JP7572170B2 - Data compression system and data compression method - Google Patents

Description

The present invention relates to a data compression evaluation system that evaluates data compression methods suitable for situation analysis techniques.

High-precision cameras, LiDAR, and other sensors have become available at low cost, making it possible to obtain high-quality data in a variety of situations. Meanwhile, in terms of analysis technology, the improvement in the numerical calculation capabilities of GPUs has made it possible to perform a large number of matrix operations. These advances in data acquisition and analysis technology have significantly improved analysis accuracy. However, due to capacity limitations of recording devices and communication terminals, data obtained from the sensor recording terminal is compressed before being recorded and transmitted, which can lead to a decrease in the accuracy of subsequent analysis. One of the causes of the decrease in analysis accuracy is that, although important features in the data differ depending on the analysis method, data is compressed uniformly without considering importance. Therefore, it is necessary to compress data in a way that leaves as much data as possible that can suppress a decrease in analysis estimation accuracy.

Patent document 1 (JP 2017-91234 A) describes a vehicle control device that is mounted on a vehicle and controls the vehicle, and is characterized in that it includes a diagnosis unit that diagnoses the processing results of a program provided in the vehicle control device, an information extraction unit that extracts information from the diagnosis results by the diagnosis unit so as to reduce data volume, and a transmission unit that transmits the information extracted by the information extraction unit to a center provided outside the vehicle (see claim 1).

In addition, Patent Document 2 (JP 2018-186391 A) describes a transfer device having a receiving unit that receives data from a sensor device, a selecting unit that selects a conversion method to be applied to the received data from among a plurality of conversion methods for reducing the amount of data transferred to a server device, a converting unit that applies the conversion method selected by the selecting unit to the received data, a transmitting unit that transmits the converted data to the server device, and a first estimating unit that estimates a relationship between the conversion method and the load condition or the analysis accuracy based on the load condition of a network used to transfer data when a previously selected conversion method was applied, or a history of the analysis accuracy of the data by the server device when the conversion method was applied, and the selecting unit selects the conversion method to be applied to the received data based on the relationship (see claim 1).

JP 2017-91234 A JP 2018-186391 A

Dijk, Tom van, and Guido C. H. E. de Croon. 2019. "How Do Neural Networks See Depth in Single Images?" arXiv. http://arxiv.org/abs/1905.07005.

The vehicle control device described in Patent Document 1 has a diagnosis unit that diagnoses the processing results of a program installed in the vehicle control device, and an information extraction unit that extracts information from the diagnosis results by the diagnosis unit to reduce the data volume, but is limited to diagnosing the processing results of the program installed in the vehicle control device and extracting data only when there is an abnormality, and does not consider how to deal with sensor data such as images that cannot be determined as normal or abnormal and should be extracted. Furthermore, high-precision images and point cloud data required to improve analysis accuracy tend to be large in size, and compression is essential to store data for a certain period of time, but such compression is not assumed.

In the transfer device described in Patent Document 2, the model estimation unit uses a combination of past load conditions, analysis accuracy, and data conversion policies stored in the information storage unit to estimate a model of the relationship between the data conversion policy and load conditions, and the relationship between the data conversion policy and analysis accuracy, and the data conversion unit executes a conversion process to reduce the amount of sensor data. Although the data that is effective in improving accuracy differs for each situation analysis method, compression is performed without taking this into account, resulting in a uniform loss of information and a decrease in analysis accuracy.

For example, Non-Patent Document 1 describes that different situation analysis methods use different data to improve estimation accuracy. Non-Patent Document 1 shows that in a situation analysis method using Deep Learning, which estimates distance using a monocular image as input, the vertical movement of an object has the greatest effect on estimation accuracy. In other words, when estimating distance using Deep Learning, creating a vertically long image that retains as much vertical information as possible and compressing it will enable more accurate estimation in the restored image.

The present invention aims to provide a data compression system that learns a data conversion model suitable for a situation analysis method and a compression model and restoration model suitable for that conversion model, enabling improved analysis accuracy compared to compression and restoration without data conversion, and to provide a data compression evaluation system that evaluates data compression methods suitable for situation analysis methods.

A representative example of the invention disclosed in the present application is as follows: That is, a data compression system constituted by a computer having a calculation device that executes calculation processing and a storage device accessible by the calculation device, comprising a data compression evaluation system, an analysis data input unit to which data to be analyzed is input, a data conversion unit that converts the data input to the analysis data input unit, a data compression unit that compresses the data converted by the data conversion unit, and a data restoration unit that restores the data compressed by the data compression unit, the data compression evaluation system generating learning data used for training a transformation model, a compression model, and a restoration model , an analysis method for analyzing a distance to an object indicated by the learning data, and a true value corresponding to the learning data , a transformation model that transforms the learning data , a compression model that compresses the data transformed by the transformation model, and a restoration model that restores the data compressed by the compression model, and converting the data input to the data input unit using the transformation model without decreasing analysis accuracy. the data conversion unit converts data input to the analysis data input unit using the conversion model selected by the evaluation unit, the data compression unit compresses the data converted using the compression model, restores the compressed data using the restoration model, and uses the restored data to estimate a distance to an object and calculate a first analytical value; an evaluation unit compares a first analytical value indicating the distance to the object analyzed by the model generation unit with the input true value, and selects each of a conversion model, a compression model, and a restoration model that minimizes a difference between the first analytical value and the true value; and a situation analysis unit estimates a distance to an object using data restored by the data restoration unit, wherein the data conversion unit converts data input to the analysis data input unit using the conversion model selected by the evaluation unit, the data compression unit compresses the data converted by the data conversion unit using the compression model selected by the evaluation unit, and the data restoration unit restores the data compressed by the data compression unit using the restoration model selected by the evaluation unit .

According to one aspect of the present invention, it is possible to improve the accuracy of analysis. Problems, configurations, and effects other than those described above will become clear from the explanation of the following examples.

1 is a block diagram showing an example of a data compression system for a vehicle recorder to which the present invention is applied; 4 is a block diagram showing a detailed configuration of a conversion/compression/restoration model generation unit. FIG. 13 is a flowchart of a process executed by a conversion/compression/restoration model generation unit. FIG. 13 is a diagram showing an example of a restoration evaluation table (expressed on an interval scale or the like). FIG. 13 is a diagram showing an example of a restoration evaluation table (expressed on a nominal scale or the like). FIG. 13 is a diagram showing an example of a restoration evaluation table (when expressed on an interval scale, a nominal scale, etc.). FIG. 13 is a diagram showing an example of a restoration evaluation table (when expressed on an interval scale, a nominal scale, etc.). 11 is a flowchart showing details of step 120. FIG. 13 is a diagram showing an example of a data conversion evaluation table (when expressed on an interval scale or the like). FIG. 13 is a diagram showing an example of a data conversion evaluation table (when expressed on an interval scale, a nominal scale, etc.). 11 is a flowchart showing details of step 150.

The following describes the examples with reference to the drawings.

Example 1
In this embodiment, a situation analysis method is given, and a data compression system that is suitable for application to vehicles is described, which performs data conversion, compression, and restoration model learning that are effective for improving the accuracy of the analysis, and performs a situation analysis method using the learned model.

Figure 1 is a block diagram showing an example of a data compression system for a vehicle recording device to which the present invention is applied.

The data compression system shown in FIG. 1 includes a learning/test data input unit 10 to which learning data and test data are input, a conversion/compression/restoration model generation unit 20 to learn a conversion, compression, and restoration model that matches the situation analysis method, a situation analysis value evaluation unit 30 to evaluate the model, an analysis data input unit 40 to which data to be analyzed is input, a data conversion unit 50 to convert the analysis data by the conversion model most evaluated by the situation analysis value evaluation unit 30, a data compression unit 60 to compress the output of the data conversion unit with the compression model generated by the conversion/compression/restoration model generation unit 20, a memory 70 to store the output of the compressed data, a data restoration unit 80 to restore the output of the memory with the restoration model generated by the conversion/compression/restoration model generation unit 20, and a situation analysis unit 90 to perform situation analysis from the output of the data restoration unit 80. The conversion/compression/restoration model generation unit 20 and the situation analysis value evaluation unit 30 constitute the learning part. The learning part, the learning/test data input unit 10, and the situation analysis unit 90 constitute a data compression evaluation system. The recording terminal device is made up of the data conversion unit 50, data compression unit 60, and memory 70. Note that the recording terminal device is configured as described above for analyzing data offline after the fact, but when analyzing data online in real time, the recording terminal device is made up of the data conversion unit 50, data compression unit 60, memory 70, data restoration unit 80, and situation analysis unit 90.

The learning/test data input unit 10 receives input of data used in the learning part. Specifically, the learning/test data input unit 10 receives input of data used for learning, test data for evaluating the learned model, a situation analysis method, and a true value corresponding to the value used in the situation analysis method. The learning data and test data are images captured by a monocular camera or a stereo camera, point cloud data acquired by LiDAR, control data, inertial data, and the like. The situation analysis method is distance estimation for calculating the distance to an object, and grasping surrounding objects in a bad environment such as at night. For example, when it is desired to know the distance between a colliding object and the vehicle in order to investigate a collision accident, the situation analysis method is distance estimation. In the case of distance estimation, the true value is the distance from the vehicle to the object.

The conversion/compression/restoration model generation unit 20 learns a conversion model that retains valid data so as not to reduce analysis accuracy in accordance with the situation analysis method input to the learning/test data input unit 10, and a compression model and restoration model that correspond to the conversion. The detailed configuration of the conversion/compression/restoration model generation unit 20 will be described with reference to FIG. 2.

The situation analysis value evaluation unit 30 determines the optimal conversion model, compression model, and restoration model by comparing the output from the conversion/compression/restoration model generation unit 20 with the true value data input to the learning/test data input unit 10.

The analysis data input unit 40, like the learning/test data input unit 10, receives inputs of sensor data (analysis data) such as images collected by the vehicle and a situation analysis method. The models used in the data conversion unit 50, data compression unit 60, and data restoration unit 80 are selected to suit the input situation analysis method.

The data conversion unit 50 converts the analysis data input to the analysis data input unit 40 using a conversion model that was generated by the conversion/compression/restoration model generation unit 20 and that was highly evaluated by the situation analysis value evaluation unit 30. Details of the method for selecting a conversion model will be described later with reference to FIG. 8. In addition, the conversion model used by the data conversion unit 50 can be dynamically changed depending on the situation analysis method. For example, when more accurate analysis becomes possible due to technological advances, a conversion model suitable for the situation analysis method can be learned by the conversion/compression/restoration model generation unit 20 and the situation analysis value evaluation unit 30, and the data conversion unit 50 can replace the conventional model with the model newly generated by learning.

The data compression unit 60 converts the output of the data conversion unit 12 with a compression model that is learned by the conversion/compression/restoration model generation unit 20 and selected by the situation analysis value evaluation unit 30. However, a compression model selected so that the size of the data output from the data compression unit 60 is smaller than the recording capacity of the memory 70 is used. In addition, the compression model used by the data compression unit 60 can be dynamically changed depending on the situation analysis method. For example, when more accurate analysis becomes possible due to technological advances, a compression model suitable for the situation analysis method can be learned by the conversion/compression/restoration model generation unit 20 and the situation analysis value evaluation unit 30, and the data compression unit 60 can replace the conventional model with a model newly generated by learning.

The memory 70 is a storage area that stores the data output from the data compression unit 60.

The data restoration unit 80 restores the pre-compression image from the compressed data stored in the memory 70 using the restoration model generated by the conversion/compression/restoration model generation unit 20 and highly evaluated by the situation analysis value evaluation unit 30. In addition, the restoration model used by the data restoration unit 80 can be dynamically changed depending on the situation analysis method. For example, when more accurate analysis becomes possible due to technological advances, a restoration model suitable for the situation analysis method can be learned by the conversion/compression/restoration model generation unit 20 and the situation analysis value evaluation unit 30, and the data restoration unit 80 can replace the conventional model with the model newly generated by learning.

The situation analysis unit 90 executes the situation analysis method input to the learning/test data input unit 10. For example, in the case of distance estimation, the distance to the object is calculated. In addition, among the situation analysis values calculated by the situation analysis unit 90, those that have a large difference from the true value are added to the learning data, and the conversion/compression/restoration model generation unit 20 and the situation analysis value evaluation unit 30 re-learn using new learning data, and the newly generated model is used in the data conversion unit 50, data compression unit 60, and data restoration unit 80, thereby reinforcing the model and improving the analysis accuracy.

The data compression system is configured with a computer having a processor (CPU), memory, auxiliary storage device, and a communication interface (not shown). The data compression system may also have an input/output interface for inputting and outputting data through the operation of an operator.

The processor is a computing device that executes programs stored in memory. The processor executes various programs to realize the functions of each section of the data compression system (e.g., the learning/test data input section 10, the conversion/compression/restoration model generation section 20, the situation analysis value evaluation section 30, the analysis data input section 40, the data conversion section 50, the data compression section 60, the data restoration section 80, the situation analysis section 90, etc.). Note that some of the processing performed by the processor by executing the programs may be executed by other computing devices (e.g., hardware such as ASICs and FPGAs).

Memory includes ROM, which is a non-volatile storage element, and RAM, which is a volatile storage element. ROM stores unchanging programs (e.g., BIOS), etc. RAM is a high-speed, volatile storage element such as DRAM (Dynamic Random Access Memory), and temporarily stores programs executed by the processor and data used when executing the programs.

The auxiliary storage device is, for example, a large-capacity, non-volatile storage device such as a magnetic storage device (HDD) or a flash memory (SSD). The auxiliary storage device also stores data used by the processor when it executes a program, and the program executed by the processor. In other words, the program is read from the auxiliary storage device, loaded into memory, and executed by the processor to realize each function of the data compression system.

A communication interface is a network interface device that controls communications with other devices according to a specific protocol.

The input/output interface is an interface to which input devices such as a keyboard and mouse are connected and which receives input from an operator, but it may also be an interface that supports voice input, etc. The output interface is an interface to which output devices such as a display device and printer are connected and which outputs the results of program execution in a format that can be viewed by the operator. Note that a terminal connected to the data compression system via a network may provide the input device and output device.

The programs executed by the processor of the data compression system are provided to the data compression system from removable media (CD-ROM, flash memory, etc.) or via a network, and are stored in a non-volatile auxiliary storage device, which is a non-transitory storage medium. For this reason, the data compression system should have an interface for reading data from removable media.

The data compression system is a computer system that is configured on one physical computer, or on multiple computers that are configured logically or physically, and may operate on a virtual computer built on multiple physical computer resources.

Figure 2 is a block diagram showing the detailed configuration of the conversion/compression/restoration model generation unit 20.

The conversion/compression/restoration model generation unit 20 includes a learning/test data conversion unit 21 that extracts data that contributes to improving the accuracy of the situation analysis method from the collected learning data and test data, a converted data situation analysis unit 22 that calculates a situation analysis value from the data output from the learning/test data conversion unit 21 using the situation analysis method, a compression model generation unit 23 that generates a compression model for compressing the data output from the learning/test data conversion unit 21, a restoration model generation unit 24 that generates a restoration model for restoring the data output from the compression model generation unit 23, and a restoration data situation analysis unit 25 that calculates a situation analysis value from the restoration data output from the restoration model generation unit 24 using the situation analysis method.

The learning/test data conversion unit 21 converts the learning data and test data input to the learning/test data input unit 10 so that the parts that contribute to improving the accuracy of the situation analysis method are maximized.

The converted data status analysis unit 22 calculates a status analysis value using a status analysis method from the learning data converted by the learning/test data conversion unit 21. The calculated status analysis value is compared with the true value in the status analysis value evaluation unit 30, and the conversion model capable of calculating a value closest to the true value is selected as the data conversion model.

The compression model generation unit 23 generates a compression model for compressing the data converted by the generative learning and test data conversion unit 21 so that the data is smaller than the recording capacity of the memory 70.

The restoration model generation unit 24 generates a restoration model for restoring the data compressed by the compression model generation unit 23 so that it is close to the data output from the training/test data conversion unit 21.

The restored data status analysis unit 25 compresses and restores the learning data converted by the learning/test data conversion unit 21 using the compression model generated by the compression model generation unit 23 and the restoration model generated by the restoration model generation unit 24, and then calculates an analysis value by a status analysis method. The compressed and restored learning data is used to calculate a status analysis value from the restored data using a status analysis method similar to that used by the converted data status analysis unit 22. The calculated status analysis value is compared with the true value in the status analysis value evaluation unit 30, and the converted model, compressed model, and restoration model that have been calculated to have a value closest to the true value are provided to the data conversion unit 50, data compression unit 60, and data restoration unit 80, respectively.

Figure 3 is a flowchart of the process executed by the conversion/compression/restoration model generation unit 20.

First, in step 100, the conversion/compression/restoration model generation unit 20 sets initial values, for example, S=50, N=1, and T=0, and proceeds to step 110. S is the number of trials to search for a data conversion model in step 120, N is the evaluation order that determines the data conversion model to be fixed in step 120, and T is the allowable time in step 190 for inputting test data, converting, compressing, and restoring, calculating the situation analysis value, and determining whether to re-learn. However, the values of S and N are natural numbers, and T is a positive value.

In step 110, the learning data, situation analysis method, true value, and situation analysis method for learning the conversion model, compression model, and restoration model are input, and the process proceeds to step 120. If multiple situation analysis methods are input, they are prioritized. For example, if, of the two situation analysis methods, the calculation accuracy of situation analysis method 1 is prioritized over the calculation accuracy of situation analysis method 2, then the priority of situation analysis method 1 is 1 and the priority of situation analysis method 2 is 2. Furthermore, if the priority is not taken into consideration, both situation analysis method 1 and situation analysis method 2 are prioritized 1.

In step 120, S conversion models are executed, the situation analysis value is calculated for each conversion model, and compared with the true value, the model with the Nth highest evaluation ranking is fixed, and the process proceeds to step 130. Details of the process in step 120 will be described with reference to FIG. 8.

In step 130, the same learning data as in step 110 is input again, and the process proceeds to step 140.

In step 140, the training data input in step 130 is converted using the data conversion model fixed in step 120, and the process proceeds to step 150.

In step 150, the data converted in step 140 is compressed and restored, and a compression model and a restoration model are learned so that the restored result is close to the data converted in step 140 (i.e., the data before compression). Details of the process in step 140 are described with reference to FIG. 11.

In step 160, test data, a situation analysis method, and true values for evaluating the learned conversion model, compression model, and restoration model are input, and the process proceeds to step 170.

In step 170, the test data is converted, compressed, and restored using the conversion model fixed in step 120 and the compression model and restoration model learned in step 150 to calculate a situation analysis value, the calculated situation analysis value is compared with the true value, the comparison result is stored in the restoration evaluation table, and the process proceeds to step 180.

In step 180, the comparison results stored in the restoration evaluation table are referenced, and the conversion model, compression model, and restoration model that minimize the difference between the situation analysis value and the true value are selected, and the process proceeds to step 190.

In step 190, if the execution time of the processes in steps 120 to 180 is less than the allowable time T, the process proceeds to step 200 to re-learn the compression model and the restoration model. On the other hand, if the execution time of the processes in steps 120 to 180 is greater than the allowable time T, the process ends the learning of the compression model and the restoration model, and provides the conversion model, compression model, and restoration model selected in step 180 for use in the data conversion unit 50, the data compression unit 60, and the data restoration unit 80.

In step 200, the evaluation rank N of the conversion model is incremented by 1 to update the value of N, and the process proceeds to step 120. For example, if N=1 in step 190, N=2 in step 200, and in steps 120 to 180, the compression model and restoration model are trained using the training data converted by the conversion model with the second smallest difference from the true value.

Figures 4, 5, 6, and 7 are diagrams showing examples of restored evaluation tables that store the results of learning. Figure 4 shows a case where the value calculated by the situation analysis method is an interval scale (e.g., distance estimation), Figure 5 shows a case where the value calculated by the situation analysis method is a nominal scale (e.g., estimating whether the object for which the emergency brake has been applied is a vehicle, a person, or a motorcycle, and setting 0 if incorrect and 1 if correct), Figure 6 shows a case where there are multiple situation analysis methods, the situation analysis values are both interval and nominal scales, and there are no values that are the same in the overall evaluation that takes both into consideration, and Figure 7 shows a case where the situation analysis method etc. are the same as in Figure 6, but the values are the same in the overall evaluation.

Figure 4 shows a restoration evaluation table that indicates the difference 210 from the true value using a situation analysis method (interval scale, etc.). The restoration evaluation table shown in Figure 4 indicates the difference (220-240) between the true value and the situation analysis value obtained by compressing and restoring data obtained by converting test data using each of three types of conversion models 1-3. In the case shown in Figure 4, conversion model 1 (220) with the smallest difference between the situation analysis value and the true value is selected in step 180.

Figure 5 shows a restoration evaluation table showing the number of correct answers 250 when compared with the true value using a situation analysis method (nominal scale, etc.). The restoration evaluation table shown in Figure 5 shows the number of correct answers (260, 270) when comparing the situation analysis values and true values obtained by compressing and restoring data obtained by converting test data using each of two types of conversion models 1 and 2. In the case shown in Figure 5, the model 2 (270), which has the highest number of correct answers, is selected in step 180.

Figure 6 is a diagram showing a restored evaluation table showing an overall evaluation taking into account both of a plurality of situation analysis methods (interval scale, nominal scale). The restored evaluation table shown in Figure 6 includes a difference 280 between the situation analysis value and the true value in situation analysis method 1 (interval scale, etc.), a number of correct answers 290 of the situation analysis value compared to the true value in situation analysis method 2 (nominal scale, etc.), an evaluation ranking 300 in situation analysis method 1 (interval scale, etc.) in which the difference 280 between the situation analysis value and the true value in situation analysis method 1 (interval scale, etc.) is ranked from 1 in descending order of the smallest, an evaluation ranking 310 in situation analysis method 2 (nominal scale, etc.) in which the number of correct answers 290 of the situation analysis value compared to the true value in situation analysis method 2 (nominal scale, etc.) is ranked from 1 in descending order, and an overall evaluation 320 calculated from the two evaluation rankings 300, 310.

Figure 6 shows the evaluation of the situation analysis values obtained by compressing and restoring the data obtained by converting the test data using each of the three types of conversion models 1 to 3 (330 to 350). The overall evaluation 320 is calculated using number (1) based on the priority of the situation analysis method in step 110 of Figure 3. Note that αk in number (1) is set so that the sum of all the values is 1, as in number (2). For example, if there are two situation analysis methods, the priority of situation analysis method 1 is 1 and the priority of situation analysis method 2 is 2, then k = 2 in number (1), and α1 = 0.7 and α2 = 0.3 are set so that the model with the highest analysis accuracy of situation analysis method 1 is selected, and the overall evaluation 320 calculated using number (1) is calculated. In this case, the conversion model 1 (330) with the smallest overall evaluation 320 is selected in step 180.

Figure 7 shows a restoration evaluation table in which the same situation analysis methods and evaluation result values as in Figure 6 are recorded. This shows an example of a restoration evaluation table in which both situation analysis method 1 and situation analysis method 2 have a priority of 1, that is, α1 = 0.5, α2 = 0.5. Using equation 1, the overall evaluation 400 is calculated from the evaluation rank 380 in situation analysis method 1 (interval scale, etc.) and the evaluation rank 390 in situation analysis method (nominal scale, etc.). As a result, conversion model 1 (410) and conversion model 2 (420) have the same value of 1.5. In this case, the conversion model, compression model, and restoration model whose situation analysis value and true value are closest to each other in the input order of the situation analysis method are selected. For example, in the case shown in FIG. 7, when the error 360 from the true value in situation analysis method 1 (interval scale, etc.) is compared between conversion model 1 (410) and conversion model 2 (420), conversion model 1 (410) is closer to the true value, so conversion model 1 (410) is selected in step 180 of FIG. 3.

Figure 8 is a flowchart showing the details of step 120 in Figure 3.

First, the conversion/compression/restoration model generation unit 20 converts the learning data input in step 110 in step 121 and proceeds to step 122. For example, when image data is input, the size, color information, etc. are the objects of data conversion. Conversion models that convert image size include reduction while maintaining the aspect ratio, and creating a vertically long image by changing the aspect ratio. Conversion models that convert color information include conversion to a black and white image, and conversion to a two-color image of red and blue from which only green information is deleted. For example, when it is desired to grasp an object that has undergone emergency braking at night, the entire image is dark and information such as the car's lights is important, so it is advisable to select an image conversion model that retains color information even when the image size is reduced, rather than a black and white image. The reason for performing such data conversion is that the data required for the situation analysis method is only a part of the whole, and much data useful for situation analysis is left, and the information is not lost due to compression. For example, Non-Patent Document 1 shows that in Deep Learning, which estimates distance using a monocular image as input, the vertical movement of the object has the greatest effect on the estimation accuracy. In other words, when estimating distance using Deep Learning, if you compress the image into a vertically long image that retains as much vertical information as possible, you can estimate the distance with high accuracy using the restored image.

In step 122, a situation analysis value is calculated from the data converted in step 121, and the process proceeds to step 123. Methods for calculating the situation analysis value include, for example, a model that uses Deep Learning to directly estimate distance from an image, and a model that uses camera parameters to estimate distance from the geometric relationship between the image and space.

In step 123, the situation analysis value calculated in step 122 is compared with the true value, the comparison result is recorded in the data conversion evaluation table, and the process proceeds to step 124. The data conversion evaluation table will be described in Figures 9 and 10.

In step 124, if the number of attempts S is other than 0, proceed to step 125; if the number of attempts S is 0, proceed to step 126.

In step 125, the number of trials S is updated by decrementing it by 1, and steps 121 to 123 are repeated until the number of trials S reaches 0, i.e., S trials of data conversion are performed, and the number of rows in the data conversion evaluation table becomes S.

In step 126, the model ranked Nth in the overall evaluation of the data conversion evaluation table is fixed as the data conversion model. In other words, if N is the initial value 1, the conversion model that enables calculation of the situation analysis value closest to the true value is fixed, and the process proceeds to step 130.

Figures 9 and 10 are diagrams showing examples of data conversion evaluation tables created or updated in step 123 of Figure 8. Figure 9 shows the case where the values calculated for one situation analysis method are interval scales, and Figure 10 shows the case where the values calculated for two situation analysis methods are interval scales and nominal scales. The configuration of the situation analysis methods is the same as the restoration evaluation tables in Figures 4 to 8.

Figure 9 is a diagram showing an example of a data conversion evaluation table in which an error 430 from the true value in a situation analysis method (interval scale, etc.) is recorded. The data conversion evaluation table shown in Figure 9 records the difference between the situation analysis value and the true value when each of conversion model 1 (440), ..., conversion model 49 (450), and conversion model 50 (460) is used. In the case shown in Figure 9, conversion model 50 (460) with the smallest difference between the situation analysis value and the true value is the conversion model fixed in step 126 of Figure 8.

Figure 10 is a diagram showing a data conversion evaluation table showing an overall evaluation taking into account both of a plurality of situation analysis methods (interval scale, nominal scale). The data conversion evaluation table shown in Figure 10 includes a difference 500 between the situation analysis value and the true value in situation analysis method 1 (interval scale, etc.), a number of correct answers 510 of the situation analysis value compared to the true value in situation analysis method 2 (nominal scale, etc.), an evaluation rank 520 in situation analysis method 1 (interval scale, etc.) in which the difference 500 between the situation analysis value and the true value in situation analysis method 1 (interval scale, etc.) is ranked from 1 in descending order of the smallest, an evaluation rank 530 in situation analysis method 2 (nominal scale, etc.) in which the number of correct answers 510 of the situation analysis value compared to the true value in situation analysis method (nominal scale, etc.) is ranked from 1 in descending order of the largest, and an overall evaluation 540 calculated from the two evaluation ranks 520 and 530. If the priority of situation analysis method 1 is 1 and the priority of situation analysis method 2 is 2, k = 2 in equation (1), and α1 = 0.7 and α2 = 0.3 are set so that a model with high analysis accuracy of situation analysis method 1 is selected, and the result of calculation using equation (1) is recorded in overall evaluation 540. Among converted model (470), ..., converted model 49 (480), and converted model 50 (490), converted model 49 (480) with the smallest overall evaluation 540 is the converted model fixed in step 126.

Figure 11 is a flowchart showing the details of step 150 in Figure 3.

First, in step 151, the transformation/compression/restoration model generation unit 20 compresses the data transformed in step 140, and proceeds to step 152. For example, some compression models use successive convolutional layers, etc., to reduce the image size.

In step 152, it is determined whether the data compressed in step 151 satisfies the memory capacity and communication capacity limitations. If the capacity limitations are met, the process proceeds to step 153. If the capacity limitations are not met, the process returns to step 151 and compresses the data converted using a different compression method so as to satisfy the capacity limitations.

In step 153, the data compressed in step 151 is restored, and the process proceeds to step 154. For example, if the image size was reduced using a convolutional layer in step 151, the restoration is performed using a restoration model that uses deconvolution to increase the image size.

In step 154, the data restored in step 153 is trained to reproduce the data before it was compressed in step 151. If the data is image data, the learning model may compare the values of each pixel between the restored image and the image before compression, and adjust the parameters of the compression model and restoration model so that the differences become smaller.

As described above, the data compression evaluation system of this embodiment includes a learning/test data input unit 10 to which learning data and test data are input, a conversion model that converts data including learning data and test data, a compression model that compresses data converted by the conversion model, and a restoration model that restores data compressed by the compression model, and a conversion/compression/restoration model generation unit 20 that uses the conversion model, compression model, and restoration model to convert, compress, and restore the data input to the learning/test data input unit 10 in accordance with a situation analysis method to calculate a situation analysis value, and The system includes a situation analysis value evaluation unit 30 that compares the situation analysis value calculated by the above with the learning data and test data input to the learning/test data input unit 10, and selects or modifies the conversion model, compression model, and restoration model that are optimal for the situation analysis method, a data restoration unit 80 that restores the compressed data using the restoration model selected or modified by the situation analysis value evaluation unit 30, and a situation analysis unit 90 that analyzes the situation using the data output from the data restoration unit 80. This makes it possible to learn in advance the conversion, compression, and restoration of data that can improve the accuracy of the situation analysis, and allows data useful for analysis to be stored in a memory with a limited capacity, and data can be transmitted even if there are restrictions on communication capacity. This improves the accuracy of analysis. For example, it is suitable for application to vehicle accident verification, determining the timing of maintenance and improving the efficiency of inspections of aging buildings, advance warning of dangerous behavior of infants and elderly people, checking the operation of manufacturing robots in factories, and understanding the growth status of plants and farmed fish in primary industries.

In addition, the situation analysis value evaluation unit 30 evaluates the conversion model generated by the learning/test data conversion unit 21 for each situation analysis method, and when learning the compression model and the restoration model, the conversion model is used in descending order of evaluation by the learning/test data conversion unit 21. Therefore, the most suitable conversion model can be used first for learning the compression model and the restoration model, and the appropriate model can be selected quickly and reliably.

In addition, the learning/test data conversion unit 21 learns from the data input to the learning/test data input unit 10 in a way that retains as much data as possible that contributes to improving the accuracy of the situation analysis method, and generates a conversion model, so that a conversion model that includes a lot of different important data for each situation analysis method can be selected.

In addition, data acquired from a moving object such as a vehicle is input to the learning/test data input unit 10, so accidents can be verified by analyzing the sensor data acquired by the vehicle.

The situation analysis unit 90 also inputs data calculated by the situation analysis method that has a large difference from the true value to the learning and test data input unit 10, so that the model can be reinforced by re-learning using the learning data that has a large difference from the true value, thereby improving the analysis accuracy.

The data compression system of this embodiment includes any one of the data compression evaluation systems described above, an analysis data input unit 40 to which data to be analyzed is input, a data conversion unit 50 that converts the data input to the analysis data input unit 40 using a conversion model selected or modified by the situation analysis value evaluation unit 30, a data compression unit 60 that compresses the data converted by the data conversion unit 50 using a compression model selected or modified by the situation analysis value evaluation unit 30, and a memory 70 that stores the compressed data. The models used by the data conversion unit 50, the data compression unit 60, and the data restoration unit 80 can be dynamically changed, so that model improvements and changes in analysis methods can be accommodated by replacing the models. For example, when a more accurate analysis method is proposed, the models can be updated in line with technological advances by re-learning the conversion model, compression model, and restoration model suitable for that analysis method.

The present invention is not limited to the above-described embodiments, but includes various modified examples and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to having all of the configurations described. Furthermore, part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Furthermore, the configuration of another embodiment may be added to the configuration of one embodiment. Furthermore, part of the configuration of each embodiment may be added, deleted, or replaced with other configurations.

Furthermore, each of the configurations, functions, processing units, processing means, etc. described above may be realized in part or in whole in hardware, for example by designing them as integrated circuits, or may be realized in software by a processor interpreting and executing a program that realizes each function.

Information such as programs, tables, and files that realize each function can be stored in a storage device such as a memory, hard disk, or SSD (Solid State Drive), or in a recording medium such as an IC card, SD card, or DVD.

In addition, the control lines and information lines shown are those considered necessary for explanation, and do not necessarily represent all control lines and information lines necessary for implementation. In reality, it is safe to assume that almost all components are interconnected.

10...Learning/test data input section, 20...Conversion/compression/restoration model generation section, 30...Situation analysis value evaluation section, 40...Analysis data input section, 50...Data conversion section, 60...Data compression section, 70...Memory, 80...Data restoration section, 90...Situation analysis section

Claims (6)

A data compression system including a computer having an arithmetic unit that executes arithmetic processing and a storage device that is accessible by the arithmetic unit,
A data compression evaluation system;
an analysis data input section to which data to be analyzed is input;
a data conversion unit that converts data input to the analysis data input unit;
a data compression unit that compresses the data converted by the data conversion unit;
a data recovery unit that recovers the data compressed by the data compression unit,
The data compression evaluation system includes:
a data input unit for inputting learning data used for learning a transformation model, a compression model, and a restoration model, an analysis method for analyzing a distance to an object indicated by the learning data, and a true value corresponding to the learning data;
a model generation unit that generates a conversion model for converting the learning data, a compression model for compressing data converted by the conversion model, and a restoration model for restoring data compressed by the compression model, converts data input to the data input unit using the conversion model so as to leave valid data so as not to reduce analysis accuracy, compresses the converted data using the compression model, restores the compressed data using the restoration model, and estimates a distance to an object using the restored data to calculate a first analytical value;
an evaluation unit that compares a first analytical value indicating a distance to an object analyzed by the model generation unit with the input true value, and selects each of a conversion model, a compression model, and a restoration model that minimizes a difference between the first analytical value and the true value;
a situation analysis unit that estimates a distance to an object by using the data restored by the data restoration unit,
the data conversion unit converts the data input to the analysis data input unit using the conversion model selected by the evaluation unit;
the data compression unit compresses the data converted by the data conversion unit using the compression model selected by the evaluation unit;
A data compression system, wherein the data restoration unit restores the data compressed by the data compression unit using a restoration model selected by the evaluation unit. 2. The data compression system of claim 1,
The model generation unit
a training data conversion unit that generates the conversion model and converts data using the generated conversion model;
a compression model generation unit that generates a compression model for compressing the data converted by the learning data conversion unit, and compresses the converted data using the generated compression model;
a restoration model generation unit that generates a restoration model for restoring data compressed by the compression model generation unit and restores the compressed data using the generated restoration model;
a restored data analysis unit that estimates a distance to an object using the data restored by the restored model generation unit to calculate a first situation analysis value;
a converted data analysis unit that uses the data converted by the data conversion unit to estimate a distance to an object and calculates a second analytical value. 3. A data compression system according to claim 2, comprising:
A data compression system characterized in that the data conversion unit generates the conversion model by learning from the data input to the data input unit in a manner that retains as much data as possible that contributes to improving accuracy in the analysis method. 2. The data compression system of claim 1,
A data compression system, wherein data acquired from a vehicle is input to the data input unit. 2. The data compression system of claim 1,
A data compression system, wherein the situation analysis unit inputs learning data having a large difference between the first analytical value and the true value to the data input unit. A data compression method executed by a data compression evaluation system configured by a computer having a calculation device that executes calculation processing and a storage device accessible by the calculation device, comprising:
a data compression evaluation procedure;
an analysis data input step in which data to be analyzed is input;
a data conversion step in which the arithmetic device converts the data input in the analysis data input step;
a data compression step in which the arithmetic device compresses the data converted in the data conversion step;
a data restoration step for restoring data compressed by the data compression step,
In the data compression evaluation step,
learning data used for learning a transformation model, a compression model, and a restoration model, an analysis method for analyzing a distance to an object indicated by the learning data, and a data input procedure for inputting a true value corresponding to the learning data;
a model generation step in which the arithmetic device generates a conversion model for converting the learning data, a compression model for compressing data converted by the conversion model, and a restoration model for restoring data compressed by the compression model, converts data input in the data input step using the conversion model so as to leave valid data so as not to reduce analysis accuracy, compresses the converted data using the compression model, restores the compressed data using the restoration model, and estimates a distance to an object using the restored data to calculate a first analytical value;
an evaluation step in which the arithmetic device compares a first analytical value indicating a distance to an object analyzed in the model generation step with the input true value, and selects each of a conversion model, a compression model, and a restoration model that minimizes a difference between the first analytical value and the true value;
a situation analysis step in which the arithmetic device estimates a distance to an object by using the data restored in the data restoration step;
In the data conversion step, the arithmetic device converts the data input in the analysis data input step using the conversion model selected in the evaluation step;
In the data compression step, the arithmetic device compresses the data converted in the data conversion step using the compression model selected in the evaluation step;
A data compression method, characterized in that in the data restoration step, the arithmetic device restores the data compressed in the data compression step using the restoration model selected in the evaluation step.

Images