example, as follows.

  First, the switching determination unit 32 refers to the processing device resource information storage unit 35, and loads the load factor L before switching each processing device 20 from the load of each analysis processing and the value of the processing resource of each processing device 20. The load factor L before switching is calculated. Next, the switching determination unit 32 refers to the per-load accuracy prediction information storage unit 34, and for each analysis process, is a predicted value of accuracy in the prediction period when switching is not performed according to the pre-switching load factor L. Calculate the accuracy prediction value before switching. The switching determination unit 32 sets the smallest one of the predicted accuracy values before switching of each analysis process as the accuracy score before switching.

  Next, the switching determination unit 32 calculates a switching accuracy score for each switching pattern (step S12). Here, the switching accuracy score is a numerical value of the degree of accuracy of analysis processing at the time of switching and after switching in the prediction period. The switching accuracy score of the switching pattern is the minimum value in the switching accuracy score for each analysis process calculated for each analysis process. Here, the switching accuracy score for each analysis process is the time average value of the switching accuracy prediction value and the post-switching accuracy prediction value calculated for each analysis processing.

  Here, the switching determination unit 32 sets the maximum value (1.0 in the example of FIG. 4) that can be used to predict the switching accuracy of the analysis process that is not the switching target. The reason for doing so is that the analysis processing that is not the target of switching does not have accuracy degradation associated with switching. Further, as described above, the switching determination unit 32 uses the value of the switching accuracy prediction information storage unit 33 for the analysis process to be switched.

  The switching determination unit 32 calculates the post-switching accuracy prediction as follows. First, the switching determination unit 32 refers to the processing device resource information storage unit 35 and determines the load factor of each processing device 20 after switching from the load of each analysis processing and the value of the processing resource of each processing device 20. A post-switching load factor L, which is L, is calculated. At this time, the switching determination unit 32 calculates the post-switching load factor L based on the arrangement on each processing device 20 of each process after switching according to the switching pattern. Furthermore, the switching determination unit 32 refers to the load accuracy prediction information storage unit 34 and calculates a post-switching accuracy prediction according to the post-switching load factor L of the processing device 20 that is executed after switching for each analysis process.

  Next, the switching determination unit 32 refers to the switching accuracy prediction information storage unit 33 and acquires the length of the switching period. Next, the switching determination unit 32 subtracts the length of the switching period from the length of the prediction period, calculates the length of the prediction period after switching, and the time average value of the switching accuracy prediction value and the switching accuracy prediction value A switching accuracy score for each analysis process is calculated.

  In addition, about the analysis process which is not the object of switching, the switching determination part 32 is good also considering the post-switching precision prediction value as a switching precision score for every analysis process. That is, for the analysis processing that is not the switching target, the prediction period and the post-switching prediction period may be premised.

  Finally, the switching determination unit 32 selects an optimal switching pattern based on the pre-switching accuracy score and the switching accuracy score (step S13). The switching determination unit 32 selects a switching pattern as follows. First, the switching determination unit 32 selects a switching pattern having a maximum switching accuracy score. When the switching accuracy score is larger than the pre-switching accuracy score, the switching determination unit 32 determines to perform switching based on the selected switching pattern. When the switching accuracy score is equal to or lower than the pre-switching accuracy score, the switching determination unit 32 determines that switching is not performed.

[effect]
There is an analysis process that temporarily lowers the processing accuracy when the processing device 20 to be executed is switched. For example, there is an analysis process in which learning is performed again from input data after switching. The control device 30 according to the present embodiment can efficiently perform switching of the processing device 20 for such analysis processing. That is, when controlling the load by switching the processing device 20, the control device 30 enables a plurality of processing devices 20 to execute a plurality of analysis processes while suppressing a decrease in accuracy.

  The reason is that the switching determination unit 32 determines a switching pattern based on switching accuracy prediction information that represents a change in accuracy due to switching.

  Such a control device 30 is particularly effective when the analysis processing is stateful and one analysis processing cannot be executed by being distributed to a plurality of processing devices 20. Further, the analysis process is not limited to the analysis process. The control system 50 according to the present embodiment is intended for general analysis processing that temporarily reduces processing accuracy when the processing device 20 to be executed is switched.

[Second Embodiment]
[Constitution]
FIG. 10 is a configuration diagram of the control system 50 according to the present embodiment. The processing device 20 according to the present embodiment is different from the first embodiment in that an environment analysis process is also executed in addition to the analysis process. The processing apparatus 20-a in FIG. 10 shows an example in which an analysis process a and an environment analysis process a, and an analysis process b and an environment analysis process b are executed. A processing device 20-b in FIG. 10 shows an example in which an analysis process c and an environment analysis process c are executed.

  The analysis process and the environment analysis process constitute a pair as a switching unit, and the environment analysis process receives data from the data distribution apparatus 10 associated with the paired analysis process, executes the environment analysis, A certain environmental condition E is output to the switching determination unit 32. Here, the environmental condition E is a feature quantity that affects the accuracy of the analysis process, such as the illumination condition of the video data and the number of persons in the video data.

  Further, the control device 30 according to the present embodiment is different from the first embodiment in that it includes a switching accuracy prediction information storage unit 36 for each environment instead of the switching accuracy prediction information storage unit 33.

  Other components are the same as those in the first embodiment. Constituent elements similar to those in the first embodiment are denoted by the same reference numerals as those in FIG. 1, and detailed description thereof is omitted.

  The switching determination unit 32 acquires the environmental condition E from the environmental analysis process, calculates the switching accuracy prediction value according to the environmental condition E, and then combines the analysis process to be switched with the switching destination processing apparatus. A switching pattern representing is determined.

  The switching accuracy prediction information storage unit 36 for each environment stores accuracy switching accuracy prediction information for each analysis process. The environmental switching accuracy prediction information includes switching accuracy prediction depending on the environmental condition E and data indicating the switching period. FIG. 11 shows an example of the environment switching accuracy prediction information stored in the environment switching accuracy prediction information storage unit 36.

  The switching time prediction in the example of FIG. 11 is expressed using a function with the environmental condition E as a parameter as a relative value when the maximum value of accuracy is 1.0, when the switching is performed. Yes. For example, when a person in the video data takes an abnormal action, the analysis process detects the abnormal action and issues an alert. When the number of persons is 0, abnormal behavior is not detected and the correct alert rate does not change by switching. If the number of persons in the video data increases, the number of persons who take abnormal actions will increase probabilistically, and the alert rate of correct alerts will drop by switching. That is, as exemplified in the switching accuracy prediction column of the table of FIG. 11, the switching accuracy prediction information can be expressed as a function whose accuracy decreases as the environmental condition E, ie, the number of people in the video data increases. .

  The environment switching accuracy prediction information is not limited to the example illustrated in FIG. 11, and may be expressed using other functions or table formats. The coefficients and tabular values of these functions may be values calculated from accuracy values measured by executing switching, or values calculated from theoretical values or the like.

  The environmental condition E is not limited to the number of persons in the video data. The environmental condition E may be, for example, the intensity of movement of a person in the video data. When there is almost no movement, since the person in the video data is almost stationary, the probability that the person enters the prohibited area for a certain period thereafter is low. On the other hand, when the movement is intense, the movement of the person in the video data has occurred, and the probability that the person will enter the prohibited area in a certain period thereafter is relatively high. In the former case, even if switching is performed, the alert rate of the correct alert does not decrease much. In the latter case, if the switching is performed, there is a high possibility that the alert rate of the correct alert is lowered. Therefore, the switching accuracy prediction can be expressed as a function whose accuracy decreases in accordance with the degree of intenseness of motion between video frames.

[Operation]
FIG. 12 is a flowchart showing the operation of the control device 30 according to the present embodiment. Also, FIG. 9 used in the description of the first embodiment is an operation in which the switching determination unit 32 determines a switching pattern representing a combination of an analysis process to be switched and a switching destination processing apparatus (step S2). ) In detail. In the following description, details of operations similar to those of the first embodiment are omitted.

  The operation of the control device 30 according to the second embodiment is different from that of the first embodiment in the following points. The first difference is that the switching determination unit 32 acquires the environmental condition E from each environmental analysis process (step S4). The second difference is that the switching determination unit 32 calculates a switching accuracy predicted value according to the environmental condition E and then calculates a switching accuracy score for each switching pattern (step S12). . The third difference is that the switching execution unit 31 switches the environmental analysis process as well as the analysis process (step S3).

  The switching determination unit 32 calculates a switching accuracy score for each switching pattern according to the environmental condition E as follows. First, the switching determination unit 32 refers to the environment-specific switching accuracy prediction information storage unit 36 based on the value of the environmental condition E acquired from each environment analysis process, and calculates the switching accuracy prediction value of each analysis process. . Thereafter, the switching determination unit 32 calculates a switching accuracy score for each switching pattern in the same manner as the operation in the first embodiment.

  Based on the switching pattern, the switching execution unit 31 switches the processing device 20 for the analysis processing to be switched and the environment analysis processing that forms a pair with the analysis processing. For example, when switching from the processing device 20-a to the processing device 20-b for the analysis processing b, the switching determination unit 32 also performs the environmental analysis processing b for the processing device 20-a to the processing device 20-b. Switch to.

[effect]
The control device 30 according to the second embodiment switches while suppressing a decrease in accuracy even when the change in accuracy when the processing device 20 that executes the analysis processing is switched varies depending on the content of the data to be analyzed. Can be executed efficiently.

  The reason is that the switching determination unit 32 calculates the switching accuracy prediction value based on the environmental condition E according to the content of the data output by the environment analysis process, and uses it for calculating the switching accuracy score.

[Third embodiment]
[Constitution]
FIG. 13 is a configuration diagram of the control device 30 according to the present embodiment. The control device 30 according to the present embodiment is different from the first embodiment in that an accuracy request information storage unit 37 is provided. Other components are the same as those in the first embodiment. Constituent elements similar to those in the first embodiment are denoted by the same reference numerals as those in FIG. 1, and detailed description thereof is omitted.

  The accuracy request information storage unit 37 stores user accuracy requirements for each analysis process. FIG. 14 shows an example of a user accuracy request stored in the accuracy request information storage unit 37.

  The example of FIG. 14 represents the accuracy request as a relative value when the maximum value is 1.0. For example, when a person in the video data takes an abnormal action, the analysis process detects the abnormal action and issues an alert. If the user wants all possible alerts to be raised, in this example, the user sets the accuracy requirement to 1.0. If the alert should be issued with a probability of 4/5, in this example, the user sets the accuracy requirement to 0.8. The accuracy request information is not limited to the example shown in FIG. 16, and may be stored using other values and formats.

[Operation]
The operation of the control device 30 according to the present embodiment will be described with reference to the flowcharts of FIGS. 8 and 9 used in the description of the first embodiment. The operation of the control device 30 according to the present embodiment differs from the operation of the first embodiment in the following points. Detailed description of operations similar to those of the first embodiment is omitted.

  The difference is that the switching determination unit 32 calculates the accuracy score before switching in consideration of the accuracy request (step S11) and calculates the switching accuracy score (step S12).

  The switching determination unit 32 calculates the pre-switching accuracy score as follows (step S11). First, the switching determination unit 32 refers to the processing device resource information storage unit 35, and loads the load factor L before switching each processing device 20 from the load of each analysis processing and the value of the processing resource of each processing device 20. Is calculated before switching.

  Next, the switching determination unit 32 refers to the per-load accuracy prediction information storage unit 34, and for each analysis process, according to the pre-switching load factor L, the pre-switching accuracy prediction representing the predicted value of accuracy before switching. Calculate the value. Next, the switching determination unit 32 refers to the accuracy request information storage unit 37, and calculates, for each analysis process, a pre-switching accuracy requirement satisfaction rate that represents the degree to which the accuracy requirement before switching is satisfied using the following equation: To do.

Pre-switching accuracy requirement fulfillment rate = pre-switching accuracy predicted value / accuracy request The switching determination unit 32 sets the minimum one of the pre-switching accuracy requirement satisfaction rates for each analysis process as the pre-switching accuracy score of the switching pattern.

  Moreover, the switching determination part 32 calculates a switching precision score as follows (step S12). The switching determination unit 32 sets the minimum one of the switching accuracy scores for each analysis process calculated for each analysis process as the switching accuracy score of the switching pattern. The switching determination unit 32 uses the time average of the switching accuracy required satisfaction rate and the post-switching accuracy required satisfaction rate calculated for each analysis process as the switching accuracy score for each analysis process.

  Here, the switching determination unit 32 calculates the switching accuracy requirement satisfaction rate as follows. First, the switching determination unit 32 calculates a switching accuracy prediction value for each analysis process in the same manner as in the first embodiment. Next, the switching determination unit 32 refers to the accuracy requirement information storage unit 37 and calculates a switching-time accuracy requirement satisfaction rate that represents the degree of satisfaction of the accuracy requirement at the time of switching for each analysis process using the following equation: To do.

Switching accuracy requirement satisfaction rate = switching accuracy predicted value / accuracy requirement The switching determination unit 32 calculates the post-switching accuracy requirement satisfaction rate as follows. First, the switching determination unit 32 calculates a post-switching accuracy prediction value for each analysis process in the same manner as in the first embodiment. Next, the switching determination unit 32 refers to the accuracy request information storage unit 37, and uses each of the following processes to calculate a post-switching accuracy requirement satisfaction rate that represents the degree of satisfaction of the accuracy requirement after switching. calculate.

Post-switching accuracy requirement satisfaction rate = post-switching accuracy prediction value / accuracy request
[effect]
The control device 30 according to the third embodiment can select an optimal switching pattern so as to satisfy the accuracy requirement of the user as much as possible.

  The reason is that the switching determination unit 32 calculates the accuracy satisfaction rate before switching, at the time of switching, and after switching based on the accuracy request stored in the accuracy request information storage unit 37, and calculates the accuracy score before switching and the switching accuracy score. It is because it uses for.

[Fourth embodiment]
FIG. 28 is a configuration diagram of the control device 30 according to the present embodiment. The control device 30 includes a switching determination unit 32.

  The switching determination unit 32 performs switching target analysis processing when one of a plurality of analysis processes (hereinafter referred to as switching target analysis processing) being executed by the first processing device 20 is switched to the second processing device 20 and executed. A switching accuracy score indicating accuracy is calculated. The switching determination unit 32 calculates the switching accuracy score based on the accuracy predicted value during the switching process and the accuracy predicted value after the switching is completed, and based on the calculated switching accuracy score, the switching target analysis process Decide whether to perform switching.

  In the present embodiment, the elements corresponding to the switching execution unit 31 in the first embodiment are provided in other devices not shown. The control device 30 according to the present embodiment outputs the determined result to the other device. Moreover, in this Embodiment, the switching determination part 32 has memorize | stored in the inside the precision prediction value in switching process, and the precision prediction value after switching completion as a setting parameter.

  There is an analysis process that temporarily lowers the processing accuracy when the processing device 20 to be executed is switched. For example, there is an analysis process in which learning is performed again from input data after switching. The control device 30 according to the present embodiment can efficiently perform switching of the processing device 20 for such analysis processing. That is, when controlling the load by switching the processing device 20, the control device 30 enables a plurality of processing devices 20 to execute a plurality of analysis processes while suppressing a decrease in accuracy.

  The reason is that the switching determination unit 32 determines to execute switching based on accuracy prediction information during switching processing that represents a change in accuracy due to switching.

  As mentioned above, each embodiment described so far is an example of the embodiment of the present invention, and the scope of the present invention is not limited only to the above-described embodiment, and various modifications can be made without departing from the gist of the present invention. It is possible to implement in a form that has been modified.

  For example, the load measurement unit 21, the execution unit 22, and the switching execution unit 31 and the switching determination unit 32 of the control device 30 can be realized by a logic circuit including firmware. The switching accuracy prediction information storage unit 33, the load accuracy prediction information storage unit 34, the processing device resource information storage unit 35, the environment switching accuracy prediction information storage unit 36, and the accuracy requirement information storage unit 37 of the control device 30 are: For example, it can be realized by a magnetic disk device.

  Further, the processing device 20 or the control device 30 can be realized by a computer device 40 including a program 43.

  FIG. 15 is a configuration diagram of the computer device 40. The computer device 40 includes a processor 41, a main storage unit 42, and an auxiliary storage device 44 that are connected to each other via a bus 45. Here, for example, the main storage unit 42 is a semiconductor memory, and the auxiliary storage device 44 is a magnetic disk device. The main storage unit 42 stores a program 43.

  The program 43 that implements the processing device 20 is executed by the processor 41, thereby causing the processor 41 to function as the load measurement unit 21 and the execution unit 22, and causing the processor 41 to perform analysis processing.

  In addition, the program 43 that implements the control device 30 is executed by the processor 41, thereby causing the processor 41 to function as the switching execution unit 31 and the switching determination unit 32. In this case, the switching accuracy prediction information storage unit 33, the load accuracy prediction information storage unit 34, the processing device resource information storage unit 35, the environment switching accuracy prediction information storage unit 36, and the accuracy request information storage unit of the control device 30. 37 can be realized by the auxiliary storage device 44.

[Specific example of the first embodiment]
FIG. 16 is a diagram illustrating a specific example of the configuration of the control system 50 according to the first embodiment.

  In this specific example, the data distribution device 10 is a camera 10e connected to a network, and inputs captured moving image data to the processing device 20.

  In this specific example, the processing device 20 is realized by the computer device 40. The execution unit 22 is an execution control program 22e, and activates, stops, and mediates communication of each analysis program.

  The analysis process a, the analysis process b, and the analysis process c shown in FIG. 1 are executed by the analysis program a, the analysis program b, and the analysis program c in FIG. 16, respectively, and the camera 10e-a, the camera 10e-b, and the camera 10e. The moving image data acquired from -c is analyzed. Each analysis program performs facial feature extraction, person tracking, and crowd behavior analysis, and outputs the analysis result to the control device 30 through network communication. Each analysis program outputs the average processing time for each frame in the latest fixed time to the load measurement program 21e by inter-program communication.

  The load measurement unit 21 is a load measurement program 21e. The load measurement program 21e acquires the average processing time for each frame of each analysis program and calculates the load for each analysis program. Here, the load for each analysis program is the number of cores of the processor required for each analysis program, and the load measurement program 21e calculates the load by the following equation.

Load per analysis program = frame rate x processing time per frame For example, if the frame rate is 10 frames / second and the processing time per frame is 0.2 seconds, the analysis program consumes 2 processor hours per second. Therefore, the required number of cores is 2.

  The load measurement program 21e outputs the load for each analysis program to the control device 30 through network communication.

  The execution control program 22e, each analysis program, and the load measurement program 21e are part of a program 43 provided in the computer device 40 that implements the processing device 20.

  In this specific example, the control device 30 is realized by a computer device 40 different from the processing device 20.

  The switching determination unit 32 is a switching determination program 32e. The switching determination program 32e acquires the load of each analysis program from the load measurement program 21e, and determines a switching pattern representing a combination of the analysis program to be switched and the switching destination processing device 20. Thereafter, the switching determination program 32e outputs the determination result to the switching execution program 31e by inter-program communication.

  The switching execution unit 31 is a switching execution program 31e, and receives an instruction from the switching determination program 32e and switches processing programs.

  In this specific example, the switching accuracy prediction information storage unit 33, the load accuracy prediction information storage unit 34, and the processing device resource information storage unit 35 are auxiliary storage devices 44 of the computer device 40, and store various information in a database. doing.

  The switching execution program 31e and the switching determination program 32e are part of the program 43 provided in the computer device 40 that implements the control device 30.

[Operation]
Next, the overall operation of the control system 50 of this specific example will be described with reference to FIGS.

  The switching determination program 32e acquires the load of each analysis program from the load measurement program 21e by network communication (step S1). Based on the load, the switching determination program 32e determines a switching pattern representing a combination of the analysis program to be switched and the switching destination processing device, and transmits the switching pattern to the switching execution program 31e by inter-program communication (step S2). The switching execution program 31e switches the processing device of the analysis program to be switched based on the switching pattern (step S3).

  Here, the switching operation will be described in detail. First, the switching execution program 31e instructs the execution control program 22e of the original processing device 20 to stop execution of the analysis program to be switched by network communication. In the original processing device 20, the execution control program 21e stops the execution of the analysis program and notifies the load measurement program 21e of the stop. Next, the switching execution program 31e moves the analysis program from the original processing device 20 to the switching destination processing device 20.

  Here, the movement operation will be described in detail. The switching execution program 31e transfers a program, setting information, and the like necessary for executing the analysis program from the original processing device 20 to the switching destination processing device 20. In the processing apparatus 20 at the switching destination, the execution control program 22e sets up the analysis apparatus so that the processing program 20 can execute the analysis program.

  Note that a program, setting information, and the like necessary for executing the analysis program may be arranged in advance in each processing device 20 and set up in an executable state. In this case, the switching execution program 31e performs the movement operation. Is omitted.

  Next, the switching execution program 31e instructs the execution control program 22e of the switching destination processing apparatus 20 to execute the analysis program. In the switching destination processing apparatus, the execution control program 22e starts the execution of the analysis program and notifies the load measurement program 21e of the start. The control device 30 repeatedly executes steps S1, S2, and S3 at predetermined intervals.

  FIG. 9 is a flowchart showing in detail an operation (step S2) in which the switching determination program 32e determines a switching pattern representing a combination of the analysis program to be switched and the processing apparatus 20 that is the switching destination.

  FIGS. 17, 18, and 19 exemplify values of switching accuracy prediction information, load-specific accuracy prediction information, and processing device resource information in this specific example. FIG. 20 is a table illustrating the load of each analysis program output by the load measurement program 21e. FIG. 21 is a table showing specific calculated values of the pre-switching accuracy prediction value and the pre-switching accuracy score based on the values exemplified in FIGS. 17 to 20. FIG. 22 is a table showing specific calculated values of the switching accuracy predicted value, the post-switching accuracy predicted value, and the switching accuracy score for each switching pattern based on the values illustrated in FIGS. 19 to 20. Here, the operation in which the switching determination program 32e determines the switching pattern will be described using specific numerical values.

  The switching determination program 32e calculates a pre-switching accuracy score based on the analysis load information (step S11). Here, the accuracy score before switching is obtained by quantifying the degree of accuracy of the analysis processing in the prediction period when switching is not performed.

  The calculation of the accuracy score before switching will be specifically described. Here, the configuration before switching is as shown in FIG. Referring to FIG. 16, before switching, the analysis program a and the analysis program b are executed by the processing device 20-a. As shown in FIG. 19, the processing resource of the processing device 20-a is 4, and as shown in FIG. 20, the loads of the analysis program a and the analysis program b are 2.5, respectively. The load factor L of a is 1.25 as follows.

(2.5 + 2.5) /4=1.25
Similarly, the load factor L of the processing device 20-b is 0.500. From the value of the accuracy prediction information for each load in FIG. 18, the accuracy prediction value before switching of the analysis program a is 0.375 as follows.

1-0.5 × 1.25 = 0.375
Similarly, the predicted accuracy values before switching of the analysis program b and the analysis program c are 0.875 and 1.00. The accuracy score before switching is 0.375, which is the smallest of the predicted accuracy values before switching of each analysis processing program.

  Next, the switching determination program 32e calculates a switching accuracy score for each switching pattern (step S12). Here, the switching accuracy score is a numerical value of the degree of accuracy of analysis at the time of switching and after switching in the prediction period.

  The calculation of the switching accuracy score will be specifically described. Here, in FIG. 22, a switching pattern P1 indicates a switching pattern for switching the analysis program a to the processing device 20-b from the configuration shown in FIG. The switching pattern P2 indicates a switching pattern for switching the analysis processing program b to the processing device 20-b from the configuration illustrated in FIG. The switching pattern P3 indicates a switching pattern for switching the analysis program c to the processing device 20-a from the configuration illustrated in FIG.

  In the switching pattern P1, the analysis program a is a switching target, and the switching accuracy prediction value is 0.800 as shown in FIG. 22 from the switching accuracy prediction information of FIG. The analysis program b and the analysis program c are not switching targets, and the switching accuracy prediction value is 1.00 respectively.

  Further, after switching in the switching pattern P1, the processing device 20-a executes the analysis program b, and the processing program 20-b executes the analysis program a and the analysis program c. The post-switching load factor L of the processing device 20-a is 0.625 as follows, referring to the processing device resource information in FIG. 19 and the load value in FIG.

2.5 / 4 = 0.625
Similarly, the post-switching load factor L of the processing device 20-b is 1.13.

  From the accuracy prediction information for each load in FIG. 18, the post-switching accuracy prediction value of the analysis program a is 0.438 as follows.

1-0.5 × 1.13 = 0.438
Similarly, the post-switch accuracy prediction values of the analysis program b and the analysis program b are 1.00 and 0.719.

  When the prediction period is 20 seconds, the switching accuracy score for each analysis process of the analysis program a is 10 seconds because the switching period is 10 seconds from FIG. 17, and the prediction period after switching is 10 seconds. 0.618.

0.800 × (10/20) + 0.438 × (10/20) = 0.618
Similarly, the switching accuracy score for each analysis process of the analysis programs b and c is 1.00 and 0.859.

  Therefore, in the switching pattern P1, the switching accuracy score is 0.618 which is the minimum of the switching accuracy scores for each analysis process. Similarly, the switching accuracy scores of the switching patterns P2 and P3 are calculated as 0.543 and 0.500 as shown in FIG.

  Next, the switching determination program 32e selects an optimal switching pattern based on the pre-switching accuracy score and the switching accuracy score (step S13).

  The operation for determining the switching pattern will be specifically described with reference to FIGS. Referring to FIG. 22, the switching pattern having the maximum switching accuracy score is P1, and the switching accuracy score is 0.618. Referring to FIG. 21, the pre-switching accuracy score is 0.375, and the switching accuracy score is larger than the pre-switching accuracy score. Therefore, the switching determination program 32e determines the switching pattern P1 as an optimal switching pattern, and instructs the switching execution program 31e to switch.

  Referring to FIG. 22, from the post-switching load factor L, the degree of load bias after switching is similarly low in switching patterns P1 and P2, and high in switching pattern P3. That is, the load after switching is leveled in the switching patterns P1 and P2. From the predicted accuracy value at the time of switching, the switching pattern P1 has the smallest deterioration in accuracy at the time of switching. That is, in this specific example, the switching determination program 32e can select the switching pattern P1 that efficiently executes a plurality of analysis programs in the plurality of processing devices 20 while suppressing a decrease in accuracy.

[Specific example of the second embodiment]
FIG. 23 is a diagram illustrating a specific example of the configuration of the control system 50 according to the second embodiment. The control system 50 of this specific example is different from the specific example of the first embodiment in that the processing device 20 includes an environment analysis program. The environment analysis process a, the environment analysis process b, and the environment analysis process c shown in FIG. 10 are executed by the environment analysis program a, the environment analysis program b, and the environment analysis program c shown in FIG. The environment analysis program a, the environment analysis program b, and the environment analysis program c constitute a pair with the analysis program a, the analysis program b, and the analysis program c, respectively.

  In addition, the control device 30 in this specific example is different from the specific example of the first embodiment in that it includes an environment-specific switching accuracy prediction information storage unit 36. In this specific example, the switching accuracy prediction information storage unit 36 for each environment is an auxiliary storage device 44 of the computer device 40 and stores information in a database. Other components are the same as the specific example of the first embodiment. The same components as those in the specific example of the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

  The processing device 20-a includes an environment analysis program a and an environment analysis program b. The processing device 20-b includes an environment analysis program c. The environment analysis program a, the environment analysis program b, and the environment analysis program c acquire moving image data from each of the cameras 10-a, 10-b, and 10-c through network communication. Each environmental analysis program analyzes the number of persons E in the video data as the environmental condition, and outputs the analysis result to the switching determination program 32e by network communication.

  The switching determination program 32e obtains the number of persons E from each environment analysis program, calculates the switching accuracy prediction value according to the number of persons E, and then switches between the analysis program to be switched and the switching destination processing apparatus. A switching pattern representing the combination is determined.

  FIG. 24 shows switching accuracy prediction information for each environment stored in the switching accuracy prediction information storage unit 36 for each environment. The accuracy prediction information at the time of switching for each environment uses a function corresponding to the number of persons E in the video data as a relative value when the accuracy is predicted when the maximum value is 1.0. It expresses. For example, when a person in moving image data takes an abnormal action, the analysis processing detects the abnormal action and issues an alert. When the number of persons E is 0, abnormal behavior is not detected and the correct alert rate does not change by switching. If the number of persons E in the moving image data increases, the number of persons who take an abnormal action will increase probabilistically, and the correct alert rate will decrease by switching. That is, as exemplified in the switching accuracy prediction column of the table of FIG. 24, the switching accuracy prediction information can be expressed as a function whose accuracy decreases with an increase in the environmental condition of the number of persons E in the video data. it can.

[Operation]
Next, the overall operation of the control system 50 of this specific example will be described with reference to the flowchart of FIG. Detailed descriptions of operations similar to those of the specific example of the first embodiment are omitted.

  The operation of this example is different from the operation of the example of the first embodiment in the following points. The first difference is that the switching determination program 32e acquires environmental conditions from each environmental analysis program (step S4). The second difference is that the switching determination program 32e calculates a switching accuracy predicted value in accordance with the environmental conditions and then calculates a switching accuracy score for each switching pattern (step S12). The third difference is that the switching execution program 31e switches the environmental analysis program paired with the target analysis program in addition to the target analysis program (step S3).

  The switching determination program 32e calculates a switching accuracy score for each switching pattern as follows according to environmental conditions. First, the switching determination program 32e refers to the switching accuracy prediction information for each environment based on the value of the number of persons E, which is an environmental condition acquired from each environmental analysis program, and calculates the switching accuracy prediction value of each analysis program. To do. Here, assuming that the number of persons E obtained by analysis by each environment analysis program is one, the accuracy prediction value at the time of switching of the analysis program a is 0.8 as follows.

1.0−0.2 × 1 = 0.8
Similarly, the predicted accuracy values at the time of switching between the analysis program b and the analysis program c are 0.2 and 0.5.

  Next, the switching determination program 32e calculates a switching accuracy score for each switching pattern in the same manner as the operation in the specific example of the first embodiment. When the switching accuracy score is calculated in the same manner as the specific example of the first embodiment, the switching accuracy score has the values shown in the table of FIG. The switching determination program 32e determines the switching pattern P1 as the optimal switching pattern, as in the specific example of the first embodiment.

  The switching execution program 31e switches from the processing device 20-a to the processing device 20-b for the analysis program a to be switched and the environment analysis program a that forms a pair with the analysis program a (step S3).

  In this specific example, the switching determination program 32e does not hold the switching accuracy predicted value as a fixed value, but dynamically generates it according to environmental conditions and uses it for selecting a switching pattern. For this reason, the switching determination program 32e can efficiently execute a plurality of analysis programs with the plurality of processing devices 20 while suppressing a decrease in accuracy even when the accuracy when switching varies depending on the contents of data to be analyzed. The switching pattern P1 to be executed can be selected.

[Specific example of the third embodiment]
The control device 30 of this specific example differs from the specific example of the first embodiment in that it includes an accuracy request information storage unit 37. In this specific example, the accuracy request information storage unit 37 is an auxiliary storage device 44 of the computer device 40 and stores information in a database. Other components are the same as the specific example of the first embodiment. The same components as those in the specific example of the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

[Operation]
The operation of the control system 50 of this example is different from the operation of the example of the first embodiment in the following points. The difference is that the switching determination program 32e calculates an accuracy score before switching in consideration of accuracy request information (step S11 in FIG. 9), and calculates a switching accuracy score (step S12).

  The operation of the switching determination program 32e determining the switching pattern will be described using specific numerical values. FIG. 25 shows accuracy request values for each analysis program included in the accuracy request information. FIG. 26 is a table showing specific calculated values of the pre-switching accuracy requirement satisfaction rate and the pre-switching accuracy score based on the accuracy requirements shown in FIG. FIG. 27 is a table showing specific calculated values of the switching accuracy required satisfaction rate, the post-switching accuracy required satisfaction rate, and the switching accuracy score.

  The calculation of the pre-switching accuracy score in step S11 will be specifically described. Here, the configuration before switching is as shown in FIG. Similarly to the specific example of the first embodiment, the pre-switching accuracy prediction value is calculated as 0.875 for the analysis program b. Here, referring to the value of the accuracy requirement information in FIG. 25, the accuracy requirement of the analysis program b is 0.4, and the accuracy requirement satisfaction rate before switching is 2.19 as follows.

0.875 / 0.4 = 2.19
Similarly, the required accuracy rates before switching between the analysis program a and the analysis program c are 0.375 and 1.25, respectively. In this situation (FIG. 26), the pre-switching accuracy score is calculated as 0.375, which is the minimum value of the pre-switching accuracy required satisfaction rate.

  The calculation of the switching accuracy score in step S12 in FIG. 9 will be specifically described. Here, in FIG. 27, a switching pattern P1 indicates a switching pattern for switching the analysis program a to the processing device 20-b from the configuration shown in FIG. The switching pattern P2 indicates a switching pattern for switching the analysis program b to the processing device 20-b from the configuration illustrated in FIG. The switching pattern P3 indicates a switching pattern for switching the analysis program c to the processing device 20-a from the configuration illustrated in FIG.

  In the switching pattern P1, similarly to the specific example of the first embodiment, the switching accuracy prediction value is calculated as 1.00 for the analysis program b. Here, referring to the value of the accuracy requirement information shown in FIG. 25, the accuracy requirement of the analysis program b is 0.4, and the accuracy requirement satisfaction rate at the time of switching is 2.50 as shown below.

1.00 / 0.4 = 2.50
Similarly, the accuracy requirement satisfaction rate at the time of switching between the analysis program a and the analysis program c0 is 0.800 and 1.25, respectively.

  In addition, in the switching pattern P1, the post-switching accuracy prediction value is calculated as 1.00 for the analysis program b, as in the specific example of the first embodiment. Here, referring to the value of the accuracy requirement information shown in FIG. 25, the accuracy requirement of the analysis program b is 0.4, and the switching accuracy requirement satisfaction rate is 2.50 as follows.

1.00 / 0.4 = 2.50
Similarly, the required accuracy rates after switching between the analysis program a and the analysis program c are 0.438 and 0.898, respectively.

  From FIG. 17, the switching period of the analysis program b is 10 seconds. Therefore, when the prediction period is 20 seconds, the post-switching prediction period is 10 seconds, and when the time average is taken, the switching accuracy score for each analysis process is 2.50 as follows.

2.50 × (10/20) + 2.50 × (10/20) = 2.50
Similarly, the switching accuracy scores for each analysis process of the analysis program a and the analysis program c are 0.619 and 1.07. Therefore, in the switching pattern P1, the switching accuracy score is 0.619 which is the minimum value among the switching accuracy scores for each analysis process. Similarly, the switching accuracy scores of the switching patterns P2 and P3 are calculated as 1.00 and 0.500 as shown in FIG.

  22 and 27 are compared, the switching determination program 32e selects the arrangement pattern P1 in FIG. 22 in the specific example of the first embodiment, and selects the arrangement pattern P2 in FIG. 27 in this specific example. That is, in this specific example, by using the accuracy request information, the switching determination program 32e can select a switching pattern in accordance with the accuracy requirement of the user.

DESCRIPTION OF SYMBOLS 10 Data distribution apparatus 10-a Data distribution apparatus 10-b Data distribution apparatus 10-c Data distribution apparatus 10e Camera 10e-a Camera 10e-b Camera 10e-c Camera 20 Processing apparatus 20-a Processing apparatus 20-b Processing apparatus 21 Load measurement unit 21e Load measurement program 22 Execution unit 22e Execution control program 30 Control device 31 Switching execution unit 31e Switching execution program 32 Switching determination unit 32e Switching determination program 33 Switching accuracy prediction information storage unit 34 Load accuracy prediction information storage unit 35 Resource information storage unit 36 Accuracy prediction information storage unit upon switching for each environment 37 Accuracy requirement information storage unit 40 Computer device 41 Processor 42 Main storage unit 43 Program 44 Auxiliary storage device 45 Bus 50 Control system

Claims (7)

A switching accuracy score indicating the accuracy of the switching target analysis process when one of a plurality of analysis processes (hereinafter referred to as switching target analysis process) being executed in the first processing apparatus is switched to the second processing apparatus and executed. , Calculated based on the predicted accuracy value during the switching process and the predicted accuracy value after the switching is completed, and determines whether or not to switch the switching target analysis process based on the calculated switching accuracy score Switching decision means to
Switching accuracy prediction information storage means for storing switching accuracy prediction information for calculating the accuracy prediction value during the switching processing based on the environmental condition value of the analysis processing,
The switching determination means acquires the environmental condition value from the first processing device, and calculates a predicted value during the switching process using the switching accuracy prediction information.
Control device. The accuracy requirements information storage means for storing the accuracy requirements of the switching target analysis, further comprising,
Said switching determining means, said switching accuracy score, fulfillment rate for the accuracy requirements of precision prediction values in the switching process, and is calculated based on the fill rate for the accuracy requirements of precision prediction value after completion of switching, wherein Item 2. The control device according to Item 1 . Resource information storage means for storing the processing resource amount of the second processing device;
Switching execution means for executing switching from the first processing device to the second processing device in the switching target analysis processing,
The switching determination unit acquires the processing amount of the switching target analysis process from the first processing device, and completes switching of the switching target analysis processing from the processing amount and the processing resource amount of the second processing device. The control device according to claim 1 or 2, which calculates a subsequent accuracy prediction value. A control device according to any one of claims 1 to 3 ;
The first processing device;
A control system including one or more of the second processing devices that execute a plurality of the analysis processes,
The switching determination means a1) the switching of the switching target analysis process for each of all combinations for switching any of the analysis processes being executed by the processing apparatus in the control system to any other processing apparatus. An accuracy score is calculated, and a2) for the analysis processing other than the switching target analysis processing, the switching accuracy score indicating the accuracy of the analysis processing after the switching of the switching target analysis processing is completed, and the accuracy prediction after the switching is completed B) A control system that calculates based on a value, and b) selects a combination having the highest accuracy score during the analysis process and decides to switch the switching target analysis process according to the combination. A switching accuracy score indicating the accuracy of the switching target analysis process when one of a plurality of analysis processes (hereinafter referred to as switching target analysis process) being executed in the first processing apparatus is switched to the second processing apparatus and executed. , Calculated based on the predicted accuracy value during the switching process and the predicted accuracy value after the switching is completed, and determines whether or not to switch the switching target analysis process based on the calculated switching accuracy score And
Storing the switching accuracy prediction information for calculating the accuracy prediction value during the switching process based on the environmental condition value of the analysis processing in the switching accuracy prediction information storage means;
A method of obtaining the environmental condition value from the first processing device and calculating a predicted value during the switching process using the switching accuracy prediction information . Store the accuracy requirement of the switching target analysis processing in accuracy requirement information storage means,
6. The method according to claim 5, wherein the switching accuracy score is calculated based on a satisfaction rate of the accuracy prediction value during the switching process with respect to the accuracy requirement and a satisfaction rate of the accuracy prediction value after the switching is completed with respect to the accuracy requirement. A switching accuracy score indicating the accuracy of the switching target analysis process when one of a plurality of analysis processes (hereinafter referred to as switching target analysis process) being executed in the first processing apparatus is switched to the second processing apparatus and executed. , Calculated based on the predicted accuracy value during the switching process and the predicted accuracy value after the switching is completed, and determines whether or not to switch the switching target analysis process based on the calculated switching accuracy score Switch decision process to
The computer having the switching accuracy prediction information storage means for storing the switching accuracy prediction information for calculating the accuracy prediction value during the switching processing based on the environmental condition value of the analysis processing is executed,
Further, the computer acquires the environmental condition value from the first processing device, and calculates the predicted value during the switching process using the switching accuracy prediction information, and executes the switching determination process. program.

Images