KR102602400B1 - Method and Apparatus for Box Level Postprocessing For Accurate Object Detection - Google Patents

Abstract

A box-level post-processing method and device for accurate object detection are presented. The box-level post-processing method for accurate object detection proposed in the present invention includes the steps of acquiring images by photographing a plurality of objects through a camera, and obtaining boxes for detecting a plurality of objects from the acquired images through an object detection unit. , A step of correcting the reliability value of the fake object box and the reliability value of the real object box due to overlap between multiple objects among the boxes for multiple object detection using the foreground pixel information in the box through the correction unit, and the corrected reliability value. It includes a step of readjusting the positions of the boxes for detecting a plurality of objects through the object detection unit.

Description

Box level postprocessing method and apparatus for accurate object detection {Method and Apparatus for Box Level Postprocessing For Accurate Object Detection}

The present invention relates to an image processing method and device that uses convolutional neural network-based object detection for accurate object detection.

Considering the lack of workers in pig farms (in Korea, one worker manages an average of 2,000 pigs) and the high mortality rate of pigs (in Korea, about 5 million pigs die annually), detailed management of individual pigs is necessary. For this reason, the need for monitoring pig farms using information technology is increasing [1]. However, even with the application of continuously evolving convolutional neural network-based object detection technology [2], there are limitations in accurately detecting pigs in crowded pig houses due to reasons such as occlusion between pigs.

With the development of convolutional neural network technology, pig monitoring in pig farms is possible through object detection. An object detection method for monitoring pigs in a pig pen according to the prior art is the YOLOv4 object detection method, which obtains bounding boxes corresponding to the detected object from images obtained from a camera. YOLOv4, a cost-effective object detector with an appropriate combination of time and accuracy, uses a reliability value that indicates how accurately the currently detected box information captures the class and object size from the object information that comes out after going through Soft-nms. In addition, to separate objects used in the method of readjusting the reliability value, the adaptiveThreshold technique is used to separate objects in the image and rotate the image. YOLOv4 achieves the highest accuracy relative to processing speed with MS COCO [4], a representative object detection public database.

With the continuous development of convolutional neural network-based deep learning technology, YOLOv4 accurately detects most objects (confidence value of boxes is over 80%), but fake object boxes are created due to the overlap of objects in a complex and crowded structure due to multiple objects. A case may occur where has a higher confidence value than the real object box.

Therefore, in order to solve this problem, a correction method is needed that lowers the reliability value of fake objects using foreground pixel information within a box and increases the reliability value of overlapping objects using information in adjacent boxes.

The technical problem that the present invention aims to achieve is object detection using convolutional neural network technology. In order to improve the accuracy of object detection, the box level post-processing method adjusts the reliability values of incorrect boxes among the boxes that are output from the detector. The purpose is to provide processing methods and devices.

In one aspect, the box-level post-processing method for accurate object detection proposed in the present invention includes the steps of acquiring images by photographing a plurality of objects through a camera, and detecting a plurality of objects through an object detector from the acquired images. Obtaining boxes, correcting the reliability value of the fake object box and the reliability value of the real object box due to overlap between a plurality of objects among the boxes for detecting a plurality of objects through a correction unit using foreground pixel information in the box. and readjusting the positions of boxes for detecting a plurality of objects through an object detection unit according to the corrected reliability value.

The step of correcting the reliability value of the fake object box and the reliability value of the real object box due to overlap between multiple objects among the boxes for multiple object detection through the correction unit using the foreground pixel information in the box is to determine whether the fake object box is real or not. In order to determine whether it is an object box, the degree of foreground pixels included in the box is calculated. If the foreground pixels included in the box are less than a predetermined standard, it is judged to be a fake object box, and the foreground pixels included in the box are determined to be fake object boxes. If it exceeds a predetermined standard, it is judged to be a real object box, and the foreground and background pixels within the box are separated using an adaptive threshold method for the real object box.

According to an embodiment of the present invention, the reliability value of the fake object box and the reliability value of the real object box are calculated according to the separated foreground pixels and background pixels using the ratio of the size of the foreground area within the box to the total size of the box. Calibrate the reliability value.

According to an embodiment of the present invention, when the reliability value of the real object box is calculated to be low due to overlapping areas among the real object boxes, the reliability value is calculated using the ratio of the size of the non-overlapping area within the box to the total size of the box. Correct.

According to an embodiment of the present invention, in order to correct the reliability value of the overlapping area among the real object boxes, the box related to the object detected in the original image is matched with the box related to the object detected in the rotation image obtained by rotating the original image. , the adaptive threshold method is used for the original image and the rotated image to separate the foreground and background pixels within the box, and then the reliability value for the box containing more foreground pixels among the original image and the rotated image is corrected.

In another aspect, the box-level post-processing device for accurate object detection proposed in the present invention includes an image collection unit that acquires images of a plurality of objects captured through a camera, and detection of a plurality of objects from the acquired images. An object detection unit that acquires boxes for detecting multiple objects and readjusts the positions of the boxes for multiple object detection according to the reliability value corrected through the correction unit, and a fake object box due to overlap between multiple objects among the boxes for multiple object detection. It includes a correction unit that corrects the reliability value of and the reliability value of the real object box using foreground pixel information in the box.

According to embodiments of the present invention, in order to solve the case where a fake object box has a higher reliability value than a real object box due to overlap of objects in a complex and congested structure due to a plurality of objects, foreground pixel information within the box is used. The accuracy of object detection can be improved by proposing a correction method that lowers the reliability value of fake objects and increases the reliability value of overlapping objects using information from adjacent boxes.

1 is a flowchart illustrating a box-level post-processing method for accurate object detection according to an embodiment of the present invention.
Figure 2 is a diagram showing a reliability correction algorithm according to an embodiment of the present invention.
Figure 3 is a flowchart illustrating a background separation process for a plurality of objects according to an embodiment of the present invention.
Figure 4 is a diagram showing a camera installed for image acquisition according to an embodiment of the present invention.
Figure 5 is a diagram showing the process of calculating the size (S _{foreground_pixel} ) of the foreground area of a box to separate the object and the background according to an embodiment of the present invention.
FIG. 6 is a diagram illustrating a reliability value correction process for lowering the reliability value of a fake object box according to an embodiment of the present invention.
FIG. 7 is a diagram illustrating a reliability value correction process for increasing the reliability value of a real object box according to an embodiment of the present invention.
Figure 8 is a diagram illustrating the process of matching after applying the reliability correction algorithm to the original image and the rotated image, respectively, according to an embodiment of the present invention.
Figure 9 is a diagram for explaining the process of separating objects in an image using an adaptiveThreshold technique according to an embodiment of the present invention.
FIG. 10 is a diagram illustrating a box-level post-processing device for accurate object detection according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

1 is a flowchart illustrating a box-level post-processing method for accurate object detection according to an embodiment of the present invention.

The proposed box-level post-processing method for accurate object detection includes the steps of acquiring images by photographing a plurality of objects using a camera (110), and acquiring boxes for detecting a plurality of objects from the acquired images through an object detection unit. (120), correcting the reliability value of the fake object box and the reliability value of the real object box due to overlap between a plurality of objects among the boxes for detecting a plurality of objects through the correction unit using the foreground pixel information in the box (130) ) and a step 140 of readjusting the positions of the boxes for detecting a plurality of objects through the object detection unit according to the corrected reliability value.

In step 110, images are obtained by photographing a plurality of objects using a camera. For example, in order to photograph multiple objects through a camera, photographing can be done through a top-view camera that photographs from above and below.

In step 120, boxes for detecting a plurality of objects are obtained from the acquired image through an object detection unit. Each reliability value can be obtained by operating a detector for images acquired from installed cameras.

In step 130, the reliability value of the fake object box and the reliability value of the real object box due to overlap between a plurality of objects among the boxes for detecting a plurality of objects are corrected through the correction unit using foreground pixel information in the box.

In step 130, the degree of inclusion of foreground pixels in the box is first calculated to determine whether it is a fake object box or a real object box. If the foreground pixels included in the box are less than a predetermined standard, it is judged as a fake object box, and if the foreground pixels included in the box are more than the predetermined standard, it is judged as a real object box. Afterwards, an adaptive threshold method is used for the real object box to separate the foreground and background pixels within the box.

According to the separated foreground pixels and background pixels, the reliability values of the fake object box and the real object box are corrected using the ratio of the size of the foreground area within the box to the total size of the box.

Additionally, if the reliability value of the real object box is calculated to be low due to overlapping areas among the real object boxes, the reliability value is corrected using the ratio of the size of the non-overlapping area within the box to the total size of the box.

To correct the reliability value for the overlapping area among the real object boxes, the box for the object detected in the original image is matched with the box for the object detected in the rotated image that rotated the original image. Afterwards, the adaptive threshold method is used for the original image and the rotated image to separate the foreground and background pixels within the box, and then the reliability value for the box containing more foreground pixels among the original image and the rotated image is corrected. .

In step 140, the positions of boxes for detecting a plurality of objects are readjusted through the object detection unit according to the corrected reliability value.

Hereinafter, a box-level post-processing method for accurate object detection according to an embodiment of the present invention will be described in more detail with reference to FIGS. 2 to 9. For detailed explanation of the object detection method and device proposed in the present invention, it will be described as an example by applying it to object detection (that is, pigs) in a pig pen. Object detection within a pig pen is only an example and is not limited to this, and the object detection method and device proposed in the present invention can be applied to object detection in various fields.

Figure 2 is a diagram showing a reliability correction algorithm according to an embodiment of the present invention.

For detailed explanation of the object detection method and device proposed in the present invention, it will be described as an example by applying it to object detection (that is, pigs) in a pig pen. First, images are obtained by photographing a plurality of objects using a camera. The camera according to an embodiment of the present invention may be installed as a top-view camera that shoots from above to below to capture images of pigs in a pig pen.

Boxes for detecting multiple pigs are obtained through an object detection unit from images of pigs in the pig pen acquired through a camera. The reliability value for each box can be obtained by running a detector on the image acquired from the installed camera. The reliability value for each box can be corrected using the algorithm shown in FIG. 2 for the calculated reliability value. At this time, through the correction unit, the reliability value of the fake object box and the reliability value of the real object box due to overlap between the plurality of pigs among the boxes for detecting the plurality of pigs are corrected using the foreground pixel information in the box.

The object detection unit according to an embodiment of the present invention used YOLOv4, a cost-effective object detector with an appropriate combination of time and accuracy. In YOLOv4, the confidence value that indicates how accurately the class and object size are captured in the currently detected box information from the object information that comes out after going through Soft-nms is used. In addition, to separate objects used in the method of readjusting the reliability value, an adaptive threshold technique is used to separate objects in the image and rotate the image to correct the reliability value.

According to an embodiment of the present invention, the reliability value of the fake object box and the reliability value of the real object box are calculated according to the separated foreground pixels and background pixels using the ratio of the size of the foreground area within the box to the total size of the box. The reliability value can be corrected.

In the case of a fake object box, you can use the Decrease_Confidence( b ) function and Opencv to separate the object from the background and obtain the size of the foreground area (S _{foreground_pixel} ) within the box.

On the other hand, in the case of a real object box, you can use the Increase_Confidence( b ) function to obtain the size of the non-overlapped area (S _{not_overlapped} ) within the box, excluding the overlapped box.

Afterwards, the reliability correction algorithm is applied to the original image and the rotated image respectively, and then the Euclidean distance is calculated for the result detected in the original image and the result detected in the rotated image, and the detection is performed in the original image. You can match the box about the detected object with the box about the object detected in the rotated image that rotated the original image.

After separating the foreground and background pixels within the box using an adaptive threshold method for the original image and the rotated image, a detection result including a large amount of background is obtained from the detected result. Afterwards, the reliability value of the object containing more background among the original image and the rotated image is corrected. After correcting the reliability value, the box position is readjusted according to the reliability value.

Figure 3 is a flowchart illustrating a background separation process for a plurality of objects according to an embodiment of the present invention.

For detailed explanation of the object detection method and device proposed in the present invention, it will be described as an example by applying it to object detection (that is, pigs) in a pig pen.

The present invention proposes a box-level post-processing method that evaluates the reliability values of boxes that are the output of a convolutional neural network-based pig detector and corrects the reliability values of incorrect boxes. First, images are obtained by photographing a plurality of objects using a camera.

Boxes for detecting multiple pigs are obtained through an object detection unit from images of pigs in the pig pen acquired through a camera. By operating a detector on images acquired from installed cameras, the confidence score for each box can be obtained.

The object detection unit according to an embodiment of the present invention used YOLOv4, a cost-effective object detector with an appropriate combination of time and accuracy. In YOLOv4, the confidence value that indicates how accurately the class and object size are captured in the currently detected box information from the object information that comes out after going through Soft-nms is used.

The reliability threshold (T) can be specified using the reliability value obtained after going through Soft-nms in YOLOv4, and the reliability value can be adjusted according to predetermined conditions as follows.

For example, if the calculated reliability value of the box is < T-10, the box can be deleted and the next object (in other words, a pig) can be searched.

If T+10 < the calculated reliability value of the box, the next object can be searched after maintaining the box.

If T < the calculated confidence value of the box < T+10, the confidence value of the box can be decreased by calling the Decrease_Confidence() function.

If T-10 < the calculated confidence value of the box < T, the confidence value for the box can be increased by calling the Increase_Confidence() function.

If the calculated reliability value of the box is < T, the next object can be searched after deleting the box.

If T < the calculated reliability value of the box, the next object can be searched after maintaining the box.

The above process can be repeated for all object boxes in the class. After repeating the process for all object boxes in the class, the captured image can be rotated and the above process can be repeated again for all object boxes in the class.

After completing the iteration for all object boxes in the class for the rotated image, the box for the object detected in the original image and the original image were rotated to correct the reliability value for the overlapping area among the real object boxes. Match n:m boxes related to objects detected in the image. Afterwards, the adaptive threshold method is used for the original image and the rotated image to separate the foreground and background pixels within the box, and then the reliability value for the box containing more foreground pixels among the original image and the rotated image is corrected. .

For example, among the matching results, the detection result containing the most foreground pixels can be selected and the confidence value can be readjusted by calling the Rotation_Confidence() function for the selected detection result. Among the matching results, the matching information with the object with the highest reliability value can be used to reflect this in the final detection result.

Figure 4 is a diagram showing a camera installed for image acquisition according to an embodiment of the present invention.

The camera 410 according to an embodiment of the present invention may be installed as a top-view camera that captures images from above and below to capture images of pigs in a pig pen.

The image acquired from the camera 410 in the pig pen can be applied to the YOLOv4 object detection unit according to an embodiment of the present invention to obtain bounding boxes corresponding to the detected object. YOLOv4 can achieve high accuracy compared to processing speed with MS COCO [4], a representative object detection public database. In other words, with the continuous development of convolutional neural network-based deep learning technology, YOLOv4 accurately detects most pigs, but due to the complex structure of the pig pen, there is an issue that boxes that detect fake pigs in the background are included (False Positive). , FP) remain. If, in order to remove these fake pigs, you simply remove boxes with low confidence values, another problem (False Negative, FN) arises in that other boxes with low confidence values, such as occluded pigs, are also removed. Occurs. In order to solve these problems, in the present invention, when it is difficult to determine whether a pig is a fake pig or a real pig based on the reliability value of the current box, the reliability value is corrected using the foreground pixel information within the box and the information of adjacent boxes. Perform.

Figure 5 is a diagram showing the process of calculating the size (S _{foreground_pixel} ) of the foreground area of a box to separate the object and the background according to an embodiment of the present invention.

First, in order to determine whether a box with an ambiguous confidence value is a fake pig box corresponding to the background, it is calculated how many foreground pixels are included in the box. Although the ratio of foreground pixels in a box varies depending on the pig's location, if it is a box corresponding to a fake pig, pixels below a predetermined standard are assumed to be foreground pixels. To distinguish pixels within a box into foreground and background, an adaptive threshold method [5] is applied, and the reliability value of the box determined based on this standard is compared to the total size of the box (S _total ) (510). Call the Decrease_Confidence() function using the size of the foreground area of the box (S _{foreground_pixel} ) (520) and correct it with the following formula:

FIG. 6 is a diagram illustrating a reliability value correction process for lowering the reliability value of a fake object box according to an embodiment of the present invention.

For the fake pig box using the YOLOv4 object detection unit according to an embodiment of the present invention, the confidence value is calculated by calling the Decrease_Confidence() function that uses the size of the foreground area (S _{foreground_pixel} ) of the box compared to the total size of the box (S _total ). Correct.

As shown in Figure 5, the reliability value of the fake pig box was 75%. It can be confirmed that the confidence value corrected by calling the Decrease_Confidence() function for the confidence value of this fake pig box is corrected to 63%, as shown in FIG. 6.

FIG. 7 is a diagram illustrating a reliability value correction process for increasing the reliability value of a real object box according to an embodiment of the present invention.

According to an embodiment of the present invention, the reliability value may be corrected using adjacent box information. To determine whether a box with an ambiguous reliability value is an overlapping pig box corresponding to the foreground, first determine whether there is a box overlapping with the corresponding box. In this case, the ratio of foreground pixels within the box will be more than half, and the correction process for lowering the confidence value described above will not be applied. In other words, in the case of overlapping pig boxes, it can be assumed that the reliability value was calculated to be low due to the adjacent pig box, and based on this assumption, the reliability value of the box is calculated as the non-overlapping area of the box compared to the total size (S _total ) (710). Call the Increase_Confidence() function using the size (S _{not_overlapped} ) (720) and correct it with the following formula:

Referring to FIG. 7, the total size (S _total ) (710), the size of the non-overlapped area of the corresponding box (S _{not_overlapped} ) (720), and adjacent boxes (730) are shown. The size of the non-overlapped area (S _{not_overlapped} ) 720 can be calculated using information on the adjacent box 730.

Additionally, the overall correction process of the reliability value using foreground pixel information and adjacent box information can be corrected according to the algorithm of FIG. 2 described above.

Figure 8 is a diagram illustrating the process of matching after applying the reliability correction algorithm to the original image and the rotated image, respectively, according to an embodiment of the present invention.

After applying the reliability correction algorithm to the original image and the rotated image respectively, assume that the results detected in the original image (A, B, C, D) and the results detected in the rotated image (a, b, c, d) are do. At this time, using the Euclidean Distance, the object detected as A in the original is rotated and then detected as a and b in the image, and the object detected as B in the original is rotated and then detected as c in the image. They are matched as detected, and after the objects detected as C and D in the original are rotated, n:m matching can be done with the objects detected as D in the image.

Figure 9 is a diagram for explaining the process of separating objects in an image using an adaptiveThreshold technique according to an embodiment of the present invention.

After separating objects using an adaptiveThreshold technique for the original image and rotated image, a detection result that includes more foreground is selected from the detected results.

The reliability value of the box containing the pixels of the object excluding the background among the original image and the rotated image is corrected using the following equation:

Here, P _{overlapped_Pixel} represents the size (910) of the overlapped area between the original image and the rotated image, and P _{not_overlapped_Pixel} represents the size (920) of the non-overlapping area between the original image and the rotated image.

Objects in the rotated image and the non-rotated image, excluding the background, can be represented by an overlapped portion 910 and a non-overlapping portion 920 as shown in FIG. 9 . Referring to FIG. 9, since the overlapped portion 910 is two pixels overlapping, only 50% of the pixel value of the non-overlapping portion is reflected when calculated by setting Reflectivity = 0.5.

FIG. 10 is a diagram illustrating a box-level post-processing device for accurate object detection according to an embodiment of the present invention.

The proposed box-level post-processing device for accurate object detection includes an image collection unit 1010, an object detection unit 1020, and a correction unit 1030.

The image collection unit 1010 acquires images of a plurality of objects photographed through a camera. For example, in order to photograph multiple objects through a camera, photographing can be done through a top-view camera that photographs from above and below.

The object detection unit 1020 obtains boxes for detecting a plurality of objects from the acquired image through the object detection unit. Each reliability value can be obtained by operating a detector for images acquired from installed cameras. The object detector 1020 according to an embodiment of the present invention can use YOLOv4, a cost-effective object detector that appropriately matches time and accuracy.

The correction unit 1030 corrects the reliability value of the fake object box and the reliability value of the real object box due to overlap between multiple objects among the multiple object detection boxes using foreground pixel information within the box.

The correction unit 1030 corrects the reliability value of the fake object box and the reliability value of the real object box due to overlap between multiple objects among the multiple object detection boxes using foreground pixel information within the box.

7The correction unit 1030 first calculates the degree of inclusion of foreground pixels in the box in order to determine whether it is a fake object box or a real object box. If the foreground pixels included in the box are less than a predetermined standard, it is judged as a fake object box, and if the foreground pixels included in the box are more than the predetermined standard, it is judged as a real object box. Afterwards, an adaptive threshold method is used for the real object box to separate the foreground and background pixels within the box.

The correction unit 1030 determines the reliability value of the fake object box and the real object box according to the separated foreground pixels and background pixels by using the ratio of the size of the foreground area within the box to the total size of the box. Correct.

In addition, if the reliability value of the real object box is calculated to be low due to overlapping areas among the real object boxes, the correction unit 1030 uses the ratio of the size of the non-overlapping area within the box to the total size of the box to determine the reliability value. Correct.

To correct the reliability value for the overlapping area among the real object boxes, the box for the object detected in the original image is matched with the box for the object detected in the rotated image that rotated the original image. Afterwards, the adaptive threshold method is used for the original image and the rotated image to separate the foreground and background pixels within the box, and then the reliability value for the box containing more foreground pixels among the original image and the rotated image is corrected. .

The object detection unit 1020 readjusts the positions of boxes for detecting a plurality of objects through the object detection unit according to the reliability value corrected through the correction unit 1030.

The device described above may be implemented with hardware components, software components, and/or a combination of hardware components and software components. For example, devices and components described in embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It may be implemented using one or more general-purpose or special-purpose computers, such as a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. A processing device may execute an operating system (OS) and one or more software applications that run on the operating system. Additionally, a processing device may access, store, manipulate, process, and generate data in response to the execution of software. For ease of understanding, a single processing device may be described as being used; however, those skilled in the art will understand that a processing device includes multiple processing elements and/or multiple types of processing elements. It can be seen that it may include. For example, a processing device may include a plurality of processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are possible.

Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or may be processed independently or collectively. You can command the device. Software and/or data may be used on any type of machine, component, physical device, virtual equipment, computer storage medium or device to be interpreted by or to provide instructions or data to a processing device. It can be embodied in . Software may be distributed over networked computer systems and thus stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the medium may be specially designed and configured for the embodiment or may be known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -Includes optical media (magneto-optical media) and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include machine language code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

As described above, although the embodiments have been described with limited examples and drawings, various modifications and variations can be made by those skilled in the art from the above description. For example, the described techniques are performed in a different order than the described method, and/or components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or other components are used. Alternatively, appropriate results may be achieved even if substituted or substituted by an equivalent.

Therefore, other implementations, other embodiments, and equivalents of the claims also fall within the scope of the claims described below.

<Reference materials>

[1] S. Matthews, et al., “Early Detection of Health and Welfare Compromises through Automated Detection of Behavioral Changes in Pigs,” The Veterinary Journal , Vol. 217, pp. 43-51, 2016.

[2] L. Liu, et al., “Deep Learning for Generic Object Detection: A Survey,” International Journal of Computer Vision , Vol. 128, pp. 261-318, 2020.

[3] A. Bochkovskiy, C. Wang, and H. Liao, “YOLOv4: Optimal Speed and Accuracy of Object Detection,” arXiv preprint arXiv:2004.10934 , 2020.

[4] T. Lin, et al., “Microsoft COCO: Common Objects in Context,” In Proceedings of the European Conference on Computer Vision (ECCV) , 2014.

[5] I. Blayvas, A. Bruckstein, and R. Kimmel, “Efficient Computation of Adaptive Threshold Surfaces for Image Binarization,” Pattern Recognition , Vol. 18, no. 1, pp. 89-101, 2006.

Claims (10)

Obtaining images by photographing a plurality of objects using a camera;
Obtaining boxes for detecting a plurality of objects from the acquired image through an object detection unit;
Correcting, through a correction unit, the reliability value of the fake object box and the reliability value of the real object box due to overlap between a plurality of objects among the boxes for detecting a plurality of objects using foreground pixel information within the box; and
A step of readjusting the positions of boxes for detecting a plurality of objects through an object detection unit according to the corrected reliability value.
Including,
The step of correcting, through the correction unit, the reliability value of the fake object box and the reliability value of the real object box due to overlap between a plurality of objects among the boxes for detecting a plurality of objects using the foreground pixel information in the box,
In order to determine whether it is a fake object box or a real object box, the degree of inclusion of foreground pixels in the box is calculated. If the foreground pixels included in the box are less than a predetermined standard, it is judged to be a fake object box, and the box is judged to be a fake object box. If the included foreground pixels are greater than a predetermined standard, it is judged to be a real object box, and an adaptive threshold method is used for the real object box to separate the foreground and background pixels within the box.
Object detection method. delete According to paragraph 1,
According to the separated foreground pixels and background pixels, the reliability value of the fake object box and the real object box are corrected using the ratio of the size of the foreground area within the box to the total size of the box.
Object detection method. According to paragraph 3,
If the reliability value of the real object box is calculated to be low due to overlapping areas among the real object boxes, the reliability value is corrected using the ratio of the size of the non-overlapping area within the box to the total size of the box.
Object detection method. According to paragraph 4,
To correct the reliability value for the overlapping area among the real object boxes, the box for the object detected in the original image is matched with the box for the object detected in the rotated image where the original image is rotated,
After separating the foreground and background pixels in the box using an adaptive threshold method for the original image and the rotated image, the reliability value for the box containing more foreground pixels among the original image and the rotated image is corrected.
Object detection method. An image collection unit that acquires images of a plurality of objects photographed through a camera;
an object detection unit that acquires boxes for detecting a plurality of objects from the acquired image and readjusts the positions of the boxes for detecting the plurality of objects according to the reliability value corrected through the correction unit; and
A correction unit that corrects the reliability value of the fake object box and the reliability value of the real object box due to overlap between multiple objects among the boxes for multiple object detection using foreground pixel information within the box.
Including,
The correction unit,
In order to determine whether it is a fake object box or a real object box, the degree of inclusion of foreground pixels in the box is calculated. If the foreground pixels included in the box are less than a predetermined standard, it is judged to be a fake object box, and the box is judged to be a fake object box. If the included foreground pixels are greater than a predetermined standard, it is judged to be a real object box, and an adaptive threshold method is used for the real object box to separate the foreground and background pixels within the box.
Object detection device. delete According to clause 6,
Correction part,
According to the separated foreground pixels and background pixels, the reliability value of the fake object box and the real object box are corrected using the ratio of the size of the foreground area within the box to the total size of the box.
Object detection device. According to clause 8,
Correction part,
If the reliability value of the real object box is calculated to be low due to overlapping areas among the real object boxes, the reliability value is corrected using the ratio of the size of the non-overlapping area within the box to the total size of the box.
Object detection device. According to clause 9,
Correction part,
To correct the reliability value for the overlapping area among the real object boxes, the box for the object detected in the original image is matched with the box for the object detected in the rotated image where the original image is rotated,
After separating the foreground and background pixels in the box using an adaptive threshold method for the original image and the rotated image, the reliability value for the box containing more foreground pixels among the original image and the rotated image is corrected.
Object detection device.