KR101851823B1 - Method for classifying children and adult bt using head and body height in images - Google Patents

Abstract

The present invention relates to a method and apparatus for detecting a pedestrian area from a video image, detecting a head area, calculating a ratio therebetween, and determining whether or not an adult is below a predetermined threshold value. How to identify adults ".
The method of distinguishing between a child and an adult using the ratio of the height to the height of the head in the video image proposed by the present invention,
Detecting a pedestrian area from a video image as a first rectangular area;
Detecting a head area of the detected pedestrian as a second rectangular area;
Determining a height ratio of the head area to a height of the pedestrian area as a height ratio of a second rectangle to a height of the first rectangle;
And judging that the child is an adult when the height ratio is less than or equal to a predetermined threshold value;
According to the present invention, since a pedestrian area is detected from a video image, a head area is detected, a ratio between them is determined, and then it is determined whether an adult is below a predetermined threshold value. .

Description

[0001] The present invention relates to a method for classifying children and adults using a ratio of height to head height in a video image,

The present invention relates to a method and apparatus for detecting a pedestrian area from a video image, detecting a head area, calculating a ratio therebetween, and determining whether or not an adult is below a predetermined threshold value. How to identify adults ".

Object detection is a technology related to image processing that deals with sample detection from semantic objects. Such object detection has a long history in image research.

As technology advances and demands in the substantive field increase, research into this area continues for the benefit of mankind.

By object detection, pedestrian, vehicle, and gender classification, or age classification can be identified as digital images and images.

Various applications have been developed in various fields related to computer vision in relation to object detection.

One of the main purposes of object detection is to find the exact location of the desired target in the field.

On the other hand, the process of detecting objects and understanding the situation by image processing is required to be quick and reliable.

One of the most important steps of object detection is feature selection. In order to detect objects in real time, the feature should be robust, discriminative and easy to calculate [1].

In recent years, the social role of children has become prominent, and children are generally characterized by not paying attention to objects such as street cars or other people.

For this reason, children in particular situations such as traffic situations require more attention than adults, and children can be used to construct data sets in areas such as health, sociology, and traffic [2].

Classifying people's appearances between adults and children is a subclass that detects pedestrians with some sort of mileage from object detection. Getting information about physical appearance in society is needed in many fields.

For that purpose, applications that need to determine location and presence have been developed, and many effective methods for age detection have been proposed.

However, in the case of a variety of conventional applications, there is a problem that the method of distinguishing adults and children from each other is not properly conducted.

1. K. Lee, C.H. Lee, S.A. Kim, Y.H. Kim, "Fast Object Detection Based on Color Histograms and Local Binary Patterns," presented at the TENCON IEEE Region 10 Conference, Cebu, Nov. 19-22, 2012. 2. H. Weda, M. Barbieri, "Automatic Children Detection in Digital Images," presented at the IEEE International Conference on Multimedia and Expo, Beijing, July. 2-5, 2007 3. Y.H. Kwon and N. da Vitoria Lobo (April 1999), Age classification from facial images, Computer Vision and Image Understanding, Vol. 74, No. 1, pp. 1-21, Available: http://www.idealibrary.com 4. W.B Horng, C.P. Lee and C.W. Chen (2001), Classification of age groups based on facial features, Tamkang Journal of Science and Engineering, Vol. 4, No. 3, pp. 183-192. 5. A. Lanitis, C. Draganova, and C. Christodoulou. "Comparing different classifiers for automatic age estimation," IEEE Transactions on Systems, Man and Cybernetics, Part B, 34 (1): 621-628, 2004. 6. X. Geng, Z. Zhou, Y. Zhang, G. Li, and H. Dai. "Learning from facial aging patterns for automatic age estimation" Proceedings of ACMMultimedia'06, pp.307-316, 2006. 7. Viola and Jones, "Rapid object detection using a boosted cascade of simple features ", Computer Vision and Pattern Recognition, 2001

In the present invention, a method for distinguishing between children and adults is proposed.

To this end, the present invention proposes a method of distinguishing an adult from a child by obtaining the ratio of the height of the pedestrian and the head size obtained from the image.

In other words, in the present invention, a method of detecting an adult (adult or child) using a ratio of height to head height of a pedestrian is focused on a medical report that a head size of a person is reduced as compared with a child's height It is aimed to propose.

Therefore, the main frame of the present invention is to detect a pedestrian in the field and obtain the ratio of the height of the pedestrian to the height of the head of the pedestrian, and if necessary, a reference table relating to the classification of the person's age may be used.

For reference, in the present invention, when a specific body part displayed on a screen is covered, it is omitted and a pedestrian which is not covered is classified.

The method of distinguishing between a child and an adult using the ratio of the height to the height of the head in the video image proposed by the present invention,
Detecting a pedestrian area from a video image as a first rectangular area;
Detecting a head area of the detected pedestrian as a second rectangular area;
Determining a height ratio of the head area to a height of the pedestrian area as a height ratio of a second rectangle to a height of the first rectangle;
And judging that the child is an adult when the height ratio is less than or equal to a predetermined threshold value;

According to the present invention, since a pedestrian area is detected from a video image, a head area is detected, a ratio between them is determined, and then it is determined whether an adult is below a predetermined threshold value. .

Figure 1 is an example of a Har-like feature set.
2A to 2F are diagrams that capture the detection results for children and adults.

Hereinafter, the present invention will be described.

The present invention is configured as follows.

First, a related work related to the types of features that can be used for adult and child detection will be described, and then an algorithm proposed by the present invention will be described.

[Related tasks]

A. Pre-work

Currently, there are some published research examples on predicting the age of a person from an image.

As a first example, Kwon & Lobo [3] began to predict the age of a person from a digital facial image, which describes how to classify into three groups: babies, young adults, and adults of all ages. It is not 100% accurate in a limited database when applied in a real environment. Other techniques based on the theory of craniofacial development have been proposed to completely categorize into similar age groups [3].

However, this approach was also disappointing due to limited data sets, which makes it very difficult to calculate the robustness when running under real data set conditions.

Meanwhile, W.B. Horng, C.P. Lee and C.W. There are other methods proposed by Chen [4]. This method considers an additional group as an adult of intermediate age for the article of Kwon & Lobo. However, the greatest weakness of this method is that it does not detect children and adults under 40 years of age. That is, only the baby can be detected. Nevertheless, the results were satisfactory. However, the analysis for characterization is quite complex.

Lanitis [5] also presented an active appearance model.

To estimate age, the combination parameter of shape and texture is removed using a classifier. We compared the age estimation accuracy of simple quadratic fitting, shortest distance classifier and neural network with those classifiers.

Geng [6] proposed statistical modeling of age patterns. This modeling is based on the idea that multiple images of different ages are available for each person.

B. Histogram of Oriented Gradients Features

When detecting a human body part, various features can be used. HOG is a feature used for object detection.

The main frame is to count the occurrence of the defined portion of the image in the gradient direction. The key to the HOG descriptor is that in a digital image, the local object appearance and shape can be described as a distribution of edge directions or intensity. In order to obtain better accuracy, it is necessary to normalize the local histogram.

Haar-Like Feature

The Haar-like feature may be given as another example using features for object detection. The name comes from the Haar wavelet, and the features were used for the first real-time face detection.

Viola and Jones developed a Har-like feature from the idea of using a Har wavelet. [7]

The Haar-like feature takes into account the rectangular regions close to each other in the detection window, adds the intensity of the pixels in each region, and evaluates the difference of these totals. These cars used to categorize parts of the image.

Figure 1 is an example of a Har-like feature set.

In Fig. 1, a Har-like feature set can be seen. In the detection process of the Viola-Jones object detection framework, a Har-like feature is calculated with the window of the target size behind. Next, the threshold for separating non-objects is compared to this difference. The main advantage of the Har-like feature lies in the computational speed.

Hereinafter, an algorithm according to the present invention for detecting a child and an adult in a digital image will be described.

The sequential steps for adult children classification proposed in the present invention are as follows. 1) The method of adult children classification includes detection of a pedestrian zone, 2) detection of a head region, 3) Length ratio), and 4) an adult determination step.

The algorithm of the present invention is composed of several steps, which can be divided into two main parts.

The first part is to detect the head area of the pedestrian and the pedestrian.

The second part is to calculate and determine the classification of children and adults.

In a digital image, it trains with Haar-like features with an Adaboost algorithm that can be used for efficient and pedestrian detection problems to detect pedestrian areas.

There are various objects in a digital image.

It is possible to detect the head of a pedestrian or a pedestrian in a digital image when the ROI (Region of Interest) is a pedestrian or a pedestrian's head.

For execution we used a Harcascade detector.

After detecting the body, our new area of interest will be the body of the pedestrian. Thereafter, it was necessary to detect the head of the pedestrian.

At this stage we used a Harcascade detector.

By performing these steps, the first two steps shown in the block diagram end.

In the third step, the head-to-body ratio needs to be calculated as shown in equation (1) below.

r = _IH / I _B , 0 < r < 1 (1)

I _B and I _H are the height of the body and head, respectively. r is the ratio of I _B to I _H.

The main reason for adding these calculation factors is that the human head height becomes relatively smaller than the height as it grows. Since the absolute value of the human height can not be measured from the pixels of the digital image, the present invention attempts to solve this problem by applying the relative measurement like the above-described method of calculating the head-to-height ratio.

According to this technical idea, the present invention makes it possible to classify children and adults.

For example, if the ratio of the head to the body is about 0.2, the pedestrian can be classified as a child, and if the ratio is about 0.15, the pedestrian can be classified as an adult. Of course, there may be an error here, but children and adults can distinguish between adults and children based on whether or not they belong to the relevant section after establishing a predetermined interval.

In order to perform this classification operation, the present invention uses a default cascade of OpenCV.

Based on these algorithms, the present invention performs experiments for verification of the present invention.

In the experimental phase, various images of the INRIA data set and the CVC-CER-01 (Computer Vision Center) pedestrian data set were used, and all training and testing procedures used Microsoft Visual Studio 2010 and OpenCV2.2.

Regarding the output image, the adult pedestrian zone is marked with a green rectangle and the child pedestrian zone is marked with a blue rectangle. The red rectangular area indicates the body of the pedestrian (first rectangle: pedestrian area), and the green and blue rectangular areas indicate the head area of the pedestrian (second rectangle: head area).

In some images, some false positives and correct negatives have been observed, which is an acceptable margin of error.

Figs. 3A to 3F are diagrams that capture detection results for children and adults.

Figures 3a and 3b can be good examples of successful adult detection.

All adults, as mentioned above, are represented by our goal, the green rectangle.

In Figures 3c and 3d some false positive and true negatives can be seen.

In Fig. 3c, the algorithm detects two children, one for each child and one for adults, but not in the image anyway.

In Figure 3D, everything is correctly detected, but the woman's head is missing, which may be due to the inadequacy of our data set.

We obtain the correct child detection results in Figures 3e and 3f.

The present invention classifies adults and children from digital images. Although the height of the pedestrian could not be precisely estimated, the approach of the present invention was found to be able to effectively classify adults and children.

In the present invention, the head and the kidney area of a pedestrian were detected by applying Haar Cascade, which is a widely used technique in object detection, and then the ratio of the height to the head was obtained.

In order to realize the present invention, even if all known technologies for detecting an existing pedestrian area and all known technologies for detecting a head area of a pedestrian are used and the technical idea of the present invention is implemented by combining them, It should be considered to fall within the scope.

Claims (1)

A method for distinguishing a child from an adult using a ratio of height to head height in a video image,
Detecting a pedestrian area from a video image as a first rectangular area;
Detecting a head area of the detected pedestrian as a second rectangular area;
Determining a height ratio of the head area to a height of the pedestrian area as a height ratio of the second rectangle to a height of the first rectangle;
And judging the child as an adult when the height ratio is less than a predetermined threshold value; and determining the child as an adult if the height ratio is less than a predetermined threshold value.

Images