KR101866381B1 - Apparatus and Method for Pedestrian Detection using Deformable Part Model - Google Patents

Abstract

In the present invention, a plurality of articles using a DPM (Deformable Part Model) use an average of four pedestrian parts, and the smallest part is three, but the present invention uses only two parts An image dividing unit for dividing the pedestrian learning data of an input image into images of two parts of an upper half body and a lower half body, A HOG feature extraction unit for extracting a Histogram of Oriented Gradient (HOG) feature common to the input images to detect an image of a pedestrian; and a display unit for displaying a detailed shape information of a pedestrian image of the HOG feature extracted from the HOG feature extraction unit A DPM matching unit which performs matching with learning data of two parts divided in the image division unit, And a pedestrian detection unit for detecting a pedestrian image from a neighboring facility to a pedestrian area covered with a part of the body through matching.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to a pedestrian detection apparatus and a pedestrian detection method using a deformable part model,

The present invention relates to an apparatus and method for detecting a pedestrian, and more particularly, to a method and apparatus for detecting a pedestrian even when occlusion occurs due to obstacles of a pedestrian due to various objects such as a car, a pedestrian, Apparatus and method.

Pedestrian detection is currently used in autonomous vehicles or ADAS and pedestrian protection systems (PPS). The present invention uses an image photographed in a road driving environment for a pedestrian traveling on a street or a pedestrian using a pedestrian crossing focusing on a driver protection system.

The pedestrian detection process first extracts features from the acquired images, and then creates a function with features common to pedestrian images through machine learning. And finally detecting the pedestrian area through matching in the target image.

At this time, the detection of pedestrians is a very challenging task and various studies are being carried out because the detection of pedestrians has a variety of shape changes depending on the action, unlike face and automobile. Haar features, HOG, and LBP are used as representative image features, and these features are learned by various machine learning methods.

The method used in almost all pedestrian detection systems is based on the HOG (N. Dalal and B. Triggs, "Histograms of Oriented Gradients for Human Detection", Proc. IEEE Conf. Computer Vision and Pattern Recognition, Vol. -893, 2005). However, since the detector using the existing HOG (Histogram of Oriented Gradient) feature uses a single object model, it can be detected only when the entire body of the pedestrian exits. In other words, the detection rate is significantly lower when a part of the body is covered by a vehicle or a nearby facility.

In order to overcome such a problem, it is proposed to use the method of Felzenszwalb (PFFelzenszwalb, RBGirshick, D. McAllester, D. Ramanan, " Object Detection with Discriminatively Trained Part Based Model ", Pattern Analysis and Machine Intelligence, Vol.32, pp.1627-1645, 2010) proposed a DPM (Deformable Part Model). This technology has the advantage of detecting not only the whole image of an object but also the parts constituting the object, so that even a part of the object can be detected.

Felzenszwalb, who first devised the DPM (Deformable Part Model), divided pedestrians into five parts (Head, Right-shoulder, Left-shoulder, Upper-leg and Lower-leg) In the conventional technique used, an object is detected by dividing the portion of the pedestrian by an average of four.

However, there is a problem in that the amount of processing for object detection through the preprocessing and detection of the feature points of the image is increased, while the detection rate of the object can be improved as the number of such divided parts increases for the detection of the object. The detection speed is lowered and the detection time of the object takes longer.

Registered Patent Publication No. 10-1419837 (registered date 2014.07.09) Published Japanese Patent Application No. 10-2011-0073924 (published on June 30, 2011)

Accordingly, the present invention has been devised to solve the above-mentioned problems. In the DPM (Deformable Part Model), several articles of pedestrians are used as an average of four parts, and the smallest parts are three The present invention aims at providing a pedestrian detecting apparatus and method using a deformable part model by simply using two parts.

Other objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned can be clearly understood by those skilled in the art from the following description.

According to another aspect of the present invention, there is provided a pedestrian detection apparatus using a deformable part model, the pedestrian detection method comprising the steps of dividing pedestrian learning data of an input image into two parts of an upper body and a lower body, A HOG feature extraction unit for extracting a HOG (Histogram of Oriented Gradient) feature common to the input images and detecting an image of a pedestrian; and a HOG feature extraction unit for extracting the HOG feature extracted from the HOG feature extraction unit through learning A DPM matching unit for matching detailed shape information of a pedestrian image with learning data of two parts divided in the image division unit; And a pedestrian detection unit for detecting a pedestrian image up to the covered pedestrian zone.

Preferably, the HOG feature detection unit includes a feature detection unit for detecting each feature in the image through the pedestrian learning data of the input image, and a feature function generation unit for generating a function having the features common to pedestrian images through the learning .

Preferably, the DPM matching unit includes a transformation unit that transforms an image of each part classified into the upper half body and the lower half body in the learned model and an anchor position defined as a position relative to the HOG feature extracted from the HOG feature extraction unit And a binary classifier for calculating an average precision based on the position (distance).

를 이용하여 평균 정밀도를 산출하며, 이때, 상기

는 학습된 필터들에 의한 점수를 나타내고, 상기

는 부분(part)(x, y)들의 위치 변형에 의한 벌칙을 나타내는 것을 특징으로 한다.Preferably, the binary classifier comprises:

The average precision is calculated using the equation

Represents the score by the learned filters,

(X, y) represents a penalty due to the positional deformation of the part (x, y).

According to another aspect of the present invention, there is provided a method for detecting a pedestrian using a deformable part model, the method comprising: (A) extracting a Histogram of Oriented Gradient (HOG) (B) dividing the pedestrian learning data of the input image into two parts of an upper body and a lower body through an image partitioning part; (C) A step of matching detailed shape information of a pedestrian image of the HOG feature with the learning data of the two divided parts through a DPM matching unit; and (D) And detecting a pedestrian image to a covered pedestrian zone.

Preferably, the step (A) includes the steps of extracting a Histogram of Oriented Gradient (HOG) characteristic from the collected image, and detecting a pedestrian image by generating a function having characteristics common to pedestrian images through machine learning And the like.

)들의 합에서 부분(part)(x, y)들의 위치 변형에 의한 벌칙(

)을 뺀 점수를 이용하여 평균 정밀도를 산출하여 매칭을 수행하는 것을 특징으로 한다.Preferably, the step (C) further comprises:

(X, y) by the displacement of the position of the part (x, y)

) Is subtracted from the average precision, and the matching is performed.

The pedestrian detection apparatus and method using the deformable part model according to the present invention as described above have the following effects.

The present invention is robust even when occlusion occurs, and is a detector using a simple pedestrian part. This detector improves the detection rate and can be applied to a system that improves the detection rate of pedestrians.

And if you use CUDA or parallel processing to increase detection speed because of high estimation accuracy, it can be developed as full automatic system.

1 is a block diagram showing a configuration of a pedestrian detecting apparatus using a deformable part model according to an embodiment of the present invention
2 is a block diagram showing a state in which a pedestrian image is divided into images of two parts through an image dividing unit in FIG.
3 is a flowchart for explaining a pedestrian detection method using a deformable part model according to an embodiment of the present invention.
FIG. 4 is a view for explaining the process of detecting a pedestrian image in FIG.
5 is a diagram for explaining a process of performing matching through the DPM matching unit in FIG.

Other objects, features and advantages of the present invention will become apparent from the detailed description of the embodiments with reference to the accompanying drawings.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A preferred embodiment of a pedestrian detection apparatus and method using a deformable part model according to the present invention will now be described with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. It is provided to let you know. Therefore, the embodiments described in the present specification and the configurations shown in the drawings are merely the most preferred embodiments of the present invention and are not intended to represent all of the technical ideas of the present invention. Therefore, various equivalents It should be understood that water and variations may be present.

1 is a block diagram showing a configuration of a pedestrian detecting apparatus using a deformable part model according to an embodiment of the present invention.

As shown in FIG. 2, the pedestrian detection apparatus detects the pedestrian learning data of the input image in the upper half and lower half of the image, as shown in FIG. 2, A HOG feature extraction unit 200 for extracting a Histogram of Oriented Gradient (HOG) characteristic common to input images and detecting an image of a pedestrian, A DPM matching process for matching the detailed shape information of the pedestrian image of the HOG feature extracted by the HOG feature extraction unit 200 with the learning data of the two parts in the image divider 100, And a pedestrian detection unit 300 for detecting a pedestrian image from neighboring facilities to a pedestrian zone where a part of the body is covered through matching performed by the DPM matching unit 300. [ And a detection unit 400.

The HOG feature detection unit 200 includes a feature detection unit 210 for detecting each feature in the image through the pedestrian learning data of the input image, and a function generation unit 210 for generating a function having features common to pedestrian images And a feature generating unit 220 for generating a feature function.

Since the HOG characteristic detecting unit 200 uses a single object model, the entire body of the pedestrian must be entirely detected to detect the pedestrian area. In other words, the detection rate is significantly reduced when a part of the body is covered by a vehicle or a nearby facility.

In order to solve this problem, the DPM matching unit 300 extracts images of each part classified into the upper half body and the lower half body from the learned model through the binary classifier 310 and the images extracted from the HOG feature extraction unit 200 The average precision is calculated based on the anchor position and the deformation position (distance) defined by the relative position with respect to the HOG feature.

The binary classifier 310 calculates the average precision through the following equation (1).

)들의 합에서 부분(part)(x, y)들의 위치 변형에 의한 벌칙(

)을 뺀 점수가 된다.As in Equation (1) above, the average precision is determined by the score

(X, y) by the displacement of the position of the part (x, y)

).

The performance index comparing the case of using only the HOG feature detector 200 and the case of using up to the DPM matching unit 300 was calculated by an average precision which is an index of the PASCAL VOC. As a result, it was confirmed that the detection rate was improved by 1.5 AP as shown in Table 1 below.

This shows that the detection rate is 10% higher than the existing algorithm when a part of the pedestrian is occlusion occluded by 10 ~ 30% to an object.

The operation of the pedestrian detecting device using the deformable part model according to the present invention will be described in detail with reference to the accompanying drawings. Like reference numerals in FIG. 1 or FIG. 2 denote the same members performing the same function.

3 is a flowchart illustrating a pedestrian detection method using a deformable part model according to an embodiment of the present invention.

Referring to FIG. 3, the HOG feature extraction unit 200 extracts a HOG (Histogram of Oriented Gradient) feature common to the images, and detects a pedestrian image (S10).

In this case, as shown in FIG. 4, the process of detecting the pedestrian image extracts HOG (Histogram of Oriented Gradient) features from the collected image, and then generates a function having features common to pedestrian images through machine learning . The detection of such a pedestrian image is a very challenging task because it has a variety of shape changes according to its motion, unlike a face and an automobile, and various studies have been made.

Next, as shown in FIG. 2, the image dividing unit 100 divides the pedestrian learning data of the input image into images of two parts of an upper body and a lower body (S20).

5, detailed information on the shape of a pedestrian image of the HOG feature extracted from the HOG feature extraction unit 200 through learning is transmitted to the image division unit 100 through the DPM matching unit 300 And performs matching with the learning data of the two divided parts (S30).

In other words, the learned data uses a start model consisting of a low-resolution root filter containing shape information of the entire object and higher-resolution part filters containing detailed shape information of the part. Each part in the learned model has a defined anchor position with respect to the root filter and each part is penalized by the given equation (1) as it gets further away from this reference position. That is, the score by the binary classifier 310 becomes a score obtained by subtracting the penalty caused by the position transformation of the parts from the sum of the scores by the learned filters.

Accordingly, the pedestrian detecting unit 400 detects the pedestrian image from the surrounding facilities to the pedestrian zone where some of the body is covered through the matching performed in the DPM matching unit 300 (S40). This improves the detection rate as described in Table 1 above.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. It will be apparent to those skilled in the art that various modifications may be made without departing from the scope of the present invention. Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims.

Claims (7)

An image dividing unit for dividing the pedestrian learning data of an input image into images of two parts of an upper half body and a lower half body,
A HOG feature extracting unit for extracting a Histogram of Oriented Gradient (HOG) characteristic common to input images and detecting an image of a pedestrian;
A DPM matching unit for matching detailed shape information of the HOG feature pedestrian image extracted by the HOG feature extraction unit with learning data of two parts divided in the image partition unit,
And a pedestrian detection unit for detecting a pedestrian image using a degree of deformation according to a deformation position from neighboring facilities to a pedestrian zone where a part of the body is covered through matching performed in the DPM matching unit,
The DPM matching unit may include a transformed position of an anchor position defined as a relative position between the images of each part classified into upper and lower halves of the learned model and the HOG feature extracted from the HOG feature extraction unit Distance), and a binary classifier for calculating an average precision is constructed.
The binary classifier may be represented by the following equation

To calculate the average precision,
At this time,

Represents the score by the learned filters,

(X, y) represents a penalty due to positional deformation of the part (x, y). The apparatus of claim 1, wherein the HOG feature detector comprises:
A feature detecting unit that detects each feature in an image through pedestrian learning data of an input image;
And a characteristic function generation unit for generating a function having features common to pedestrian images through the learning. delete delete delete delete delete

Images