Object detection device and program

Abstract

To provide an object detection apparatus and a program configured to suppress false detection.SOLUTION: An object detection apparatus 100 includes: an image adjustment unit 24 which generates adjustment data which is sensor data obtained by adjusting data with respect to sensor data acquired from a sensor which is in a constant relative positional relation with an object; a detection unit 26 which detects the object from the adjustment data and detects the object from sensor data before adjustment; a classification unit 34 which classifies each of a detection result from the adjustment data and a detection result from the sensor data before adjustment, based on regions where the object has been detected, into a removed candidate indicating removing the detection result and a maintained candidate indicating remaining the detection result; and an integration unit 36 which integrates the detection result from the adjustment data with the detection result from the sensor data before adjustment, on the basis of a classification result regarding the detection result from the adjustment data and a classification result regarding the detection result from the sensor data before adjustment.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an object detection device and a program, and more particularly to an object detection device and a program for detecting an object from sensor data. [Background technology]

[0002] Image processing devices have been known for some time that address the problem of falsely detecting an object that is not a specific object, such as a poster with a person in it, as a specific object (Patent Document 1). This image processing device reduces false detections by setting a high detection threshold within a specific area around the object that is falsely detected, and resets the detection threshold to its original value when it detects that a moving object has entered the specific area using a technique such as background subtraction, thereby correctly detecting the object to be detected when it passes through the specific area. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-125058 Summary of the Invention [Problem to be solved by the invention]

[0004] The technology described in Patent Document 1 above may not be able to handle slow-moving specific objects that cannot be detected as moving objects. Furthermore, when the detection threshold for specific regions is restored, it cannot handle the case where a partially occluded object that is not a specific object is still erroneously detected in addition to the specific object.

[0005] The present invention has been made to solve the above-mentioned problems, and has an object to provide an object detection device and a program that can suppress erroneous detection. [Means for solving the problem]

[0006] In order to achieve the above object, the object detection device of the present invention includes a data adjustment unit that adjusts sensor data acquired by a sensor that has a constant relative positional relationship with an object, and generates adjusted data, which is sensor data obtained by adjusting the data; a detection unit that detects the object from the adjusted data and detects the object from the sensor data before adjustment; a classification unit that classifies the detection results from the adjusted data into removal candidates where the detection results should be removed and maintenance candidates where the detection results should be maintained, based on the detected area, and classifies the detection results from the sensor data before adjustment into the removal candidates and the maintenance candidates, based on the detected area; and an integration unit that integrates the detection results from the adjustment data and the detection results from the sensor data before adjustment, based on the classification results of the detection results from the adjustment data and the classification results of the detection results from the sensor data before adjustment.

[0007] The program of the present invention causes a computer to function as a data adjustment unit that adjusts sensor data acquired by a sensor that has a constant relative positional relationship with an object and generates adjusted data, which is the adjusted sensor data; a detection unit that detects the object from the adjusted data and detects the object from the sensor data before adjustment; a classification unit that classifies the detection results from the adjusted data into removal candidates where the detection results should be removed and maintenance candidates where the detection results should be maintained based on the detected area, and classifies the detection results from the sensor data before adjustment into the removal candidates and the maintenance candidates based on the detected area; and an integration unit that integrates the detection results from the adjustment data and the detection results from the sensor data before adjustment based on the classification results of the detection results from the adjustment data and the classification results of the detection results from the sensor data before adjustment.

[0008] According to the present invention, a data adjustment unit adjusts sensor data acquired by a sensor having a constant relative positional relationship with an object, and generates adjusted data, which is sensor data obtained by adjusting the data. A detection unit detects the object from the adjusted data and detects the object from the pre-adjustment sensor data. A classification unit then classifies the detection results from the adjusted data into removal candidates for which the detection results should be removed and maintenance candidates for which the detection results should be maintained, based on the detected area, and classifies the detection results from the pre-adjustment sensor data into the removal candidates and the maintenance candidates, based on the detected area. An integration unit integrates the detection results from the adjusted data and the detection results from the pre-adjustment sensor data based on the classification results of the detection results from the adjusted data and the classification results of the detection results from the pre-adjustment sensor data.

[0009] In this way, data is adjusted for sensor data, objects are detected from the adjusted data and the sensor data before adjustment, the detection results from the adjusted data are classified into removal candidates and retention candidates, the detection results from the sensor data before adjustment are classified into removal candidates and retention candidates, and based on the classification results, the detection results from the adjusted data and the detection results from the sensor data before adjustment are integrated, thereby suppressing false detections.

[0010] The integration unit according to the present invention can obtain the detection result of the object by integrating the removal candidates from the detection results from the adjustment data and the retention candidates from the detection results from the sensor data before the adjustment.

[0011] The object detection device of the present invention further includes a collection unit that collects false-detection sensor data, which is sensor data for which the detection result by the detection unit is incorrect, based on whether or not the detection result by the detection unit is a region detected in a background region, and a parameter calculation unit that calculates adjustment parameters based on the detection result by the detection unit for the collected false-detection sensor data so that a correct detection result is obtained for the false-detection sensor data, and the classification unit can classify the detection results from the sensor data before adjustment into removal candidates and retention candidates based on whether or not the detected region is a background region, and can classify the detection results from the adjustment data into removal candidates and retention candidates based on whether or not the detected region is a background region.

[0012] The parameter calculation unit according to the present invention can further calculate classification parameters based on the detection results from the sensor data before adjustment by the detection unit for the collected false detection sensor data, so that the detection results for the false detection sensor data are classified as candidates for removal.

[0013] The sensor data according to the present invention is made up of elements, and the data adjustment unit can adjust the sensor data for each of the elements.

[0014] The data adjustment unit according to the present invention adjusts the data by adding noise to each of the elements, and the parameter calculation unit can calculate the noise to be added to each of the elements as the adjustment parameter.

[0015] The sensor may be an imaging device, and the sensor data may be an image. The imaging device may be installed in a fixed position.

[0016] The data adjustment unit adjusts the data by adding noise to each of the elements, and the parameter calculation unit can calculate the noise to be added to each of the elements as the adjustment parameter.

[0017] The storage medium for storing the program of the present invention is not particularly limited and may be a hard disk or a ROM. It may also be a CD-ROM, a DVD disk, a magneto-optical disk, or an IC card. Furthermore, the program may be downloaded from a server or the like connected to a network. [Effects of the Invention]

[0018] As described above, according to the present invention, data is adjusted for sensor data, objects are detected from the adjusted data and the sensor data before adjustment, the detection results from the adjusted data are classified into removal candidates and retention candidates, the detection results from the sensor data before adjustment are classified into removal candidates and retention candidates, and the detection results from the adjusted data and the detection results from the sensor data before adjustment are integrated based on the classification results, thereby achieving the effect of suppressing erroneous detections. [Brief explanation of the drawings]

[0019] [Figure 1] 1 is a block diagram showing a configuration of an object detection device according to an embodiment of the present invention. [Figure 2] FIG. 10 is a diagram illustrating an example of an erroneously detected image. [Figure 3] FIG. 10 is a diagram illustrating an example of a detection result for a captured image after adjustment. [Figure 4] FIG. 10 is a diagram illustrating an example of a detection result for a captured image before adjustment. [Figure 5] 10A and 10B are diagrams illustrating an example of integrating a detection result for a captured image after adjustment and a detection result for a captured image before adjustment. [Figure 6] FIG. 10 is a diagram showing how object detection results are output from each object detection device to a traffic control system. [Figure 7] 4 is a flowchart showing the contents of an erroneously detected image collection processing routine in the object detection device according to the embodiment of the present invention. [Figure 8] 4 is a flowchart showing the contents of an adjustment parameter calculation processing routine in the object detection device according to the embodiment of the present invention. [Figure 9] 4 is a flowchart showing the contents of an object detection processing routine in the object detection device according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0020] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described in detail below with reference to the accompanying drawings. In the present embodiment, the present invention will be described as being applied to an object detection device that detects an object from an image captured by a fixed camera.

[0021] <Outline of the embodiment of the present invention> In an embodiment of the present invention, specific objects (e.g., vehicles or pedestrians) are detected from images captured by a fixed camera (e.g., a surveillance camera installed on the side of the road) using object detection technology (neural networks such as deep learning, hereinafter referred to as DNN).

[0022] Furthermore, in the embodiment of the present invention, the image captured by the fixed camera is adjusted so that the output of the detection unit suppresses false detection from an image (background image) that does not contain the detection target, and the detection results are classified into removal candidates to be removed and retention candidates to be maintained based on simple rules. Furthermore, the image adjustment and classification of the detection results are characterized by using parameters specific to each fixed camera.

[0023] In addition, the advantage of rule-based classification of detection results is that it does not generate unnecessary false positives, but the disadvantage is that it may remove true positives other than false positives.Also, adjusting the input image using adversarial noise generated to suppress false positives has the advantage of avoiding the removal of true positives, which is a drawback of rule-based classification, but the disadvantage is that it may generate new false positives.

[0024] Therefore, in the embodiment of the present invention, by integrating the results of both, it is possible to selectively suppress erroneous detection while suppressing unintended changes in the detection results that accompany image adjustment.

[0025] <Configuration of the object detection device according to the embodiment of the present invention> As shown in FIG. 1, an object detection device 100 according to an embodiment of the present invention includes a fixedly installed imaging device 10 and a computer 20 that executes an object detection processing routine that uses an object detection technology (e.g., DNN) to detect objects (e.g., vehicles or pedestrians) from captured images output from the imaging device 10 and monitors and tracks the objects.

[0026] The imaging device 10 is, for example, a surveillance camera fixedly installed on the side of the road.

[0027] The computer 20 is configured to include a CPU that controls the entire object detection device 100, a ROM as a storage medium that stores programs such as an erroneously detected image collection processing routine, an adjustment parameter calculation processing routine, and an object detection processing routine, which will be described later, a RAM that temporarily stores data as a work area, and a bus that connects these. In such a configuration, a program for realizing the function of each component is stored in a storage medium such as a ROM or HDD, and each function is realized by the CPU executing the program.

[0028] If this computer 20 is explained in terms of functional blocks divided into function realization means determined based on hardware and software, it can be represented as a configuration including an image acquisition unit 22, an image adjustment unit 24, a detection unit 26, a collection unit 28, an image database 30, a parameter calculation unit 32, a classification unit 34, an integration unit 36, and an output unit 38, as shown in Figure 1.

[0029] The image acquisition unit 22 acquires the captured image captured by the imaging device 10 and input to the computer 20 .

[0030] The image adjustment unit 24 adjusts the pixel value of each pixel of the captured image to generate an adjusted image.

[0031] In the present embodiment, the image adjuster 24 adjusts the captured image by adding noise to each pixel, for example.

[0032] Specifically, the image captured by the imaging device 10 is represented by x, and the image after adjustment by the image adjustment means g is represented by ∨ x is calculated using the adjustment parameter φ.

[0033] JPEG0007753919000001.jpg2079 (1)

[0034] In addition, in the formula JPEG0007753919000002.jpg97 In the specification, ∨ It is expressed as x.

[0035] For example, when φ=0, JPEG0007753919000003.jpg1364 (2) is.

[0036] An example of g() is shown in equation (3).

[0037] JPEG0007753919000004.jpg1681 (3)

[0038] If the captured image x is a color image, it can be expressed as a third-order tensor, and the shape of the tensor is width × height × number of channels. a and b are tensors with the same shape as x, and the operator JPEG0007753919000005.jpg1213 indicates element-wise multiplication. In particular, when all elements of a are 1, g() is an image adjustment that adds adversarial noise.

[0039] An example of different g() is shown in equation (4).

[0040] JPEG0007753919000006.jpg17102 (4) JPEG0007753919000007.jpg1548

[0041] The operator * denotes a convolution operation, and A is a filter that performs the convolution operation, such as a convolution layer with a kernel size of 1x1 or 3x3 used in convolutional neural networks. h is a kernel function that performs a linear or nonlinear transformation.

[0042] The detection unit 26 detects the position and size of the object from each of the captured image after adjustment and the captured image before adjustment.

[0043] Specifically, the adjusted captured image ∨ The results of object detection by the object recognition means f for the image x and the captured image x before adjustment are denoted as y1 and y2.

[0044] JPEG0007753919000008.jpg3168 (5)

[0045] The object recognition means f is composed of a DNN. The object detection result y includes position information of the detection frames of multiple detected objects and predicted class probabilities.

[0046] The collecting unit 28 collects erroneously detected images, which are captured images for which the detection result is incorrect, based on the detection result by the detecting unit 26 and the correct detection result given to the captured image.

[0047] Specifically, the collection unit 28 collects images x that do not include the object to be detected in the captured image but in which the object is detected (falsely detected image). FP For example, an erroneously detected image that falls under any of the following cases 1 to 4 is identified.

[0048] Case 1: When a monitor visually checks the captured image in which an object was detected and finds that it was a false detection.

[0049] Case 2: An object is detected from an image captured in a situation where it is guaranteed that the object is not present (for example, when the imaging target area is blocked off).

[0050] Case 3: An object is detected in a predetermined background area on the captured image (e.g., an area outside the road) where it is known that no objects exist.

[0051] Case 4: When an object is detected from the captured image even though it cannot be detected using other sensors such as LIDAR (Light Detection and Ranging) or radar.

[0052] The image database 30 contains false positive images x as shown in FIG. FP Here, the set of false positive images is stored as X FP Figure 2 shows an example of an image in which a container temporarily placed in the background area (area outside the road) is mistakenly detected as a vehicle. As shown here, false detections can occur when an object is temporarily placed or the environment changes due to road construction, etc.

[0053] The parameter calculation unit 32 calculates adjustment parameters based on the detection results by the detection unit 26 for the erroneously detected images stored in the image database 30 so that correct detection results can be obtained for the erroneously detected images.

[0054] Specifically, the parameter calculation unit 32 calculates the erroneously detected image group X FP The adjustment parameters are calculated using

[0055] Since it is desirable that the image adjustment be small, for example, the adjustment parameters are optimized under the following constraints.

[0056] JPEG0007753919000009.jpg1698 JPEG0007753919000010.jpg1392 (6)

[0057] where p CAR is the set of false positive images X FP The class probability y of the object detection result of the image x included in CAR For simplicity, if one object is detected from one image x, then p CAR is calculated using the following formula:

[0058] JPEG0007753919000011.jpg2370 (7)

[0059] Note that adversarial noise that suppresses false positives may be calculated by defining a loss function for false positive images and using the backpropagation algorithm. The adjustment parameters can be determined by determining the noise intensity so as to satisfy the constraints. The adversarial noise calculation method is not limited to this example.

[0060] JPEG0007753919000012.jpg2490 JPEG0007753919000013.jpg2069 JPEG0007753919000014.jpg1163 JPEG0007753919000015.jpg1388

[0061] [Reference] T Kutsuna, S Koide, K Kawano, “Discovering Potential False Positives in Object Detection using Adversarial Attacks”, AAAI 2019 workshop on Engineering Dependable and Secure Machine Learning Systems

[0062] When the captured image is adjusted using the adjustment parameters optimized by the parameter calculation unit 32, the detection unit 26 produces detection results such as those shown in Figures 3(A), (B), and (C). Figure 3(A) shows an example in which false detections from background areas are suppressed. Figure 3(B) shows an example in which an object existing near the background area is correctly detected. Figure 3(C) shows an example in which a false detection frame that did not originally exist has newly appeared due to the conditions of the image adjustment unit g and the arrangement of surrounding objects.

[0063] In addition, the parameter calculation unit 32 calculates classification parameters based on the detection results from the captured images before adjustment by the detection unit 26 for the falsely detected images stored in the image database 30 so that the classification means can obtain correct classification results for the detection results.

[0064] Specifically, the parameter calculation unit 32 calculates the erroneously detected image group X FP The classification parameters are calculated using

[0065] For example, when the classification means c uses overlap of detection frames as a judgment condition, the false detection image group X FP The detection frame of the false positive that occurred in one image randomly selected from the set of false positive images X is selected as the detection frame to be used for the overlap calculation. FP For all other images in the image, the maximum threshold value that can classify the false positive detection window as a removal candidate can be determined. Alternatively, the threshold value can be determined so as to maximize the classification performance in a separately prepared detection dataset. The method for determining the classification parameters is not limited to this example.

[0066] The classification unit 34 classifies the detection results from the adjusted captured image into removal candidates and retention candidates based on the detected areas, and classifies the detection results from the unadjusted captured image into removal candidates and retention candidates based on the detected areas.

[0067] In this embodiment, the detection result of the object from the background area of ​​the fixed camera image is targeted for suppression. Therefore, the classification unit 34 classifies the detection result y1 from the adjusted captured image into removal candidates y based on whether the detected area is a background area or not, using the classification means c. rm and the candidate for maintenance y kp and using a classification means c, the detection result y2 from the captured image before adjustment is classified into removal candidates y based on whether the detected region is a background region or not. rm and the candidate for maintenance y kp They are classified as follows.

[0068] y 1,rm ,y 1,kp =c(y1) y 2,rm ,y 2,kp =c(y2)

[0069] More specifically, the classification means c may output a detection frame that is in the same class as the object and has an overlap with the background region that is equal to or greater than a threshold as a removal candidate, and the rest as a retention candidate.

[0070] Here, the classification means c is preset with classification parameters (e.g., thresholds) calculated using the detection results of the captured image before adjustment. The classification means c is not limited to this example. For example, a region in the image (e.g., a background region) may be specified, and detection frames with an overlap of a threshold or more may be set as candidates for removal.

[0071] The integration unit 36 ​​integrates the detection results from the adjusted captured image and the detection results from the unadjusted captured image based on the classification results of the detection results from the adjusted captured image and the classification results of the detection results from the unadjusted captured image.

[0072] Here, taking into consideration the following two points: (1) the detection result y1 from the adjusted captured image has already suppressed erroneous detection of the object, and (2) the classification means c classifies the detection result that is likely to be a erroneous detection of the object as a removal candidate, the removal candidate y1 of the classification result for the detection result of the adjusted captured image is 1,rmis considered to be a positive detection result that is erroneously removed by the classification means c. Furthermore, the detection result y2 from the captured image before adjustment is guaranteed to be the same as the detection result from the captured image after adjustment in areas that are not related to the false detection to be suppressed (areas that are not background areas), and the effect of selectively suppressing false detection by image adjustment can be obtained.

[0073] Therefore, in this embodiment, the integration unit 36 ​​obtains the detection result of the object by integrating the removal candidates from the detection results from the captured image after adjustment and the retention candidates from the detection results from the captured image before adjustment according to the following formula.

[0074] y=y 1,rm +y 2,kp

[0075] For example, the detection results from the captured image after adjustment by the detection unit 26 are as shown in Figures 3(A), (B), and (C). The detection frame in Figure 3(B) is classified as a removal candidate.

[0076] 4(A), (B), and (C) show candidates to be retained from the detection results from the captured image before adjustment, corresponding to the detection results shown in FIGS. 3(A), (B), and (C) above. FIG. 4(A) shows an example in which false detections from the background region are suppressed. FIG. 4(C) shows an example in which no new false detections occur. FIG. 4(B) shows an example in which a true detection is erroneously removed when the object overlaps with a false detection in the background region.

[0077] In addition, Figures 5(A), (B), and (C) show the results of integrating the removal candidates from the detection results shown in Figures 3(A), (B), and (C) above with the retention candidates from the detection results from the captured images before adjustment shown in Figures 4(A), (B), and (C) above.

[0078] FIG. 5(C) shows an example in which new false detections caused by the adjustment of the captured image are suppressed. FIG. 5(B) shows an example in which the detection frame that is a candidate for removal shown in FIG. 3(B) is maintained. In this way, false detections from background areas can be suppressed by taking advantage of the advantages of both the adjustment by the image adjustment means g and the classification by the classification means c. Furthermore, because the adjustment parameters are specific to each imaging device 10, they do not affect the detection performance of other imaging devices 10.

[0079] When the integrated detection result indicates that an object has been detected, the output unit 38 transmits the integrated detection result to the traffic control system 150 (see FIG. 6). FIG. 6 shows an example in which adjustment parameters and classification parameters are calculated in multiple object detection devices 100, each having an imaging device 10, and when an object is detected, the integrated detection result is transmitted to the traffic control system 150.

[0080] <Operation of the object detection device according to the embodiment of the present invention> Next, a description will be given of an erroneous detection image collection processing routine executed by the computer 20 of the object detection device 100 according to the embodiment of the present invention with reference to Fig. 7. For example, the erroneous detection image collection processing routine is executed in a situation where it is possible to guarantee that no object is present.

[0081] In step S100, a captured image captured by the imaging device 10 is acquired.

[0082] Then, in step S102, the image adjuster 24 adjusts the pixel value of each pixel of the captured image using the current adjustment parameters to generate an adjusted image.

[0083] Next, in step S104, the detection unit 26 detects an object from the adjusted image generated in step S102.

[0084] In step S106, the collection unit 28 determines whether or not an object has been detected from the adjusted image based on the detection result in step S104. If an object has not been detected from the adjusted image, the collection unit 28 determines that the detection result is correct and ends the erroneously detected image collection processing routine.

[0085] On the other hand, if the object is detected from the adjusted image, it is determined that the detection result is incorrect, and the process proceeds to step S108.

[0086] In step S108, the captured images acquired in step S100 are collected as erroneously detected images, which are captured images with erroneous detection results, and stored in the image database 30, and the erroneously detected image collection processing routine is terminated.

[0087] 8, an adjustment parameter calculation processing routine executed by the computer 20 of the object detection device 100 according to the embodiment of the present invention will be described. For example, the adjustment parameter calculation processing routine is executed every time a predetermined number of erroneously detected images are accumulated.

[0088] First, in step S110, the erroneously detected images stored in the image database 30 are acquired.

[0089] In step S112, the parameter calculation unit 32 calculates adjustment parameters so that correct detection results can be obtained for erroneously detected images stored in the image database 30, based on the detection results from the captured images after adjustment by the detection unit 26. Furthermore, the parameter calculation unit 32 calculates classification parameters so that correct classification results can be obtained for the detection results by the classification means, based on the detection results from the captured images before adjustment by the detection unit 26, for erroneously detected images stored in the image database 30.

[0090] In step S114, the adjustment parameters and classification parameters calculated in step S112 are updated as the current adjustment parameters and classification parameters, and the adjustment parameter calculation processing routine is terminated.

[0091] The erroneously detected image collection processing routine and the adjustment parameter calculation processing routine are executed in each of the plurality of object detection devices 100 each having an image capture device 10.

[0092] 9, an object detection processing routine executed by the computer 20 of the object detection device 100 according to the embodiment of the present invention will be described. For example, in each of a plurality of object detection devices 100 each having an image capture device 10, the object detection processing routine is executed every time an image is captured by the image capture device 10.

[0093] First, in step S120, the image acquisition unit 22 acquires the captured image captured by the imaging device .

[0094] Then, in step S122, the image adjuster 24 adjusts the pixel value of each pixel of the captured image using the current adjustment parameters to generate an adjusted image.

[0095] Then, in step S124, the detection unit 26 detects an object from each of the adjusted image generated in step S102 and the captured image before adjustment.

[0096] In step S126, the classification unit 34 classifies the detection results from the adjusted image into removal candidates and retention candidates based on the detected areas, and classifies the detection results from the captured image before adjustment into removal candidates and retention candidates based on the detected areas.

[0097] In step S128, the integrating unit 36 ​​integrates the detection results from the adjusted image and the detection results from the unadjusted captured image based on the classification results of the detection results from the adjusted image and the classification results of the detection results from the unadjusted captured image. If an object is detected as a result of the integration, the output unit 38 transmits the integrated detection results to the traffic control system 150.

[0098] As described above, according to the object detection device of an embodiment of the present invention, a captured image is adjusted, objects are detected from the adjusted image and the captured image before adjustment, the detection results from the adjusted image are classified into candidates for removal and candidates for retention, the detection results from the captured image before adjustment are classified into candidates for removal and candidates for retention, and based on the classification results, the detection results from the adjusted image and the detection results from the captured image before adjustment are integrated, thereby suppressing erroneous detection of objects without degrading object detection performance.

[0099] In addition, by integrating the detection results before and after adjustment of the captured image based on the classification results into removal candidates or retention candidates, it is possible to selectively suppress erroneous detections due to adjustment of the captured image while suppressing unintended changes in detection results due to adjustment of the captured image.

[0100] Furthermore, adversarial noise, which is adjusted to suppress false detections in background images, is most effective against false detections arising from background images, but is less effective in suppressing detections in images in which an object is present.

[0101] In the above embodiment, the case where only falsely detected images are collected has been described as an example, but the present invention is not limited to this. In addition to falsely detected images, correctly detected images including the target may also be collected. Specifically, the collection unit 28 collects falsely detected images, which are captured images with incorrect detection results, and correctly detected images, which are captured images including the target, based on the detection results by the detection unit 26 and the correct detection results given to the captured images. The image database 30 stores the falsely detected images x FP and positive detection image x TP The parameter calculation unit 32 calculates adjustment parameters based on the detection results by the detection unit 26 for the erroneously detected images stored in the image database 30 and the detection results by the detection unit 26 for the correctly detected images stored in the image database 30 so that correct detection results are obtained for the erroneously detected images and the correctly detected images.

[0102] Furthermore, adjustment parameters may be prepared for each image attribute at the time of image capture. Specifically, the image acquisition unit 22 acquires captured images captured by the imaging device 10 along with the image attributes at the time of image capture. Here, the image attributes may be expressed by any one or a combination of information obtained from the camera, such as camera gain, exposure time, aperture, and capture time. Furthermore, information about the imaging environment obtained from a sensor other than the camera may be used as the image attributes. For example, any one or a combination of information about weather, such as time and precipitation, the on / off status of lighting, and road traffic density calculated from the frequency of vehicle detection may be used. The image adjustment unit 24 selects adjustment parameters corresponding to the image attributes acquired by the image acquisition unit 22 and uses the selected adjustment parameters to adjust the pixel values ​​of each pixel of the captured image to generate an adjusted image. The collection unit 28 collects erroneously detected images, which are captured images with incorrect detection results, based on the detection results by the detection unit 26 and the correct detection results assigned to the captured images, and stores the erroneously detected images in the image database 30 along with the image attributes of the captured images. Falsely detected images are stored for each image attribute in image database 30. For each image attribute, parameter calculation unit 32 calculates an adjustment parameter for the image attribute based on the detection result by detection unit 26 for the falsely detected image of that image attribute stored in image database 30 so that a correct detection result can be obtained for the falsely detected image.

[0103] Furthermore, while the present invention has been described with reference to an example in which an image captured by a fixed camera is used as the target, the present invention is not limited to this. Sensor data consisting of elements acquired by a sensor whose relative positional relationship with the target is constant can be adjusted to generate adjusted data. For example, the present invention may be applied to an object detection device that captures images of products with a camera and detects defective parts of the products as targets. Images captured by the camera are adjusted using adjustment parameters and processed by an inspection algorithm configured with DNN to determine whether the product is good or bad and identify defective parts.

[0104] In the above embodiment, the target is an image in the narrow sense, which is two-dimensional (or three-dimensional in the case of a color image) information captured by an imaging device. However, the present invention is not limited to this, and the target sensor data may be an image of information acquired by a sensor. For example, the target sensor data may be a range image generated from a point cloud measured by a LIDAR or a spectral image obtained by performing a discrete Fourier transform on acoustic data.

[0105] Furthermore, although the example has been described in which the position and size of an object are identified by detecting the object, this is not limited to this, and the detection of an object may include object recognition that identifies the position and attributes of the object, and image identification that determines the attributes of the image itself. [Explanation of symbols]

[0106] 10. Imaging device 20 Computer 22 Image acquisition unit 24 Image adjustment section 26 Detector 28 Collection Department 30 Image Database 32 Parameter calculation section 34 Classification Department 36 Integration Department 38 Output section 100 Object detection device 150 Traffic Control System

Claims

1. A data adjustment unit that adjusts pixel values ​​of an image acquired by an imaging device whose installation position is fixed by adding noise to each pixel, and generates an adjusted image by adjusting the image; a detection unit that detects an object from the adjusted image and detects the object from the image before adjustment; classifying the detection results from the adjusted image into removal candidates for which the detection results should be removed and retention candidates for which the detection results should be maintained, such that detection frames that are of the same class as the object and have an overlap with the background region equal to or greater than a threshold are designated as removal candidates, and the rest are designated as retention candidates; a classification unit that classifies the detection results from the pre-adjustment image into removal candidates and retention candidates, such that a detection frame that is in the same class as the object and has an overlap with a background region equal to or greater than a threshold is set as a removal candidate, and the rest are set as retention candidates; an integration unit that obtains a detection result of the object by integrating the removal candidates from the detection results from the adjusted image and the retention candidates from the detection results from the unadjusted image based on a classification result of the detection results from the adjusted image and a classification result of the detection results from the unadjusted image; An object detection device comprising:

2. a collection unit that collects false-detection images, which are images for which the detection result by the detection unit is incorrect, based on whether the area is detected as a background area; 2. The object detection device according to claim 1, further comprising: a parameter calculation unit that calculates noise to be added to each pixel based on the detection result by the detection unit for the collected false-detection image so as to obtain a correct detection result for the false-detection image.

3. The object detection device according to claim 2, wherein the parameter calculation unit further calculates classification parameters based on the detection results from the images before adjustment by the detection unit for the collected falsely detected images so that the detection results for the falsely detected images are classified as candidates for removal.

4. Computer, a data adjustment unit that adjusts pixel values ​​of an image acquired by an imaging device whose installation position is fixed by adding noise to each pixel, and generates an adjusted image by adjusting the image; a detection unit that detects an object from the adjusted image and detects the object from the image before adjustment; classifying the detection results from the adjusted image into removal candidates for which the detection results should be removed and retention candidates for which the detection results should be maintained, such that detection frames that are of the same class as the object and have an overlap with the background region equal to or greater than a threshold are designated as removal candidates, and the rest are designated as retention candidates; a classification unit that classifies the detection results from the image before adjustment into removal candidates and retention candidates such that a detection frame that is in the same class as the object and has an overlap with a background region equal to or greater than a threshold is set as a removal candidate, and the rest are set as retention candidates; and an integration unit that obtains a detection result of the object by integrating the removal candidates among the detection results from the adjusted image and the retention candidates among the detection results from the unadjusted image based on a classification result of the detection results from the adjusted image and a classification result of the detection results from the unadjusted image. A program to function as a

Images