Mirror image determination device, mirror image determination method, and mirror image determination program

Abstract

To provide a mirror image determination device, a mirror image determination method, and a mirror image determination program that improve reliability of a camera sensor in autonomous driving.SOLUTION: A mirror image determination device includes: an acquisition unit that acquires camera recognition data; a camera recognition label generation unit that generates a camera recognition label on the basis of the camera recognition data; a vehicle data extraction unit that reads the camera recognition label line by line and extracts the line data of the image according to a vehicle type; an information storage unit that temporarily stores the line data as a table in a storage device; a glossiness determination unit that determines whether or not a target reflected in the vehicle included in the camera recognition data has glossiness; and a mirror image determination unit that determines whether the target is a mirror image on the basis of a positional relation with the mirror image determination unit, which determines whether the target is a mirror image on the basis of a positional relation between bounding boxes attached to the image.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to a mirror image determination device, a mirror image determination method, and a mirror image determination program.

Background Art

[0002] Vehicles on the side with a large specular reflection area often form mirror images of objects passing nearby. The presence of such mirror images may cause unexpected behaviors in the automatic driving of vehicles, such as stopping earlier than expected, deviating from the diagonal line to avoid a collision with a misdetected object, etc. As techniques for detecting mirror images in camera data, there are techniques disclosed in Patent Documents 1 to 3, Non-Patent Documents 1 to 2, and the like.

[0003] Patent Document 1 discloses a technique that enables appropriate recognition of each of a plurality of recognition targets even when the plurality of recognition targets overlap each other. Patent Document 2 discloses a technique for calibrating a television image by creating a mirror image of a display image using a mirror attached at a specific position, detecting the difference between the signal of the mirror image and the signal of the transmitted image, and correcting the image. Patent Document 3 discloses a technique for improving the reliability in face image recognition, which enhances the reliability by comparing the outputs of both a face image recognition learned using teacher data and a mirror image detector learned using a substantially frontal image of the teacher data. In Non-Patent Document 1, a method for determining the glossiness of an object based on the standard deviation and skewness of the luminance of an image is proposed. In Non-Patent Document 2, a technique for detecting glossiness using a DNN (Deep Neural Network) is reported.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Patent Document 3

Non-Patent Literature

[0005]

Non-Patent Literature 1

Non-Patent Literature 2

Summary of the Invention

Problems to be Solved by the Invention

[0006] As a method for detecting a mirror image in camera data, deep learning has been proposed as disclosed in Non-Patent Literature 1. However, the training data used in such supervised learning is significantly different from the camera data of the mirror image reflected in a vehicle in terms of the field of view and the type of object, etc. Therefore, if applied as it is, performance beyond what is required will be demanded. Even if learning the mirror image reflected in a vehicle from scratch, it is difficult to obtain sufficient training data for the mirror image that appears on a vehicle with a mirror finish, and there are also few opportunities to capture mirror images during daily driving. Further, since it uses a supervised learning technique that requires a huge amount of teacher data as a premise, the computational load is enormous and it is difficult to install it in a general vehicle when incorporating it.

[0007] In the automatic driving of a vehicle, although it is possible to identify shiny areas in the image acquired by a sensor, it is not possible to detect a vehicle and the mirror image of a moving object target reflected in that vehicle only with this technology. Further, since it uses a supervised learning technique that requires a huge amount of teacher data as a premise, the computational load is enormous and it is difficult to install it in a general vehicle.

[0008] In view of the above problems, an object of the present invention is to provide a mirror image determination device, a mirror image determination method, and a mirror image determination program that improve the reliability of a camera sensor in automatic driving.

Means for Solving the Problems

[0009] A first aspect of the present invention is a mirror image determination device, comprising: an acquisition unit that acquires camera recognition data; a camera recognition label generation unit that generates a camera recognition label based on the camera recognition data; a vehicle data extraction unit that reads the camera recognition label line by line and extracts image line data corresponding to the type of vehicle; an information writing unit that temporarily stores the line data in a storage device as a table; a glossiness determination unit that determines the presence or absence of glossiness of the vehicle included in the camera recognition data; and a mirror image determination unit that determines an object as a mirror image based on the positional relationship between bounding boxes attached to the image.

[0010] In the first aspect of the present invention, a setting unit for setting the bounding box in the image may be further provided.

[0011] In the first aspect of the present invention, an output unit may be further provided that collates the line number in the line data of the determination result by the mirror image determination unit with the line number included in the line data of the camera recognition label and outputs it to a sensor weakness file as mirror image data.

[0012] In the first aspect of the present invention, when one of the plurality of bounding boxes is included in the other, the mirror image determination unit may determine the completely included bounding box as a mirror image.

[0013] In the first aspect of the present invention, the line data of the camera recognition label may have a frame number of a still image, a type of object, coordinates, and a size included in the camera recognition data.

[0014] A second aspect of the present invention is a mirror image determination method, which includes an acquisition step of acquiring camera recognition data, a camera recognition label generation step of generating a camera recognition label based on the camera recognition data, a vehicle data extraction step of reading the camera recognition label line by line and extracting image line data corresponding to the type of vehicle, an information writing step of temporarily storing the line data in a storage device as a table, a glossiness determination step of determining the presence or absence of glossiness of the vehicle included in the camera recognition data, and a mirror image determination step of determining that an object is a mirror image based on the positional relationship between the bounding boxes attached to the images.

[0015] A third aspect of the present invention is a mirror image determination program, which causes a computer to realize an acquisition function of acquiring camera recognition data, a camera recognition label generation function of generating a camera recognition label based on the camera recognition data, a vehicle data extraction function of reading the camera recognition label line by line and extracting image line data corresponding to the type of vehicle, an information writing function of temporarily storing the line data in a storage device as a table, a glossiness determination function of determining the presence or absence of glossiness of an object reflected in the vehicle included in the camera recognition data, and a mirror image determination function of determining that an object is a mirror image based on the positional relationship between the bounding boxes attached to the images.

Advantages of the Invention

[0016] According to the present invention, it is possible to provide a mirror image determination device, a mirror image determination method, and a mirror image determination program that improve the reliability of a camera sensor in autonomous driving.

Brief Description of the Drawings

[0017]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Mode for Carrying Out the Invention

[0018] Next, embodiments of the present invention will be described with reference to the drawings. In the description of the drawings according to the embodiments, the same or similar parts are denoted by the same or similar reference numerals. However, it should be noted that the drawings are schematic, and the relationships such as planar dimensions are different from the actual ones. Therefore, specific dimensions should be determined with reference to the following description. Of course, there are also parts where the dimensional relationships and ratios are different between the drawings.

[0019] Furthermore, the embodiments illustrate devices and methods for embodying the technical idea of the present invention, and the technical idea of the present invention does not specify the configurations, arrangements, layouts, etc. of the respective components as the following. The technical idea of the present invention can be variously modified within the technical scope defined by the claims described in the claims.

[0020] (Embodiment) The mirror image determination device according to this embodiment is a device that efficiently searches for mirror images reflected in a vehicle that induces false detection of targets in autonomous driving, using camera recognition data as input. The mirror image determination device according to this embodiment, for the camera recognition data after object detection, 1) extracts the bbox (bounding box, the same hereinafter) information of the data with the label of the vehicle, 2) extracts the partial image data of the bbox of the vehicle, 3) determines the presence or absence of the glossiness of the partial image, 4) searches for the bbox within the bbox of the vehicle with glossiness, 5) regards the bbox within the bbox as a mirror image, and outputs the original image ID, label and bbox information, and the mirror image flag to a file, thereby searching for the mirror image reflected in the vehicle. According to the mirror image determination device according to this embodiment, it is possible to determine the mirror image of a vehicle with glossiness without performing learning using teacher data. The details are described below.

[0021] An example of the usage situation of the mirror image determination device according to this embodiment is shown in FIG. 1. Specifically, assuming an autonomous driving scenario, a camera 1 is installed in a vehicle 2 such as a passenger car. The camera 1 photographs the front of the vehicle 2 and generates camera recognition data. In the figure, the vehicle 2 is traveling on the road surface 4 and is photographing the front of the vehicle 2. Another vehicle 5 is traveling in front of the vehicle 2. In addition to this, it is also possible to photograph the entire periphery of the vehicle 2. The mirror image determination device 10 is connected to the camera 1 by wire or wirelessly, or via an Internet line or the like.

[0022] The block diagram of FIG. 2 is the mirror image determination device 10 of the embodiment, which is connected to the aforementioned camera 1. The camera 1 is provided with a light receiving unit, a conversion unit that converts the received light into image data, a communication unit that is responsible for communication with the mirror image determination device 10, a storage unit that stores the image data, a control unit that controls the camera 1, and the like are appropriately provided. In FIG. 2, it is an example in which the mirror image determination device 10 is directly connected to the camera 1 to acquire image data. In addition to this, the image data may be stored in an appropriate storage medium by the data measured by the camera 1 and acquired by the mirror image determination device 10 afterwards, or may be acquired through other devices or networks via wireless communication or the like.

[0023] The mirror determination device 10 (processing unit, computer) is hardware-wise composed of arithmetic elements such as a CPU and a GPU, and storage elements such as a RAM, a ROM, a HDD, and an SSD. Further, it is provided with various interfaces for signal reception from the camera 1 and input / output to the outside of the mirror determination device 10. Also, software-wise, each function is realized by a program loaded into the main memory. As the recording medium for storing this program, a non-transitory tangible medium such as a CD, a DVD, a semiconductor memory, a programmable logic circuit, etc. can be used. Also, this program may be supplied to the mirror determination device 10 (computer) via any transmission medium (communication network, broadcast wave, etc.) capable of transmitting the program.

[0024] The mirror determination device 10 (processing unit, computer) in the block diagram of FIG. 2 schematically shows the functional units of the same device. The same device 10 includes functional units such as an acquisition unit 100, a camera recognition label generation unit 110, a vehicle data extraction unit 120, an information storage unit 130, a glossiness determination unit 140, a mirror determination unit 150, and an output unit 160.

[0025] The acquisition unit 100 acquires camera recognition data generated by the camera 1. The camera recognition data is a video file photographed by the camera 1.

[0026] The camera recognition label generation unit 110 extracts a target from the image data of one frame in the camera recognition data acquired by the acquisition unit 100 by a recognition algorithm, and generates a camera recognition label in which a class indicating the type of the extracted target, the center coordinates and dimensions of a bounding box (hereinafter referred to as bbox) indicating the existence region of the target, and the frame number in the camera recognition data in which the target appears are described. Here, the mirror determination device 10 of the embodiment may further include a setting unit that sets a bounding box in the image in the camera recognition data acquired by the acquisition unit 100 that has been acquired.

[0027] In this embodiment, the camera recognition label is assumed to be text-formatted data having row data including the frame number of the camera recognition data, the class indicating the type of the target, and the position and dimensions of the bbox that specifies the location within the image. However, the camera recognition label may be data in any format, either text format or binary format, and may be one file per piece. As an example of text-formatted data having row data, for example, "[row data]": "[time]-[center coordinates]-[size]-…" etc. may be mentioned. As an example, the data layout of the camera recognition label is [frame number], [class of the target], [bbox center coordinate x], [bbox center coordinate y], [dimension of the bbox (width w)], [dimension of the bbox (height h)], [confidence]. Here, the frame number is the number of the frame of each image of the camera recognition data which is a video file. The class of the target is the type of the target, for example, a passenger car, a bus, a truck, etc. The confidence is the probability that the target is reflected within the bbox.

[0028] Note that, as will be described later, as long as the input processing is modified, the bbox may have a shape other than a rectangle, such as a polygon or pixels, in order to more accurately represent the existence area of the target. The above data layout example is the case where the bbox is a rectangle.

[0029] The generation of the camera recognition label by the camera recognition label generation unit 110 is executed for all frames of the camera recognition data acquired by the acquisition unit 100. As a result, in the camera recognition label, the above-described data is used as row data, and all frames and the data corresponding to the targets included in each frame are described.

[0030] The vehicle data extraction unit 120 reads the camera recognition label one row at a time, and when the class of the target in the row data is a vehicle having a gloss and capable of transferring the mirror image of another vehicle to the vehicle body, it delivers the row data to the information storage unit 130. Examples of vehicles having a gloss and capable of transferring the mirror image of another vehicle to the vehicle body include a passenger car, a bus, a truck, etc.

[0031] The information storage unit 130 temporarily stores the row data received from the vehicle data extraction unit 120 in a storage device as a table.

[0032] The extraction of information by the vehicle data extraction unit 120 and the information storage unit 130 is executed for all the objects described in the camera recognition label.

[0033] The glossiness determination unit 140 takes as input the camera recognition label, the information extracted by the vehicle data extraction unit 120 and the information storage unit 130, and the camera recognition data, determines the mirror image of the object imaged in the vehicle included in the camera recognition data, and temporarily stores the determination result in the storage device. Details will be described below.

[0034] The glossiness determination unit 140 performs processing for each piece of target information. The glossiness determination unit 140 extracts the frame number of the camera recognition label according to the row number of the camera recognition label described in the row data of the information to be extracted, and cuts out the target frame of the camera recognition data as a temporary image. Next, the standard deviation of the luminance of the region of the temporary image corresponding to the bbox shown in the camera recognition label of the same row is calculated. Finally, the region of the same temporary image is converted into edge information by a Laplacian filter, and its variance (hereinafter, this variance is referred to as a blur score) is calculated. If both the blur score and the standard deviation of the luminance exceed a predetermined threshold value, it is regarded as having glossiness, and the row number of the camera recognition label in which the object is detected is passed to the mirror image determination unit.

[0035] The calculation methods of the blur score and the standard deviation of the luminance are as follows. In Non-Patent Document 1, the correlation between human gloss perception and the tuple of the standard deviation σ L of luminance and the skewness of luminance (hereinafter referred to as the "standard deviation - skewness scale") is reported. However, in the mirror image determination device 10 according to the present embodiment, the presence or absence of glossiness is determined using the blur score and the standard deviation of the luminance. The standard deviation σ L of luminance is defined by the following formula.

[0036]

Equation

[0037] Here, L is the luminance image of the target bbox, i is the pixel index, and μ L is the average value of L, and n L is the number of pixels of L. μ L is defined by the following formula.

[0038]

Equation

[0039] The blur score, which is the variance of the edge intensity of the luminance image, is defined by the following formula.

[0040]

Equation

[0041] E is the edge intensity of the luminance image, and μ E and n E are the average of E and the number of pixels of E, respectively. E is calculated by applying an 8-neighborhood Laplacian filter to the luminance image within bbox(L) using the following formula. K8 is the kernel (coefficient matrix) of the filter.

[0042]

Equation

[0043] The asterisk operator (*) represents the convolution product operation, and K8 is represented by the following formula.

[0044]

Equation

[0045] In the blur score B and the standard deviation σ of luminance defined above L for the detection of the mirror-finished vehicle, the standard deviation σ LThe range of A is 40 to 60, and the threshold value of the blur score B is appropriately 0 to 300. The mirror image determination device 10 according to the present embodiment performs the detection process of the mirror-finished vehicle based on the above threshold value.

[0046] The mirror image determination unit 150 extracts the position and dimensions of the bbox of the corresponding row from the row number of the camera recognition label received from the glossiness determination unit 140. Let this bbox be bbox0. Extract the entire row of the camera recognition label including the frame number of the camera recognition data included in the same row. Extract the position and dimensions of the bbox included in the row of the extracted camera recognition label. Let these bboxes be bboxi (1 ≤ i ≤ n). Examine the positional relationship between bbox0 and bboxi. If either one is included in the other, determine that the object indicated by the completely included bbox is a mirror image. If they are not completely included and the bboxes overlap, determine that it is not a mirror image. Extract the row data including the bbox determined to be a mirror image in the camera recognition label and temporarily store it in the storage device as a determination result table.

[0047] The output unit 160 collates the row number in the row data of the determination result with the row number included in the row data of the camera recognition label and outputs it to the sensor weakness file as mirror image data. Specifically, the output unit 160 identifies the row number of the camera recognition label from the row number in the row data of the determination result, that is, the row number of the row data in the camera recognition label determined to be a mirror image. Next, the row number of the camera recognition label, the file name of the camera recognition label, the camera recognition video, and the string of "mirror image" as the type of sensor weakness included in the row data of the camera recognition label with the same row number are output as the row data of the sensor weakness file. of the string is output as the row data of the sensor weakness file.

[0048] Figures 3 and 4 show an example of a mirror image of a vehicle reflected in a shiny vehicle. Figure 3 is an image showing actual camera image data in grayscale. The bbox detected by the mirror image determination device according to the present embodiment is displayed in both Figures 3 and 4. Bbox 31 is a bbox extracted for a passenger car 34 which is an actual target. Bbox 32 is a bbox determined to be a target with a shiny feeling, which is a bus 35. Bbox 33 is a bbox determined to be a mirror image 36 of the passenger car 34 reflected on the side surface of the bus 35. In the examples shown in Figures 3 and 4, it is assumed that all bboxes are rectangles.

[0049] The camera recognition labels of these targets are generated by the camera recognition label generation unit 110, and the information of the passenger car 34, the bus 35, the mirror image 36 which are targets, and the information of the bbox are extracted by the vehicle data extraction unit 120 and the information storage unit 130. The presence or absence of the shiny feeling of each target is determined by the shiny feeling determination unit 140. In the examples shown in Figures 3 and 4, the bus 35 is determined to be a target with a shiny feeling. The mirror image determination unit 150 determines that the bbox 33 is completely included in the bbox 32, that is, the mirror image 36 which is the target indicated by the bbox 33 is determined to be a mirror image.

[0050] Figure 5 shows an example of a mirror image of a bicycle reflected in a shiny vehicle. The bbox detected by the mirror image determination device according to the present embodiment is also displayed in Figure 5. Bbox 51 is a bbox determined to be a target with a shiny feeling, which is a bus 56. Bbox 52 is a bbox of an actual target, which is a bicycle 57, and bbox 53 is a bbox of a bicycle 58. Bbox 54 is a bbox determined to be a mirror image 59 of the actual target, which is a bicycle 57, reflected on the side surface of the bus 56, and bbox 55 is a bbox determined to be a mirror image 60 of the bicycle 58 reflected on the side surface of the bus 56. In the example shown in Figure 5, it is assumed that all bboxes are rectangles. Similar to the examples shown in Figures 3 and 4, these targets and bboxes are detected by the mirror image determination device according to the present embodiment, and the mirror images 59 and 60 are determined to be mirror images.

[0051] FIG. 6 shows an example of a target located in front of a shiny vehicle. In the examples shown in FIGS. 3, 4, and 5, all the bounding boxes (bboxes) were assumed to be rectangles, but in the example shown in FIG. 6, the bbox is not limited to a rectangle and is assumed to include polygons. The bbox 61 is a bbox in which the bus 63, which is an actual target, is determined to be a shiny target, and the bbox 61 has a trapezoidal shape. The bbox 62 is the bbox of the motorcycle 64, which is an actual target. The bbox 62 is located in front of the bbox 61, but the bbox 62 is not completely enclosed by the bbox 61. Therefore, the mirror image determination unit 150 determines that the motorcycle 64, which is the target indicated by the bbox 62, is not a mirror image.

[0052] The mirror image determination method according to the present embodiment will be described with reference to the flowcharts shown in FIGS. 7 to 10. FIGS. 7 to 10 are flowcharts for explaining the procedure in which the mirror image determination device according to the present embodiment acquires camera recognition data generated by the camera 1 and detects a mirror image reflected on the vehicle.

[0053] As shown in FIG. 7, the mirror image determination method according to the present embodiment includes three steps: an input step (step S701), a determination step (step S702), and an output step (step S703).

[0054] In step S701, using the camera recognition data as an input, information on the extraction target to be the subject of the determination process in step S702 is generated and temporarily stored in the storage device.

[0055] In step S702, the gloss determination unit 140 determines a mirror image of a target reflected in the vehicle included in the camera recognition data, using the camera recognition label, the information on the extraction target, and the camera recognition data as inputs, and temporarily stores the determination result in the storage device.

[0056] In step S703, the line number in the row data of the determination result is collated with the line number included in the row data of the camera recognition label, and the result is output as mirror image data to the sensor vulnerability file.

[0057] Figures 8 to 10 are flowcharts for explaining the details of each of the three steps shown in FIG. 7.

[0058] FIG. 8 explains the details of the input step (step S701) shown in FIG. 7.

[0059] In step S801, the acquisition unit 100 acquires camera recognition data generated by the camera 1.

[0060] In step S802, the camera recognition label generation unit 110 generates a camera recognition label.

[0061] In step S803, the vehicle data extraction unit 120 reads the camera recognition labels one by one and delivers the row data of passenger cars, buses, and trucks to the information storage unit 130.

[0062] In step S804, the information storage unit 130 temporarily stores the row data received from the vehicle data extraction unit 120 in the storage device as a table.

[0063] Steps S803 to S804 are repeatedly executed for all targets.

[0064] FIG. 9 explains the details of the determination step (step S702) shown in FIG. 7.

[0065] In step S901, the acquisition unit 100 acquires the information to be extracted.

[0066] In step S902, the acquisition unit 100 acquires the camera recognition label and the camera recognition data.

[0067] In step S903, the glossiness determination unit 140 extracts the frame number of the camera recognition label according to the row number of the camera recognition label described in the row data of the information to be extracted, and cuts out the target frame of the camera recognition data as a temporary image.

[0068] In step S904, the glossiness determination unit 140 calculates the standard deviation of the luminance of the region of the temporary image corresponding to the bbox indicated by the camera recognition label in the same row, and calculates the blur score for the same region of the temporary image.

[0069] In step S905, if both the blur score and the standard deviation of the luminance exceed a predetermined threshold, it is determined that there is a glossiness, and the line number of the camera recognition label where the object target was detected is passed to the mirror image determination unit.

[0070] In step S906, the position and dimensions of the bbox in the corresponding row are extracted from the line number of the camera recognition label received from the glossiness determination unit 140.

[0071] In step S907, the positional relationship between bbox0 and bboxi is examined to determine whether it is a mirror image.

[0072] In step S908, the row data including the bbox determined to be a mirror image in the camera recognition label is extracted and temporarily stored in the storage device as a determination result table.

[0073] Steps S902 to S908 are repeatedly executed for all extraction targets.

[0074] FIG. 10 explains the details of the output step (step S703) shown in FIG. 7.

[0075] In step S1001, the output unit 160 identifies the line number of the camera recognition label from the line number in the line data of the determination result.

[0076] In step S1002, the output unit 160 outputs, as the line data of the sensor weakness file, the line number of the camera recognition label, the file name of the camera recognition label, the camera recognition video, and the string "mirror image" as the type of sensor weakness included in the line data of the camera recognition label with the same line number. of the string as the line data of the sensor weakness file.

[0077] The computer program of the present invention described above may be recorded on a processor-readable recording medium. As the recording medium, a "non-transitory tangible medium" , for example, a disk, a card, a semiconductor memory, a programmable logic circuit, etc. can be used.

[0078] This device is software having three processes of input, determination, and output, and is implemented as a program in the Python language. However, as long as it is a language capable of numerical analysis and data processing, it can be implemented in other languages such as C and Java. Also, as long as a language capable of numerical analysis and data processing can be executed, the type of OS installed in the computer does not matter.

[0079] As described above, according to the mirror image determination device according to the present embodiment, the mirror image that is the source of misrecognition in the camera recognition data can be removed, and the reliability of the camera sensor alone in autonomous driving can be improved. Also, without going through learning using teacher data, if camera recognition data as input is given, the mirror image of the moving object target reflected in the vehicle can be immediately determined. Since it is light as compared with large-scale models such as deep learning, it is easy and fast to incorporate into an environment with limited resources. Furthermore, the data labeled and accumulated by the present invention can also be used as teacher data for supervised learning, and by dramatically reducing the man-hours required for annotation of teacher data, development of a more accurate machine learning model can be realized in a short period of time.

[0080] Of course, the present invention includes various embodiments and the like not described herein. Therefore, the technical scope of the present invention is defined only by the invention-specific matters according to the appropriate claims from the above description.

Description of Reference Numerals

[0081] 1 Camera 2 Vehicle 4 Road surface 5 Other vehicle 10 Mirror image determination device 11 Light receiving unit 12 Conversion Unit 13 Communication Unit 14 Memory Unit 15 Control Unit 100 Acquisition Unit 110 Camera Recognition Label Generation Unit 120 Vehicle Data Extraction Unit 130 Information Memory Unit 140 Glossiness Judgment Unit 150 Mirror Image Judgment Unit 160 Output Unit 34 Passenger Car 35 Bus 36 Mirror Image

Claims

1. An acquisition unit that acquires camera recognition data, A camera recognition label generation unit that generates a camera recognition label based on the camera recognition data, A vehicle data extraction unit that reads the camera recognition labels one by one and extracts row data of an image according to the type of vehicle, An information storage unit that temporarily stores the row data as a table in a storage device, A glossiness determination unit that determines the presence or absence of glossiness of a vehicle included in the camera recognition data, A mirror image determination unit that determines a target as a mirror image based on the positional relationship between bounding boxes attached to the image A mirror image determination device characterized by comprising the above.

2. The mirror image determination device according to claim 1, further comprising a setting unit that sets the bounding box in the image.

3. The mirror image determination device according to claim 1, further comprising an output unit that collates the line numbers in the row data of the determination result with the line numbers included in the row data of the camera recognition label and outputs them to a sensor weakness file as mirror image data.

4. The mirror image determination device according to claim 1, wherein when one of the plurality of bounding boxes is included in the other, the mirror image determination unit determines the completely included bounding box as a mirror image.

5. The row data of the camera recognition label has a frame number of a still image included in the camera recognition data, a type of target, coordinates, and a size, according to the mirror image determination device according to claim 1.

6. An acquisition step of acquiring camera recognition data, A camera recognition label generation step of generating a camera recognition label based on the camera recognition data, A vehicle data extraction step of reading the camera recognition labels one by one and extracting row data of an image according to the type of vehicle, An information writing step of temporarily storing the row data as a table in a storage device, A glossiness determination step of determining the presence or absence of glossiness of a vehicle included in the camera recognition data, A mirror image determination step of determining a target as a mirror image based on the positional relationship between bounding boxes attached to the image A mirror image determination method characterized by comprising the above.

7. On a computer, An acquisition function of acquiring camera recognition data, A camera recognition label generation function of generating a camera recognition label based on the camera recognition data, A vehicle data extraction function that reads the camera recognition labels one by one and extracts row data of images according to the type of vehicle, An information writing function that temporarily stores the row data as a table in a storage device, A glossiness determination function that determines the presence or absence of glossiness of an object imaged on a vehicle included in the camera recognition data, A mirror image determination function that determines an object as a mirror image based on the positional relationship between the bounding boxes attached to the image A mirror image determination program characterized by realizing the above functions.

Images