JP7199974B2 - Parameter selection device, parameter selection method, and parameter selection program - Google Patents

Description

The present invention relates to an object detection device and an object detection method.

There is a growing need for object detection technology that detects objects based on information acquired by measuring devices (hereinafter sometimes referred to as sensors), and object detection results are used in a variety of applications according to user requirements. For example, detection of an intruder into a predetermined area for monitoring purposes, extraction of trajectories of objects for the purpose of analyzing the flow of people and traffic, counting of the number of objects, etc. can be mentioned. Monitoring cameras, distance sensors, laser radars, and the like are widely used as sensors. If the sensors can be evenly installed at appropriate locations, it will be possible to achieve highly accurate object measurement and subsequent applications required by the user.

However, in reality, the number of sensors to be used and the performance of object detection are limited due to factors such as cost and the environment of the installation site, and there are many cases where the expected measurement accuracy is not achieved. In such situations, it is necessary for engineers to visit the site and spend time adjusting parameters such as the sensor installation position, orientation, angle of view, and thresholds used for object detection in order to maintain accuracy. rice field. In view of this situation, in recent years, technology for automatically optimizing sensor parameters has been developed, but it is not realistic to optimize the parameters of a large number of sensors at once, considering the computational cost. . Therefore, there are high expectations for technology for selecting parameters to be adjusted. For example, in Japanese Unexamined Patent Application Publication No. 2002-100000, parameters relating to optimal placement conditions for surveillance cameras are automatically selected and optimized.

JP 2018-128961 A

However, in Patent Literature 1, the parameters can be efficiently optimized by selecting the arrangement conditions of the surveillance cameras so as to realize surveillance without blind spots with the minimum number of cameras, but the threshold used for object detection is etc. cannot be targeted. In addition, we select the parameters for optimal surveillance camera placement conditions according to user requirements such as surveillance. It is not possible to preselect parameters suitable for achieving the requirements.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and it is an object of the present invention to appropriately select parameters to be adjusted that are suitable for achieving user requirements for each measurement purpose when detecting an object from a measurement range. and

In order to solve such a problem, in one example of the present invention, the parameter selection device uses the algorithm for the parameter used when the object detection device detects an object within the measurement range of the measurement device. a parameter influence calculation unit that calculates a first influence on the detection accuracy of detecting the object by using the object detection device; and a parameter selection unit for selecting the parameter to be adjusted based on the first influence and the second influence. characterized by

ADVANTAGE OF THE INVENTION According to this invention, when detecting an object from a measurement range, the parameter of the adjustment object suitable for achievement of user requirements for every measurement purpose can be selected appropriately.

FIG. 2 is a functional block diagram showing the configuration of the object detection device according to the first embodiment; FIG. FIG. 4 is a diagram for explaining the processing of the processing block decomposing unit; FIG. 4 is a functional block diagram showing the configuration of a parameter contribution calculation unit; The figure which shows an example of the variation amount of the measurement accuracy calculated by the measurement accuracy variation amount calculation part. The figure which shows an example of a parameter contribution rate table. A diagram for explaining user requirements. FIG. 4 is a diagram showing an example of an analysis item table and a combination table; The figure which shows an example of an analysis item contribution rate table. 4 is a flowchart showing object detection processing according to the first embodiment; FIG. 10 is a functional block diagram showing configurations of an object detection device and a parameter selection device according to a second embodiment; FIG. 11 is a functional block diagram showing the configuration of the parameter selection device of Example 3; FIG. 11 is a functional block diagram showing the configurations of an object detection device and a parameter selection device according to a fourth embodiment; FIG. 2 is a hardware diagram showing the configuration of a computer that realizes the object detection device and parameter selection device of Examples 1 to 4;

An embodiment of the present invention will be described in detail below with reference to the drawings. In each drawing for explaining the embodiments below, the same reference numbers indicate the same or similar configurations or processes, and the descriptions thereof will be omitted. Moreover, each embodiment and each modification can be combined in whole or in part within the scope of the technical idea of the present invention and within the scope of matching.

<Structure of Object Detection Apparatus of Embodiment 1>
FIG. 1 is a functional block diagram showing the configuration of the object detection device of Example 1. As shown in FIG. The object detection device 1 of the first embodiment is a device that automatically selects parameters to be used for detecting an object detected by a sensor using sensor information 200 and user requirements 201 such as the purpose of measurement by a measurement device such as a sensor. be. In this embodiment, a case where a stereo camera is used as a sensor will be described, but the present invention is not limited to a stereo camera. application is also possible.

The object detection device 1 includes a processing block decomposition unit 2, a parameter contribution ratio calculation unit 3, a user analysis item extraction unit 4, an analysis item contribution ratio calculation unit 5, an optimization processing block determination unit 6, and a parameter selection unit. 7 , a parameter adjustment unit 8 , an object detection execution unit 9 , and an analysis execution unit 10 .

A processing block decomposition unit 2, a parameter contribution ratio calculation unit 3, a user analysis item extraction unit 4, an analysis item contribution ratio calculation unit 5, an optimization processing block determination unit 6, a parameter selection unit 7, a parameter Each functional unit of the adjustment unit 8 and the object detection execution unit 9 is realized in a sensor having an arithmetic unit, a main storage device, and an external storage device, or in a computer prepared separately from the sensor.

The processing block decomposer 2 decomposes the sensor-specific measurement algorithm included in the sensor information 200 into a plurality of processing blocks. The parameter contribution calculation unit 3 determines how much the parameters used in each of the plurality of processing blocks decomposed by the processing block decomposing unit 2 affect the accuracy of the output result of the sensor-specific measurement algorithm. Calculate the contribution shown.

The user analysis item extraction unit 4 extracts sensor analysis items required by the user from the user requirements 201 . The analysis item contribution ratio calculator 5 calculates the contribution ratio of the analysis item to each processing block decomposed by the processing block decomposing unit 2 .

The optimization processing block determination unit 6 extracts information on the analysis items extracted from the user requirements 201 by the user analysis item extraction unit 4, and the analysis item contribution ratio calculated by the analysis item contribution ratio calculation unit 5 for each processing block. Based on the information, a processing block for performing optimization processing is determined.

The parameter selection unit 7 selects a parameter having a high contribution rate in the processing block determined by the optimization processing block determination unit 6 to perform the optimization processing as a parameter to be optimized. The parameter adjustment unit 8 adjusts the parameters selected by the parameter selection unit 7 so as to improve the accuracy of object detection using a sensor-specific measurement algorithm.

The object detection execution unit 9 executes object detection using designated parameters. The analysis execution unit 10 uses the result of object detection by the object detection execution unit 9 to execute an application according to user requirements 201 .

Details of the functions of the processing block decomposition unit 2, the parameter contribution ratio calculation unit 3, the user analysis item extraction unit 4, the analysis item contribution ratio calculation unit 5, the optimization processing block determination unit 6, and the parameter selection unit 7 will be described below. do.

<Processing of processing block decomposing unit>
FIG. 2 is a diagram for explaining the processing of the processing block decomposing unit. The object detection algorithm 12 is an example of a measurement algorithm that outputs an output result 14 obtained by subjecting an input from the stereo camera 11 to object detection processing, for example. Image acquisition 12a, distortion correction 12b, parallax calculation 12c, 3D point cloud calculation 12d, and object detection 12e are examples of the processing blocks into which the object detection algorithm 12 is decomposed.

The processing block decomposition unit 2 acquires the object detection algorithm 12 of the stereo camera 11 used for object detection from the sensor information 200, and decomposes it into a plurality of processing blocks. The sensor information 200 includes information such as program source for object detection that allows the user to grasp the details of the algorithm, information on the SDK unique to the sensor used for object detection, parameters related to object detection, and the like.

FIG. 2 shows an example of decomposing the object detection algorithm 12 of the stereo camera 11 into a plurality of processing blocks. As a method of dividing the object detection algorithm 12 into a plurality of processing blocks, for example, for each function or class in the program source, for each SDK to be used, etc. , a division method in which each flowchart obtained by subdividing an algorithm using an analysis tool or the like is used as a processing block.

Each processing block will be described. The image acquisition 12a is a process of acquiring camera images by controlling the stereo camera 11, and parameters 13a such as the frame rate indicating the frequency of image acquisition and the resolution of images to be acquired are used. The distortion correction 12b is a process of removing camera-specific lens distortion from the image acquired by the image acquisition 12a. There are common methods such as removing distortion.

The parallax calculation 12c compares the two camera images from which the lens distortion has been removed by the distortion correction 12b, defines a region of a certain size that minimizes the degree of difference as a small region, and calculates the parallax from the search width. Examples of methods for calculating the dissimilarity of small regions include methods such as SAD (Sum of Absolute Difference) and SSD (Sum of Squared Difference). The parameters 13c used in the parallax calculation 12c include the size of the small region and the maximum value of the search width (maximum search width).

In the three-dimensional point cloud calculation 12d, using parameters 13d such as the installation position and installation orientation of the stereo camera 11, based on the following equations (1) and (2), camera coordinates Xc and world coordinates Xw are calculated from the image coordinates. , to derive the three-dimensional point group coordinates of the world coordinates Xw. In the following equation (2), the rotation matrix R indicates the installation attitude of the camera, and can be calculated from the pan, tilt, and roll angles of the stereo camera 11 . Also, in the following equation (2), the translation vector Xw indicates the installation position such as the camera height of the stereo camera 11 .

ただし、
(u,v)：画像座標、(u₀,v₀)：画像中心、b：基線長、f：焦点距離、d：視差、pit：ピッチ

however,
(u, v): image coordinates, (u ₀ , v ₀ ): image center, b: baseline length, f: focal length, d: parallax, pit: pitch

ただし、R：回転行列、t：並進ベクトル

where R: rotation matrix, t: translation vector

The object detection 12e judges and detects an object when the shape, volume, and the like obtained by analyzing the three-dimensional point cloud calculated by the three-dimensional point cloud calculation 12d satisfy certain reference values. The parameters 13e used in the object detection 12e include a detection threshold value, which is a constant reference value, and the number of times correction processing is performed to remove noise from the three-dimensional point group.

Although the object detection algorithm 12 of the stereo camera 11 is divided into five processing blocks in this embodiment, the number of divisions is not limited. Further, the processing contents of each processing block and the parameters to be used are not limited to those shown in FIG.

<Configuration of Parameter Contribution Rate Calculation Unit>
FIG. 3 is a functional block diagram showing the configuration of the parameter contribution calculation unit. The parameter contribution ratio calculation unit 3 has a parameter setting range determination unit 20 , a measurement accuracy variation amount calculation unit 21 , and a parameter contribution ratio table creation unit 22 .

The parameter setting range determination unit 20 receives the parameters 13a to 13e used in each processing block and determines the setting range of parameter values for calculating the contribution rate of each parameter 13a to 13e. The measurement accuracy variation calculation unit 21 calculates the measurement accuracy based on the object detection output result 14 output from the object detection execution unit 9 when the parameter is changed within the setting range acquired by the parameter setting range determination unit 20. Calculate the amount of variation. The parameter contribution rate table creation unit 22 creates a table showing the parameter contribution rate from the measurement accuracy variation amount calculated by the measurement accuracy variation amount calculation unit 21 .

Each function will be described below. The parameter setting range determination unit 20 receives as input the initial values of the parameters used in each processing block decomposed by the processing block decomposition unit 2, and the parameter setting range used by the measurement accuracy variation calculation unit 21. is determined for each parameter. The initial values input to the parameter setting range determination unit 20 include, for example, values set by the user, prescribed values on the system, and values automatically estimated by a predetermined algorithm.

In this embodiment, default values of the parameters 13a, 13c, and 13e used in the image acquisition 12a, the parallax calculation 12c, and the object detection 12e are defined by the user on the system. Estimated values by a predetermined algorithm are used as initial values of parameters 13b and 13d used in distortion correction 12b and three-dimensional point cloud calculation 12d. For example, the parameter 13b can be estimated using the Zhang calibration method, and the parameter 13d can be estimated using a method using plane information detected by the RANSAC algorithm.

As a method of determining the setting range from the initial values input to the parameter setting range determining section 20, there is a method of setting values within a predetermined range before and after the initial value as the minimum and maximum values. Here, the method of determining the minimum and maximum values may be switched depending on whether the user knows the details of the object detection algorithm 12 or not. For example, if the user knows the details of the object detection algorithm 12, the minimum and maximum values can be determined based on past experience. Parameter values may be investigated and determined. Alternatively, both methods may be combined. A method for determining the minimum and maximum values is not particularly limited.

Also, in this embodiment, discrete values are used as parameters included in the setting range. For example, in the case of a continuous parameter, the interval between the minimum value and the maximum value of the setting range is divided into a predetermined number to generate parameters with discrete values. In the case of discrete parameters, all possible values may be included in the parameter setting range.

The measurement accuracy variation calculation unit 21 changes the parameter value within the setting range of each parameter input by the parameter setting range determination unit 20 to calculate the variation in measurement accuracy by the object detection execution unit 9 . In the present embodiment, as the measurement accuracy, the object detection accuracy when inputting a true-valued test image or an evaluation index unique to the processing block is used.

As a method for calculating the accuracy of object detection, we used a test video in which the detection target was captured and which retained true value information so that the target area could be known in advance for all frames. The ratio of the number of frames in which the is detected to the total number of frames can be calculated as the object detection accuracy.

In the case of the processing block of the parallax calculation 12c, the evaluation index unique to the processing block is, for example, using a random pattern with which the parallax is relatively easily calculated as a test image, and obtaining the parallax by calculating the parallax for the number of pixels in the entire image. It may also be a value that represents the ratio of the total number of invalid parallax pixels that cannot be used as a percentage.

In addition, the original evaluation index of the processing block of the distortion correction 12b is, for example, using a test video with many straight lines such as floor tiles, and applying a straight line detection algorithm to the image after distortion correction for the number of straight lines in the test video. The ratio of the number of straight lines applied and detected may be expressed as a percentage.

In addition, for example, when the setting range of the parameter p is 10 integer values of 0 ≤ p ≤ 9, the amount of variation in the measurement accuracy is determined by fixing all parameter values other than p to the initial value, p = 0, It can be obtained by calculating the object detection accuracy in the test video when p=1, . FIG. 4 is a diagram showing an example of variation in measurement accuracy calculated by the measurement accuracy variation calculation unit. FIG. 4 shows an example in which the measurement accuracy takes the maximum value when p=2.

Note that when calculating the measurement accuracy by the measurement accuracy variation calculation unit 21, depending on whether the user can grasp the detailed algorithm of each processing block, the measurement accuracy may be the object detection accuracy or the processing block's own evaluation. You can switch which of the indicators to use. For example, for each processing block whose detailed algorithm can be grasped, test data suitable for evaluating the unique index of the processing block is prepared. For the processing blocks that cannot grasp the algorithm, test data corresponding to the object detection accuracy may be used in common among the processing blocks that cannot grasp the algorithm to calculate the measurement accuracy.

The parameter contribution rate table creation unit 22 uses the variation in measurement accuracy calculated by the measurement accuracy variation calculation unit 21 to create a parameter contribution rate table T1 for each processing block. The contribution rate of a parameter to a processing block is an index that indicates the importance of each parameter in the processing block.

As a method of calculating the parameter contribution rate from the variation amount of the measurement accuracy, for example, it is determined that the larger the maximum value of the measurement accuracy when the parameter is varied, the higher the contribution rate of the parameter, and the maximum value of each parameter is set to 0. There is a method of using a value normalized in the range from to 100 as the contribution rate. Alternatively, for example, a method of using a value normalized by using any one of the minimum value, the average value, and the variance value of the measurement accuracy, or a unique index combining any one of them, may be used as the contribution rate. The parameter contribution rate table creation unit 22 stores each parameter contribution rate calculated for each processing block in the parameter contribution rate table T1.

FIG. 5 is a diagram showing an example of a parameter contribution rate table. A parameter contribution rate table T1 shown in FIG. 5 is stored in a predetermined storage unit, and includes contribution rate tables T1a to T1e for each processing block. The parameter contribution rate table T1 includes a contribution rate table T1a storing the contribution rate of each parameter 13a in the image acquisition 12a, a contribution rate table T1b storing the contribution rate of each parameter 13b in the distortion correction 12b, and each parameter 13c in the parallax calculation 12c. Contribution rate table T1c storing the contribution rate of , Contribution rate table T1d storing the contribution rate of each parameter 13d in the three-dimensional point cloud calculation 12d, and Contribution rate table T1e storing the contribution rate of each parameter 13e in the object detection 12e including.

In this embodiment, the parameter contribution rate is calculated by the configuration of the parameter contribution rate calculation unit 3 shown in FIG. is not particularly limited.

<User requirements>
User requirements 201 will be described with reference to FIG. FIG. 6 is a diagram for explaining user requirements. User requirements 201 include a list of applications for which the user implements object detection techniques to achieve the primary goal.

FIG. 6 shows, in the measurement range 30A of the stereo camera 11 shown by the two-dimensional map 30, each measurement area 32a, 32b set around each of the facilities 31a, 31b related to the user's main purpose, A list 33 of applications that the user wants to execute is illustrated. User requirements 201 include two-dimensional map 30 and inventory 33 .

For example, if the user's main purpose is to monitor the entire measurement range 30A, the measurement area is the entire two-dimensional map 30, and it is necessary to execute an application such as moving object detection for detecting a moving person in the image. . Also, if the user's main purpose is to analyze the flow of people in the entire measurement range 30A, the measurement area is the entire two-dimensional map 30, and it is necessary to execute an application such as human trajectory estimation.

If the main purpose of the user is to grasp the utilization rate of the equipment 31a, a measurement area 32a is set around the equipment 31a, and an application such as counting the number of people in the measurement area 32a or measuring the residence time can be executed. , the number of users and the usage time of the equipment 31a can be analyzed. In addition, when the main purpose of the user is to grasp the user class of the facility 32b, the measurement area 32b is set around the facility 31b, and the number of people in the area 32b is counted. By executing applications such as a person's gender determination and a carry-back detection that detects the possession of a person's carry-back, it is possible to analyze the ratio of men and women to the total number of people and the ratio of carry-back holders.

Note that the user requirement 201 in FIG. 6 is merely an example, and is not limited to this. As a method for creating the user requirements 201, as shown in FIG. 6, the user uses a GUI or the like to set the measurement area on the two-dimensional map 30, and the entire measurement range or each set measurement area There is a method of registering an application to be executed in the list 33 and creating it. The two-dimensional map 30 and list 33 are stored in a predetermined storage unit. However, the information and method are not particularly limited as long as the range to be measured and the application to be executed can be grasped by the user. Further, the main purpose is not particularly limited as long as it is abstraction level information that can be grasped by the application corresponding to the measurement area, or information that can be converted into information that can be grasped by the application corresponding to the measurement area.

<Analysis item table and combination table>
Processing of the user analysis item extraction unit 4 will be described with reference to FIG. FIG. 7 is a diagram showing an example of an analysis item table and a combination table. FIG. 7A shows an example of an analysis item list table 40 showing a list of analysis items extracted from each processing block decomposed by the processing block decomposing unit 2. FIG. FIG. 7B shows an example of a combination table 41 of applications and analysis items. According to FIG. 7A, the analysis item list table 40 includes "object detection", "tracking", "position measurement", "behavior recognition", "image recognition", "shape recognition", and "size measurement". analysis items. The analysis item list table 40 and combination table 41 are stored in a predetermined storage unit.

The user analysis item extraction unit 4 extracts analysis items necessary for realizing an application desired by the user from the information of the user requirements 201 and the combination table 41 created in advance. The analysis item is an analysis technique (application) for acquiring necessary information in addition to object detection when executing the application.

For example, in the combination table 41, "application" and "moving body detection" are functions that can be realized only with the result of object detection, so "analysis item" is "none". Also, for example, the "application" and "residence time measurement" use the result of object detection, so in addition to object detection, the object is "tracked" and " "Position measurement" analysis items are required. In this way, from the user's main purpose in the user requirements 201, it is possible to grasp the application that uses the result of object detection and the analysis items corresponding to the application.

For example, as shown in the list 33 of FIG. 6, in the user requirement 201, when the "main purpose" is "understanding the utilization rate of the measurement area 31a", the "applications" that need to be executed are "counting people in the area" and It becomes "residence time measurement". For this reason, the user analysis item extraction unit 4 refers to the combination table 41 to refer to the "analysis item", "position measurement" and " Extract "tracking".

As for the “analysis item”, items other than the analysis items shown in the analysis item list table 40 of FIG. 7A may also be used. Also, one analysis item may be divided into a plurality of items based on the positional relationship between the measurement range and the sensor, the accuracy, and the like. For example, in object detection, when the distance from a sensor such as a camera to an object increases, the amount of information about the object decreases, or the probability that an obstacle exists between them increases, resulting in a decrease in detection accuracy. Therefore, there is a method of dividing the "analysis item" and "object detection" into "near detection" and "distant detection". In addition, for safety and security reasons, there are cases where location information with information granularity less than or greater than a predetermined level is required. There is a method of dividing into "measurement".

<Analysis item contribution rate table>
FIG. 8 is a diagram showing an example of an analysis item contribution rate table. FIG. 8 is an example of calculation of the contribution rate of the analysis item for each processing block of the distortion correction 12b and the three-dimensional point group calculation 12d by the analysis item contribution rate calculator 5. In FIG. As a method of calculating the contribution rate of the analysis item for each processing block, the method of the parameter contribution rate calculation unit 3 is followed. There is a method of calculating the contribution rate using the amount. Note that FIG. 8 assumes that the "analysis item" and "object detection" shown in FIG. 7A are subdivided into "near detection" and "distant detection".

A method of calculating the contribution rate for each processing block of "analysis item" and "tracking" will be described below.

First, a true-priced test video is generated to evaluate the "tracking" technique. Next, the parameters are changed only in the target processing block for obtaining the amount of variation in the accuracy of the "analysis item" and "tracking", and object detection and tracking using the object detection results are performed. As a method of changing parameters, there is a method of sequentially using all combination patterns of parameters used in the same processing block. Finally, the amount of change in tracking accuracy due to changes in parameters is calculated. The tracking accuracy is, for example, how much the same person can be tracked. By comparing the amount of variation in tracking accuracy of all processing blocks and the amount of variation in tracking accuracy of each processing block, the contribution rate of the tracking accuracy of each processing block to the variation in tracking accuracy of all processing blocks can be obtained.

It should be noted that when creating a combination pattern of parameters used in the processing block when determining the amount of variation in the accuracy of the analysis item, instead of using all the parameters, the parameter setting range determination unit 20 (see FIG. 3) The number of combinations may be reduced by a method of utilizing the information of

In addition to the method described in this embodiment, any method can be used as long as it can calculate the degree of influence of each processing block on the accuracy of the analysis item.

As described above, as shown in FIG. 8, the method of calculating the contribution rate of the analysis item for each processing block of the distortion correction 12b and the three-dimensional point cloud calculation 12d has been described. The contribution ratio of analysis items to each processing block is also calculated in the same manner.

As shown in FIG. 8, the user analysis item extraction unit 4 registers the contribution rate of each analysis item for each processing block calculated as described above in the analysis item contribution rate table T2. The analysis item contribution rate table T2 is stored in a predetermined storage unit, and includes contribution rate tables T2a to T2e for each processing block. The analysis item contribution rate table T2 includes a contribution rate table T2a (not shown) storing the contribution rate of each analysis item in the image acquisition 12a, a contribution rate table T2b storing the contribution rate of each analysis item in the distortion correction 12b, and a parallax A contribution rate table T2c (not shown) storing the contribution rate of each analysis item in the calculation 12c, a contribution rate table T2d storing the contribution rate of each analysis item in the three-dimensional point cloud calculation 12d, and each analysis in the object detection 12e It includes a contribution rate table T2e (not shown) that stores contribution rates of items.

Returning to the description of FIG. The optimization processing block determination unit 6 uses the output information from the user analysis item extraction unit 4 and the analysis item contribution ratio calculation unit 5 to optimize the processing block with the largest contribution ratio of the analysis items required in the user requirements 201. is determined as a conversion processing block. For example, if the user requirement 201 requires only one analysis item, there is a method of selecting the processing block with the highest contribution rate to the corresponding analysis item.

Also, when a plurality of analysis items are required, there is a method of selecting a processing block that maximizes the value obtained by multiplying each contribution rate. For example, in the case of analysis items that require "neighborhood detection" and "tracking" in FIG. In the point cloud calculation 12d, 60×30=1800. Therefore, the processing block for the three-dimensional point group calculation 12d is selected over the distortion correction 12b. Note that the number of processing blocks determined by the optimization processing block determining unit 6 is not limited to one. You may judge as a target.

The parameter selection unit 7 selects a parameter with the highest parameter contribution rate calculated by the parameter contribution rate calculation unit 3 as a parameter to be optimized in the processing block targeted for optimization. Note that the parameter to be selected is not limited to one. For example, a method may be adopted in which a threshold value for the parameter contribution rate is set in advance, and all parameters with higher parameter contribution rates than the threshold value are targeted for optimization. .

Further, when a plurality of processing blocks are output by the optimization processing block determination unit 6, the contribution rate of the analysis item of the processing block and the parameter contribution rate within the processing block are multiplied, and a value higher than a preset threshold value is obtained. A method of optimizing the indicated parameters may also be used.

For example, the processing blocks output by the optimization processing block determination unit 6 are the distortion correction 12b and the three-dimensional point cloud calculation 12d whose tables are shown in FIGS. ” is “tracking”. Multiplying each parameter contribution rate of the distortion correction 12b by the contribution rate "20" of the analysis item to the "tracking" of the distortion correction 12b gives 20×60=1200 for the "focal length" and 20×30 for the "distortion coefficient". = 600, and the "image center" is 20 x 10 = 200. When the 3D point cloud calculation 12d is calculated in the same way, the “camera height” is 30×60=1800, the “tilt angle” is 30×30=900, the “roll angle” is 30×10=300, and the “scale” is is 30×20=600. Therefore, when the preset threshold value is 1000, the "focal length" of the distortion correction 12b and the "camera height" of the three-dimensional point group calculation 12d are determined as parameters to be optimized.

<Object Detection Processing of Embodiment 1>
FIG. 9 is a flowchart illustrating object detection processing according to the first embodiment. The object detection process of the first embodiment is executed by the object detection device 1 at a predetermined timing of a user instruction.

As shown in FIG. 9, first, in step S11, the processing block decomposition unit 2 decomposes the sensor-specific measurement algorithm included in the sensor information 200 into a plurality of processing blocks.

Subsequently, the object detection device 1 repeats the loop processing of steps S12 to S14 for all processing blocks obtained by decomposing in step S11. That is, in step S12, the object detection device 1 selects one processing block obtained in step S11. Next, in step S13, the parameter contribution ratio calculator 3 calculates the contribution ratio of each parameter to the object detection algorithm 12 for one processing block selected in step S12. Next, in step S14, the analysis item contribution ratio calculator 5 calculates the contribution ratio of each analysis item to the object detection algorithm 12 for one processing block selected in step S12.

When step S14 ends, the object detection device 1 selects an unselected processing block from among the processing blocks obtained in step S11 in step S12, and performs the processing of steps S13 and S14 for the selected processing block. Run. Note that when step S14 ends and there is no unselected processing block among the processing blocks obtained in step S11, the object detection device 1 shifts the processing to step S15.

In step S15, the optimization processing block determination unit 6 extracts the analysis item extracted from the user requirement 201 by the user analysis item extraction unit 4, and the contribution of each analysis item calculated by the analysis item contribution ratio calculation unit 5 in step S14. Using the rate, the processing block with the largest contribution rate of the analysis items required by the user requirements 201 is determined as the optimization processing block to be optimized.

Subsequently, in step S16, the parameter selection unit 7 selects a parameter having a high contribution rate to the processing block for which the optimization processing block determination unit 6 determined in step S15 to perform the optimization processing as a parameter to be optimized. .

Subsequently, in step S17, the parameter adjustment unit 8 adjusts the parameters selected in step S16 so as to improve the object detection accuracy and the analysis accuracy corresponding to the analysis items.

Subsequently, in step S18, the object detection executing section 9 executes object detection using the parameters adjusted in step S17. Subsequently, in step S19, the analysis execution unit 10 uses the parameters adjusted in step S17 to execute an application according to the user requirements 201 using the result of object detection by the object detection execution unit 9 in step S18. do. When step S19 ends, the object detection device 1 ends the object detection process of the first embodiment.

According to the above-described first embodiment, in the object detection device 1 that detects an object from the measurement range, the contribution rate of the analysis item and the contribution rate of the parameter for each processing block obtained by dividing the algorithm unique to the measurement device are calculated, and the adjustment target A parameter is selected for each user requirement 201 . That is, when executing an application that satisfies the user requirements 201, a processing block with a high contribution rate of analysis items required for realizing this application is selected, and a parameter with a high contribution rate to the selected processing block is selected. This makes it possible to appropriately select parameters to be adjusted that are suitable for achieving user requirements for each measurement purpose.

In Example 1, as shown in FIG. 1, the structure which implement|achieves each function by the integrated object detection apparatus 1 was shown. However, the present invention is not limited to this, and the object detection device 1B of the second embodiment may have only the object detection execution unit 9 and the analysis execution unit 10 as shown in FIG. A parameter selection device 50B, which is a separate device, includes a processing block decomposition unit 2, a parameter contribution ratio calculation unit 3, a user analysis item extraction unit 4, an analysis item contribution ratio calculation unit 5, and an optimization processing block determination unit 6. , a parameter selection unit 7 and a parameter adjustment unit 8 . The object detection device 1B may execute object detection and analysis using the parameters selected and adjusted by the parameter selection device 50B, and output object detection result information 202 and analysis result information 203.

Furthermore, in the third embodiment, as shown in FIG. 11, another device has a processing block decomposition unit 2, a parameter contribution rate calculation unit 3, and an analysis item contribution rate calculation unit 5. By executing these processing functions, analysis item contribution rate information 204 and parameter contribution rate information 205 for each processing block are obtained. At the site, a parameter selection device 50C having a user analysis item extraction unit 4, an optimization processing block determination unit 6, and a parameter selection unit 7 selects user requirements 201, parameter contribution rate information 205, and analysis item contribution rate information 204. , the parameter to be adjusted can be selected at high speed for each user requirement 201 . An object detection device similar to the object detection device 1B of the second embodiment executes object detection and analysis using the parameters selected (or selected and adjusted) by the parameter selection device 50C, and collects object detection result information and analysis results. output information.

Further, in the first to third embodiments described above, a series of functions may be executed in consideration of the processing cost, and this embodiment will be described as the fourth embodiment. For example, FIG. 12 shows a configuration including the parameter selection device 50B and the object detection device 1B of the second embodiment shown in FIG. 10 in consideration of the processing cost.

In FIG. 12, the permissible processing cost acquisition unit 51 acquires the permissible processing cost for object detection and subsequent analysis application from the user requirements 201 and the spec information 207 of the object detection device 1D. The parameter selection device control unit 52 uses the processing cost information 206 when the object detection execution unit 9 and the analysis execution unit 10 are executed and the allowable processing cost information acquired by the allowable processing cost acquisition unit 51 to determine the parameter selection device 50B. functions are controlled.

Examples of inputs to the permissible processing cost acquisition unit 51 include user requirements 201 such as the amount of adjustment cost permissible at the site, whether the application executed by the object detection device 1D is to be processed in real time or offline, and the object In the detecting device 1D, there is hardware CPU and memory specification information 207 for executing object detection and other applications, and the unit of the processing cost is a processing frame rate (ms).

In addition, as a control method of the parameter selection device control unit 52, in addition to the amount of change in measurement accuracy due to changes in parameters when deriving the parameter contribution rate in the parameter contribution rate calculation unit 3, the amount of change in processing cost is also calculated. There is a method of reflecting it in the parameter contribution rate. Alternatively, a method of changing parameters based on processing costs such as the number of processing blocks to be optimized determined by the optimization processing block determination unit 6 and the number of parameters selected by the parameter selection unit 7, etc. Yes, but not particularly limited.

When selecting parameters, by selecting parameters that take into account the processing costs allowed for object detection and subsequent analysis applications, the efficiency of processing in the object detection device 1D is improved according to physical resources and performance. can.

<Modification>
Further, by changing the parameter adjustment unit 8 shown in FIGS. 1, 10, and 12 to a function that automatically optimizes parameters, a stereo camera that can achieve user requirements 201 without on-site parameter adjustment Eleven parameters can be estimated. In a system that requires installation of a large number of sensors (such as the stereo camera 11), the system cost can be significantly reduced.

As described above, in Examples 1 to 4, the method of selecting parameters related to the stereo camera object detection technique has been described, but the sensor is not limited as long as it performs object detection. Furthermore, in Examples 1 to 4, the parameter contribution rate and analysis item contribution rate to each processing block can be calculated using the final object detection accuracy as an index. If the sensor is capable of grasping SDK information and configurable parameter specifications, parameters to be adjusted can be selected for each user requirement. Therefore, it can be adapted to object detection devices using various sensors, and the versatility of the entire system can be improved.

Further, in the first to fourth embodiments, the method of selecting parameters related to applications related to sensor object detection technology has been described, but the technical idea of the present invention is not limited to object detection technology. For example, even if user requirements are applications related to object tracking technology, parameters to be adjusted in object tracking technology can be selected for each user requirement.

FIG. 13 is a hardware diagram showing the configuration of a computer that realizes the object detection device and parameter selection device of Examples 1-4. A computer 5000 that realizes the object detection device and the parameter selection device of Examples 1 to 4 includes an arithmetic device 5300 represented by a CPU (Central Processing Unit), a memory 5400 such as a RAM (Random Access Memory), an input device 5600 (for example keyboard, mouse, touch panel, etc.) and output devices 5700 (eg, a video graphics card connected to an external display monitor) are interconnected through memory controller 5500 . In the computer 5000, a program for realizing the object detection device or parameter selection device of Embodiments 1 to 4 is read from an external storage device 5800 such as an SSD or HDD via an I/O (Input/Output) controller 5200. Execution by cooperation of the CPU 5300 and the memory 5400 realizes the object detection device and the parameter selection device of the first to fourth embodiments. Alternatively, the program for realizing the object detection device and parameter selection device of Examples 1 to 4 may be acquired from an external computer through communication via network interface 5100. FIG.

In addition, the present invention is not limited to the above-described embodiments, and includes various modifications. For example, the above-described embodiments have been described in detail in order to explain the present invention in an easy-to-understand manner, and are not necessarily limited to those having all the described configurations. In addition, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. Moreover, it is possible to add, delete, replace, integrate, and distribute other configurations for a part of the configuration of each embodiment. Also, each process shown in the embodiment may be appropriately distributed or integrated based on processing efficiency or implementation efficiency.

1, 1B, 1D: object detection device, 2: processing block decomposition unit, 3: parameter contribution calculation unit, 4: user analysis item extraction unit, 5: analysis item contribution calculation unit, 6: optimization processing block determination unit , 7: parameter selection unit, 8: parameter adjustment unit, 9: object detection execution unit, 10: analysis execution unit, 11: stereo camera, 12: object detection algorithm, 12a: image acquisition, 12b: distortion correction, 12c: parallax Calculation, 12d: parallax calculation, 12c, 12d: 3D point cloud calculation, 12e: object detection, 13a-13e: parameters, T1: parameter contribution rate table, T1a-T1e: contribution rate table, 14: output result, 20: Parameter setting range determination unit 21: Measurement accuracy variation calculation unit 22: Parameter contribution rate table creation unit 30: Two-dimensional map 30A: Measurement range 31a, 31b: Equipment 32a, 32b: Measurement area 33: List 40: Analysis item list table 41: Combination table 50, 50B, 50C: Parameter selection device 51: Allowable processing cost acquisition unit 52: Parameter selection device control unit 200: Sensor information 201: User requirements 202: object detection result information; 203: analysis result information; 204: analysis item contribution rate information; 205: parameter contribution rate information; 206: processing cost information; : controller, 5300: CPU, 5400: memory, 5500: memory controller, 5600: input device, 5700: output device, 5800: external storage device

Claims (15)

A parameter used in an algorithm used by an object detection device to detect an object within a measurement range of the measurement device is a parameter for calculating a first degree of influence on the detection accuracy of detecting the object using the algorithm. an impact calculation unit;
an analysis item impact calculation unit that calculates a second impact that an analysis item for achieving a user requirement has on the measurement accuracy of the object detected by the object detection device;
A parameter selection device, comprising: a parameter selection unit that selects the parameter to be adjusted based on the first degree of influence and the second degree of influence. 2. The parameter selection device according to claim 1, further comprising a processing block decomposition unit that decomposes the algorithm into a plurality of processing blocks. The processing block decomposing unit,
The parameter selection device according to claim 2, wherein the algorithm is divided based on specification information of the algorithm or specification information related to the object detection device. The parameter influence degree calculation unit
3. The parameter selection device according to claim 2, wherein the first degree of influence on the algorithm or each processing block is calculated for each parameter used in each of the plurality of processing blocks. The analysis item impact degree calculation unit
5. The parameter selection device according to claim 4, wherein the second degree of influence on the algorithm or each processing block is calculated for each analysis item. The first influence degree and the second influence degree are calculated using the detection accuracy of the algorithm or the amount of variation in the output result of the processing block when the parameter is changed. 5. The parameter selection device according to 5. an optimization processing block determination unit that determines a processing block for which the second degree of influence satisfies a second predetermined condition as an optimization processing block;
The parameter selection unit
A parameter that satisfies a first predetermined condition for the first degree of influence on a processing block determined to be an optimization processing block by the optimization processing block determination unit is selected as a parameter to be optimized. 5. The parameter selection device according to 5. 2. The parameter selection device according to claim 1, further comprising a parameter adjustment unit that adjusts the parameter selected by the parameter selection unit so that the detection accuracy or the analysis accuracy is equal to or higher than a predetermined accuracy. The first degree of influence and the second degree of influence are calculated in advance,
The parameter adjustment unit
When the object detection device detects the object, the parameter is adjusted based on the information of the analysis item and the pre-calculated first degree of influence and the second degree of influence. The parameter selection device according to claim 8, characterized by: The parameter selection device according to claim 1, wherein the parameter selection unit selects the parameter based on a processing cost of the object detection device. 2. The parameter selection device according to claim 1, further comprising a user analysis item extraction unit that extracts the analysis item from the user's measurement purpose. 12. The parameter selection device according to claim 11, wherein the analysis item corresponds to an application that uses object detection results by the algorithm. The parameter selection device according to claim 1, wherein the measurement device is one of a monocular camera, a stereo camera, a TOF sensor, and a laser radar. A parameter selection method executed by a parameter selection device,
A parameter used in an algorithm used by an object detection device to detect an object within a measurement range of the measurement device is a parameter for calculating a first degree of influence on the detection accuracy of detecting the object using the algorithm. an impact calculation step;
an analysis item impact calculation step of calculating a second impact of an analysis item for achieving a user requirement on the measurement accuracy of the object detected by the object detection device;
and a parameter selection step of selecting the parameter to be adjusted based on the first degree of influence and the second degree of influence. the computer,
A parameter used in an algorithm used by an object detection device to detect an object within a measurement range of the measurement device is a parameter for calculating a first degree of influence on the detection accuracy of detecting the object using the algorithm. an impact calculation unit;
an analysis item impact calculation unit that calculates a second impact that an analysis item for achieving a user requirement has on the measurement accuracy of the object detected by the object detection device;
A parameter selection program for functioning as a parameter selection device, comprising: a parameter selection unit that selects the parameter to be adjusted based on the first degree of influence and the second degree of influence.

Images