KR102201168B1 - Method for tracking a person using an overhead camera - Google Patents

Abstract

)를 생성하는 단계, 각 클러스터 객체(

)마다 하기 (수학식 2)으로 표현되는 추적 객체(

)를 생성하는 단계, 상기 추적 객체(

)를 통하여 하기 (수학식 3)으로 표현되는 추적 오브젝터 목록(

)을 생성하는 단계, 상기 추적 오브젝터 목록(

)에서 다음 프레임의 동일한 영역 내 추적 객체를 선택하여, 상기 다음 프레임에서 선택된 추적 객체의 위치를 추정하는 단계, 상기 다음 프레임의 탐색 영역(

)을 하기 (수학식 4)와 같이 계산하는 단계 및 상기 다음 프레임의 탐색 영역 내에 존재하는 사람을 검출하는 단계를 포함한다.
본 발명은 정지된 사람을 오랜 시간 동안 탐지하는 것이 가능하고, 가려진 것을 처리하는 것, 환경의 급격한 변화를 가져오는 것 등을 보정하여 추적을 효과적이고 정확하게 수행할 수 있다.The present invention discloses a person tracking apparatus, detecting a person present in an image, and returning a cluster in which the detected person is located, a cluster object represented by the following (Equation 1) for each of the clusters (

) To create each cluster object (

) For each tracking object represented by the following (Equation 2) (

), the tracking object (

) Through the list of tracking objects represented by the following (Equation 3) (

) Generating, the tracking object list (

Selecting a tracking object in the same area of the next frame in ), estimating the position of the selected tracking object in the next frame, the search area of the next frame (

) As follows (Equation 4), and detecting a person present in the search area of the next frame.
According to the present invention, it is possible to detect a stationary person for a long time, and to perform tracking effectively and accurately by correcting for processing a hidden object, causing a sudden change in environment, and the like.

Description

Person tracking method using overhead camera {METHOD FOR TRACKING A PERSON USING AN OVERHEAD CAMERA}

The present invention relates to a method for tracking a person using an overhead camera, and more particularly, to a method for tracking and identifying a person by analyzing images and videos.

The current video analysis technology is rapidly changing the Internet of Things through research and technology development, thereby creating new opportunities. It can be said that one of the application fields of the Internet of Things (IoT) is to enable the camera to automatically recognize people and other moving objects by using machine learning algorithms on images and videos obtained from video analysis.

One of these areas of research in the Internet of Things (IoT) and computers is the technology of tracking and identifying people in images and videos. The technology of tracking a person can be applied to areas such as driving assistance systems, human walking behavior analysis and monitoring systems, and the like.

On the other hand, when it is desired to analyze the work flow in an actual environment by observing various situations of workers working in industrial sites, a technology for identifying and tracking people can be used. In this case, a side/normal view or an overhead view camera may be used. In general, an overhead view camera may more accurately track the movement.

FIG. 1(a) is an image of people obtained through a side/normal view, and FIG. 1(b) is an image of people obtained through an overhead view. 1 (a) and (b) are images of the same scene simultaneously captured by two different cameras.

In (a) and (b) of Fig. 1, people are represented by white ellipses. Referring to Figure 1 (a), it can be seen that people are covered with various equipment such as machines, electric wires, and shelf parts. In the case of using the side/normal view, it is not possible to fully grasp the movement of people. For example, when working behind a large machine, it is difficult to identify the person. In this case, it is very difficult to identify people and identify movements.

1B illustrates an image captured using an overhead view (or top view) camera equipped with a wide-angle lens. Referring to (b) of FIG. 1, it can be seen that an overhead camera equipped with a wide-angle lens is used so that images of people covered by a machine or the like can be clearly grasped. In this way, when an overhead view camera is used, objects can be projected more clearly and clearly even in a hidden or cluttered environment.

On the other hand, factors that make the process of tracking people difficult are noise in images, changes in the shape of objects, and sudden movements of objects. Such as camera movement. The conventional human tracking method has a problem in that the accuracy is very poor because the above problem cannot be completely solved.

In order to solve the above problems, an object of the present invention is to provide a method for tracking a person with high accuracy using an overhead camera.

)를 생성하는 단계, 각 클러스터 객체(

)마다 하기 (수학식 2)으로 표현되는 추적 객체(

)를 생성하는 단계, 상기 추적 객체(

)를 통하여 하기 (수학식 3)으로 표현되는 추적 오브젝터 목록(

)을 생성하는 단계, 상기 추적 오브젝터 목록(

)에서 다음 프레임의 동일한 영역 내 추적 객체를 선택하여, 상기 다음 프레임에서 선택된 추적 객체의 위치를 추정하는 단계, 상기 다음 프레임의 탐색 영역(

)을 하기 (수학식 4)와 같이 계산하는 단계 및 상기 다음 프레임의 탐색 영역 내에 존재하는 사람을 검출하는 단계를 포함한다.In order to solve the above technical problem, the present invention discloses a person tracking device, detects a person present in an image, and returns a cluster in which the detected person is located, as follows (Equation 1) for each of the clusters. The represented cluster object (

) To create each cluster object (

) For each tracking object represented by the following (Equation 2) (

), the tracking object (

) Through the list of tracking objects represented by the following (Equation 3) (

) Generating, the tracking object list (

Selecting a tracking object in the same area of the next frame in ), estimating the position of the selected tracking object in the next frame, the search area of the next frame (

) As follows (Equation 4), and detecting a person present in the search area of the next frame.

)을 생성하는 단계를 더 포함하는 것을 특징으로 한다.In addition, after the step of creating the cluster object, a list of cluster objects represented by the following (Equation 5) through the cluster object (

It characterized in that it further comprises the step of generating ).

) 및 속력(

)을 기초로 추정되는 것을 특징으로 한다.In addition, the location of the tracking object is the speed of the tracking object calculated through the following (Equation 6) (

) And speed (

It is characterized in that it is estimated based on ).

In addition, the location of the tracking object is characterized in that it is estimated by the following (Equation 7).

In addition, after the step of detecting a person in the search area of the next frame (a) if a person is not detected, the tracking object is removed, and (b) if a person is detected, the distance between the previous location and the current location Comparing, (b-1) updating the center of the tracking object and the cluster if the distance is less than 40 pixels, (b-2) generating a new tracking object if the distance is greater than 40 pixels. It features.

In addition, the process of detecting the person includes obtaining a motion map image through the following (Equation 8), dividing the motion map image into sub-areas of 8 × 8 pixels, and summing the pixel intensity values of the sub-areas. It characterized in that it is carried out in the step of detecting a person through the step and the following (Equation 9).

An effect of a method for tracking a person using an overhead camera according to embodiments of the present invention will be described as follows.

Compared to the accuracy of the conventional method (48%), the present invention can exhibit high accuracy (about 99%).

In addition, unlike the conventional method, the present invention can further accurately identify the presence or absence of a person.

In addition, according to the present invention, it is possible to detect a stationary person for a long period of time, and correcting for processing a hidden object or causing a sudden change in the environment can perform tracking effectively and accurately.

However, the effects that can be achieved by the method for tracking a person using an overhead camera according to the embodiments of the present invention are not limited to those mentioned above, and other effects not mentioned are the technology to which the present invention belongs from the following description. It will be clearly understood by those of ordinary skill in the field.

The accompanying drawings, which are included as part of the detailed description to aid in understanding of the present invention, provide embodiments for the present invention, and describe the technical idea of the present invention together with the detailed description.
FIG. 1(a) is an image of people obtained through a side/normal view, and FIG. 1(b) is an image of people obtained through an overhead view.
2 is a flowchart of a method for tracking a person according to the present invention.
3 shows a process of detecting spots according to the present invention.
4 is a test sequence for testing the performance of the present invention.
5 is a graph showing accuracy according to distances of all detected locations.
6 is a comparison between the tracking method of the present invention and the MS algorithm in Sequence 1 of FIG. 4.
7 is a comparison between the tracking method of the present invention and the MS algorithm in Sequence 2 of FIG. 4.
8 is a comparison of the tracking method of the present invention and the MS algorithm in Sequence 3 of FIG. 4.
9 is a comparison of the tracking method of the present invention and the MS algorithm in Sequence 4 of FIG. 4.

The terms or words used in this specification and claims are not limited to their usual or dictionary meanings and should not be interpreted, and the inventor should appropriately define the concept of terms in order to describe his own invention in the best way. Based on the principle that it is possible, it should be interpreted as a meaning and concept consistent with the technical idea of the present invention. Therefore, the embodiments described in the present specification and the configurations shown in the drawings are only one of the most preferred embodiments of the present invention, and do not represent all the technical ideas of the present invention, so that they can be replaced at the time of application. It should be understood that there may be various equivalents and variations. Hereinafter, a method for tracking a person using an overhead camera according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

2 is a flowchart of a method for tracking a person according to the present invention.

The present invention is based on a tracking algorithm capable of detecting the relevant moving area within an image. The present invention provides a simple yet effective human tracking method by incorporating the rHOG algorithm into this tracking algorithm. The rHOG algorithm is a widely known algorithm, and its detailed description is omitted here.

A method for tracking a person according to the present invention will be described with reference to FIG. 2.

First, a person present in a given image is detected, and a cluster in which the detected person is located is returned. In this case, using the rHOG algorithm, a person present in the image can be identified by scanning the image. At this time, the rHOG algorithm continuously scans the image, and when a person is detected, it starts generating a cluster of the detected person and the corresponding location.

)를 생성한다.That is, for each cluster, a cluster object represented by the following (Equation 1) (

).

(Equation 1)

은 캡쳐된 이미지 이름,

는 이미지의 캡쳐 시간,

는 사람이 검출된 위치,

는 클러스터의 중심(centroid)으로

으로 표현되며,

는 클러스터 구성원, 즉 검출된 사람의 구분자 모두이다.

는 클러스터의 모든 검출된 사람의 평균 위치를 나타낸다. 이후, 상기 클러스터 객체를 통하여 하기 (수학식 2)와 같이, 클러스터 오브젝터 목록(

)을 생성한다.here

Is the name of the captured image,

Is the capture time of the image,

Where the person was detected,

Is the centroid of the cluster

Is expressed as

Is all of the cluster members, i.

Represents the average location of all detected people in the cluster. Thereafter, through the cluster object, as shown in the following (Equation 2), the cluster object list (

).

(Equation 2)

은

번째 클러스터 객체이고,

은 자연수이다.here

silver

Is the first cluster object,

Is a natural number.

)마다 하기 (수학식 3)으로 표현되는 추적 객체(

)를 생성한다.Next, each cluster object (

) For each tracking object represented by the following (Equation 3) (

).

(Equation 3)

는 프레임이 기록된 시간들,

는 이에 대응한 이미지의 이름들,

은 이에 대응한 사람이 검출된 위치들,

는 해당 클러스터 객체 이름,

는 해당 추적장치의 위치이다.here

Is the times the frames were recorded,

Is the names of the corresponding images,

Is the locations where the corresponding person was detected,

Is the corresponding cluster object name,

Is the location of the tracking device.

The tracking object is introduced for each predefined cluster object. A cluster object composed of a name of an image or video, a location where people are detected, a time when a frame is recorded, a location of a tracking device, and a relationship with the tracking object may be a property of the tracking object.

)를 통해 하기 (수학식 4)와 같이 추적 오브젝터 목록(

)을 생성한다.The tracking object (

) Through the list of tracking objects as shown in (Equation 4) below (

).

(Equation 4)

은

번째 추적 객체이고,

은 자연수이다. 이후 추적 객체가 지속적으로 업데이트될 수 있다. 상기 수학식 4는 목록 형식으로 추가된 추적(Tracker) 객체를 보여준다.here

silver

Is the first tracking object,

Is a natural number. Thereafter, the tracking object may be continuously updated. Equation 4 shows a tracker object added in a list format.

)에서 다음 프레임의 동일한 영역 내의 추적 대상 객체(사람)를 선택하여, 상기 다음 프레임에서 선택된 추적 대상 객체의 위치를 추정한다. 즉, 다음 프레임의 동일한 영역 내에서, 추적 오브젝터 목록의 각 객체를 선택하여 추적 대상 객체의 위치를 계산한다. 이는 반복적으로 수행될 수 있다. Next, the tracking object list (

In ), a tracking target object (person) within the same area of the next frame is selected, and the location of the selected tracking target object in the next frame is estimated. That is, within the same area of the next frame, each object in the tracked object list is selected and the location of the tracked object is calculated. This can be done repeatedly.

)와 속력(

)(즉, 속도의 크기)는 하기 (수학식 5)를 통해 계산될 수 있다.The velocity for calculating the position of the object being tracked (

) And speed (

) (That is, the magnitude of the velocity) can be calculated through the following (Equation 5).

(Equation 5)

Then follow one of the following steps.

는 0이 된다. 그러나 이는 탐지되는 첫 번째 프레임의 경우에만 해당된다. a) If the object to be tracked is the same as one detected position

Becomes 0. However, this only applies to the first frame to be detected.

이다. 여기서

와

는 각각 거리와 시간의 차이를 나타낸다. b) When the object to be tracked has two detected locations (cluster size=2)

to be. here

Wow

Represents the difference between distance and time, respectively.

c) If the object to be tracked has more than 2 detection positions (cluster size>2), the average velocity is determined using a maximum of 10 (k=1,2,..n=10) previous detection positions. .

Next, the location of the object to be tracked is estimated and approximated in the same manner as in (Equation 6) below.

(Equation 6)

는 추적 대상 객체의 예상 위치(다음 프레임에서 추정되는 사람의 위치),

는 추적 대상 객체의 이전 위치(이전 프레임들에서의 사람의 위치)이다. 여기서

는 다음 프레임에서의 얼룩의 움직임을 추정하는데 사용된다.here

Is the expected position of the object to be tracked (the position of the person estimated in the next frame),

Is the previous position of the object to be tracked (the position of the person in the previous frames). here

Is used to estimate the motion of the blob in the next frame.

)을 하기 (수학식 7)과 같이 계산한다.Next, the search area of the next frame (

) Is calculated as the following (Equation 7).

(Equation 7)

(Pixel Window, 가로 및 세로 픽셀 크기)는

,

는 추적 객체의 위치변화이다. 본 발명에 따른 사람 추적 방법은 사람의 평균 너비를 나타내는 약 64 픽셀의 영역을 분석한다. 검출된 각 얼룩이 적어도 사람의 50 % 크기 이상인 것이 일반적이다. 따라서 이러한 정보를 사용하여 (수학식 7)과 같이 적절한 정사각형(square) 크기의 검색 영역

를 계산한다. 첫째로, 크기가

가 32 × 32 픽셀인 얼룩을 찾는다. 나머지 프레임들 내의 모든 후속 탐색들에 대해,

는 거리의 이전 궤적의 정보를 통해 갱신될 수 있다.here

(Pixel Window, horizontal and vertical pixel size)

,

Is the position change of the tracking object. The person tracking method according to the present invention analyzes an area of about 64 pixels representing an average width of a person. It is common for each spot detected to be at least 50% of the size of a person. Therefore, using this information, a search area of an appropriate square size, such as (Equation 7)

Calculate First, the size

Finds a blob of 32 × 32 pixels. For all subsequent searches in the remaining frames,

May be updated through information on the previous trajectory of the distance.

Next, a person present in the search area of the next frame is detected.

If a person is not detected in the search area, the object to be tracked is removed, and if a person is detected, the distance between the previous location and the current location is compared, and if the distance is less than 40 pixels, the center of the object to be tracked and the cluster And, if the distance is greater than 40 pixels, a new tracking object is created.

The tracked person object is updated with the person's new detected position in the frame, then repeats the previous step. For all other distances where the detected position is 40 pixels or more away from the existing tracking object, a new tracking object of the tracking module is created. Afterwards, the previous steps are repeated for the new tracking object. If the rHOG algorithm cannot detect the suitability of the existing tracking object up to 5 frames, the object is removed. The rest of the tracking object list is checked for suitability and goes through the previous steps when a non-empty list is found or the tracking process is stopped.

Meanwhile, in the present invention, a spot detection module may be additionally supplemented to improve performance. 3 shows a process of detecting spots according to the present invention.

This has a significant advantage in that only an image area with a higher possibility of movement can be found. The process of detecting a person through the spot detection module is performed in the following steps.

First, a motion map image is obtained through continuous frames as shown in Equation 8 below. 3A is an original image and (b) is a motion map image.

(Equation 8)

는 모션맵 이미지,

,

는 연속된 프레임이다.here

Is the motion map image,

,

Is a continuous frame.

Thereafter, the motion map image is divided into sub-areas of 8 × 8 pixels, and pixel intensity values of the sub-areas are summed. The image through this process is shown in (c) of FIG. 3. Thereafter, the summed pixel intensity value and the threshold value are compared to finally determine whether or not the pixel is uneven. This can be expressed as (Equation 9).

(Equation 9)

는 상기 하위 영역들의 합산된 픽셀 강도 값,

는 임계값이다. 도 3의 (d)는 임계 값인

을 적용하고 모션맵 이미지 32 × 32 픽셀을 하위영역들로 나누어 검출한 경우를 도시한 것이고, 도 3의 (e)는 임계 값이 적용되고 작은 얼룩이 제거 된 후 분할된 이미지를 도시한 것이다.here

Is the summed pixel intensity value of the sub-areas,

Is the threshold. Figure 3 (d) is the threshold value

Is applied and the motion map image is detected by dividing 32 × 32 pixels into sub-areas, and FIG. 3(e) shows the divided image after a threshold value is applied and small spots are removed.

4 is a test sequence for testing the performance of the present invention. 4A shows sequence 1, (b) shows sequence 2, (c) shows sequence 3, and (d) shows sequence 4.

The floor that limits the working area has special characteristics, but it does not affect the movement of people, and the color of people's clothes is not limited. Images (640 x 480 pixels) were captured on a 5m high PTZ camera at a rate of 20 frames per second. The specifications of the PTZ 212 camera are Fujinon lens, F1.8, fixed aperture, vertical viewing angle: 35° to 105° and horizontal viewing angle: 44° to 140°.

In order to evaluate the performance of the algorithm, the four test sequences shown in FIG. 4 were used to compare with the existing algorithm. For comparison, each test sequence was automatically inspected manually, i.e. using ground truth, and compared based on people's trajectory/path in the test video. The last test sequence shown in (d) of FIG. 4 is relatively complex compared to other sequences, because there is a lot of zigzag movement of a person while performing welding in an automobile assembly.

5 is a graph showing accuracy according to distances of all detected locations.

The x-axis represents the distances of all detected locations, and the y-axis represents the detection speed or accuracy. Referring to FIG. 5, it can be seen that the method for tracking a person according to the present invention exhibits a much higher hit rate compared to the existing algorithm. For example, when the distance is 20 pixels, the present invention has an accuracy of 99%, but the MS algorithm has an accuracy of 64%. Hereinafter, for each sequence, the tracking method of the present invention and the MS algorithm are compared in detail.

6 is a comparison between the tracking method of the present invention and the MS algorithm in Sequence 1 of FIG. 4. 6A and 6B are performed through the tracking method according to the present invention, and FIGS. 6C and 6D are results performed through the MS algorithm.

The start time jumps for the video test set are 0.5, 0.55, 0.8, and 0.65 seconds. In the missed frame, the MS algorithm can detect the user's location. 6B shows this time difference. During the test sequence, it was observed that if a person starts working and starts moving his hand several times, the MS algorithm shifts the center of the blob from the person to the elbow.

This leads to so-called "undetected positions". The weakness of the MS algorithm is that it is completely undetectable by the user in a few frames. In fact, this phenomenon was observed in the middle of the test sequence. In other words, it can be seen that a weakness appears in the detection of the location of the actual observation data.

On the other hand, the results performed through the tracking method according to the present invention are shown in (c) and (d) of FIG. 6. Referring to the numerical values shown in (c) and (d) of FIG. 6, it can be seen that the present invention can detect a person in all consecutive frames. This means that the tracking method according to the present invention has improved performance compared to the MS algorithm. Specifically, the accuracy (97%) was significantly improved than the accuracy (55%) shown through the MS algorithm. The deviation was also confirmed to have a lower deviation of 7 pixels compared to the 13 pixels shown in the MS algorithm.

7 is a comparison between the tracking method of the present invention and the MS algorithm in Sequence 2 of FIG. 4. 7A and 7B are performed through the tracking method according to the present invention, and FIGS. 7C and 7D are results performed through the MS algorithm.

In test sequence 2, it can be seen that the MS algorithm missed many locations in the video sequence. In addition, the detected stains were found to be very far away from the actual data, and were very small in size. This is shown in Figures 7(a) and (b). The size of the stain identified by the MS algorithm was 8 Х 20 pixels, and because it occupied only a small part of the human body, it can be seen that the stain does not correspond to the actual position of the person.

On the other hand, since the tracking method according to the present invention is based on the assumption representing a person of 32 占32 pixels or more, the influence of small spots on the body of a person can be ignored. Thus, all blobs below the 32Х32 pixel threshold are deleted. Referring to FIGS. 7C and 7D, it can be seen that the tracking method according to the present invention exceeds the performance of the MS algorithm in accuracy as well as the detected position. In the case of the MS algorithm, the detection speed was 35.8% within an average distance of 13 pixels, but the tracking method according to the present invention was found to be 100% at an average distance of 6 pixels.

8 is a comparison of the tracking method of the present invention and the MS algorithm in Sequence 3 of FIG. 4. 8(a) and (b) are performed through the tracking method according to the present invention, and FIGS. 8(c) and (d) are results performed through the MS algorithm.

8A and 8B show test sequence 3 for the MS algorithm. Referring to this, it can be seen that the detected stain is not located on a person's body, but is mostly located on a person's shoulder or arm, and has a small size. Most of the blobs in these images are not detected because of their size and location. This is shown in Figure 8 (a). Figure 8(b) shows the results observed through the time interval of 20 scans after 191 frames.

Compared with this MS algorithm, the tracking method according to the present invention shows a much improved detection result. The tracking method according to the present invention starts detection immediately from the beginning, and provides an improved detection function in terms of actual measurement data on the location of a target to be detected. The tracking method according to the present invention handles a large time gap such as the aforementioned 20 seconds by continuing tracking for a time interval. Compared with the MS algorithm, which does not detect a person in a stationary state, the present invention was able to detect a person continuously even though the person was in a static state. These detection results are shown in (c) and (d) of FIG. 8. Referring to this, it can be confirmed that any point in the graph is detected close to the actual measurement.

Table 1 below compares the results obtained in Sequences 1, 2, and 3 described above. Here, HR stands for accuracy and MSE stands for square root error.

(Table 1)

As a result, the present invention showed an average detection rate of 99.08%, which was higher than all other algorithms. The highest accuracy among other algorithms except for the present invention was 48.38% of the MS tracking method.

9 is a comparison of the tracking method of the present invention and the MS algorithm in Sequence 4 of FIG. 4.

Test sequence number 4 has a complex characteristic due to the sudden and dynamic change of the scenario and the difference in data within the frame. The sequence consists of 1021 frames with observed data gaps. In this frame, the person was held at a radius of 40 pixels in one direction, except for one long jump of about 8 seconds. Over time, a person's position change is observed in the same horizontal direction as the original position, which is about 100 pixels.

The total length of the video is 76 seconds, and nearly 30 seconds of data are missing. The missing 30 seconds consist of a total of 730 frames. The complexity of the environment arises because there are wires, machinery and sparks from welding. People are hidden in this environment.

The tracking method according to the present invention is compared with the other three algorithms. The trajectory of the detected position of the person can be confirmed in FIG. 9. Referring to (a) of FIG. 9, it is observed that some are located above the threshold, and there is a distance from the actual observation. This is because the sparks from the welding machine create falsely perceived spots.

9D shows actual observation data on the Y-axis versus the time on the x-axis to indicate the time interval observed in the sequence. 9C and 9B show trajectories of spot locations detected using a tracking method according to the present invention and an MS tracking algorithm for time and actual observation data, respectively.

The result of this graph shows the performance in terms of detecting the position of a person with a time interval in a video sequence frame, and is shown in Table 2.

(Table 2)

Test sequence 4 is the most complex sequence due to crossover and other disturbing factors occurring in an environment such as welding, but the tracking method according to the present invention exhibited higher accuracy than other detection algorithms. The highest accuracy found in other algorithms other than the present invention was only 31.2%, and it was found that there is a significant difference from the present invention, which showed an accuracy of 97.8%.

The present invention discloses an effective method of tracking people through an overhead camera. The present invention can effectively detect and track a stationary person even if the time interval is large.

Although the exemplary embodiments of the present invention have been described in detail above, those of ordinary skill in the art to which the present invention pertains will understand that various modifications may be made to the above-described embodiments without departing from the scope of the present invention. . Therefore, the scope of the present invention is limited to the described embodiments and should not be determined, and should not be defined by the claims to be described later, but also by those equivalents to the claims.

Claims (6)

In a person tracking method in an image captured by a camera in a person tracking device,
(A) A cluster object expressed by (Equation 1) for the cluster at the location where a person is detected in the image (

Generating );
(Equation 1)

(here

Is the name of the captured image,

Is the capture time of the image,

Where the person was detected, the center of the cluster

Is the average of the locations where a person was detected,

Is the delimiter of all detected persons)
(B) Cluster object list expressed by (Equation 2) (

Generating );
(Equation 2)

(here

silver

Cluster object (

)ego,

Is a natural number)
(C) Each cluster object (

Objects to be tracked (Equation 3) for each)

Generating );
(Equation 3)

(Here the frames were recorded

Corresponding to,

Is the names of the images,

Is the locations where a person is detected as the object to be tracked,

Is the cluster object name,

Is the location of the tracking device)
(D) List of tracking objects expressed by (Equation 4) (

Generating );
(Equation 4)

(here

silver

Object to be tracked (

)ego,

Is a natural number)
(E) List of tracking objects above (

), selecting a tracking target object in the same area of the next frame and estimating a position in the next frame based on the speed and speed of the tracking target object; And
(F) The search area of the next frame (

) As follows (Equation 5) to calculate the search area (

) Determining whether or not a person is detected,
(Equation 5)

(here

Is the predetermined horizontal and vertical pixel size,

Is the change in the location of the object to be tracked)
The (F) step,
Obtaining a motion map image through (Equation 6);
(Equation 6)

(here

Is the motion map image,

,

Are images of consecutive frames)
Dividing the motion map image into lower regions of a predetermined number of pixels and summing pixel intensity values of each of the lower regions; And
Including the step of detecting a person by removing the influence of the blob by using the determination of whether or not blob for each of the lower regions through (Equation 7),
(Equation 7)

(here

Is the summed pixel intensity value of each of the sub-areas,

Is the threshold),
In step (F), (a) if a person is not detected, the object to be tracked is removed, and (b) if a person is detected, the distance between the previous location and the current location is compared, (b-1 ) If the distance is less than or equal to a critical distance, the tracking target object and the center of the cluster are updated, and (b-2) if the distance is greater than the critical distance, a new tracking target object is generated. delete delete delete delete delete

Images