CN100435555C - High Speed Target Tracking Method and Its Circuit System - Google Patents

Abstract

The invention belongs to the field of image information processing, and relates to a high-speed target tracking method and a circuit system thereof. The process of high-speed target tracking is to first specify the target to be tracked by manual or automatic method. Secondly, the threshold window of the corresponding regional feature is set according to the regional feature of the target, and the input image is binarized with this threshold window. Then, the target is tracked by processing each frame of binarized image. When the target is found to be lost, the target loss processing is performed; when the target is found to be out of the camera's field of view, the camera is turned to aim at the target, and the field of view is rotated for post-processing. Finally, output the position of the object in each frame of the image. The invention has the characteristics of simple calculation, high speed and parallel processing. While ensuring high-speed processing, it takes into account the complexity of the target and the background, and is especially suitable for tracking high-speed moving targets. The present invention provides a structural diagram of an embodiment of a hardware implementation.

Description

High Speed Target Tracking Method and Its Circuit System

technical field

The invention relates to the technical field of image information processing, in particular to a target tracking method and a circuit system thereof.

Background technique

Real-time tracking of moving objects is an important topic in the field of image information processing. Template matching and background subtraction are two common information processing methods for extracting and tracking moving objects. Template matching is a method of using the existing target feature data as a template to search for matching points in the image containing the target to find the target. This is a target retrieval and tracking method based on the morphological characteristics of the target. The template matching method has high sensitivity and reliability when applied to tracking, and has been widely used. However, due to the large amount of computation in the target retrieval and tracking information processing of the template matching method, the system hardware circuit scale is large and the real-time processing of information is difficult. In addition, the image features of the tracked target generally change with time, so the fixed template matching method will not be applicable.

The background subtraction technique is a method to obtain the target by taking the difference between the current image and the existing background. Compared with the template matching method, this method is simpler and faster. But the difficulty of background subtraction methods lies in dealing with various disturbances and changes existing in the background. Moreover, the general background image is obtained by shooting a certain scene in advance. If the tracked target is moving in a continuous and complex large scene, it is not advisable to store the background information in advance.

With the improvement of real-time requirements, the processing speed of both template matching and background subtraction will be too slow. Recently, people have proposed the concept of a system chip that integrates image information processing functions and photosensitive unit arrays. This kind of system chip is generally called a vision chip (visionchip). Due to the limitation of chip area and power consumption, the signal processing method implemented on this system chip cannot be too complicated. Therefore, it is very important to develop a simple and effective high-speed moving target tracking method and its circuit system. Recently, the University of Tokyo proposed an information processing method for high-speed target tracking, and developed a corresponding digital information processing vision chip. The vision chip can complete the information processing of one frame of image within 1ms, and can effectively track the moving target under limited conditions. The basic process is: first select a brightness value as a critical point, and binarize the image by comparing the brightness value of the pixel point of the image with it. Then get the target in the first image. Next, a self-window capture method is used to realize target tracking, and it is judged whether it collides with other objects in the background during the tracking process. If a collision occurs, record the area of the target at this time, and then wait for the target to separate from the background object. After separation, the target was reconfirmed by comparison with the previously recorded target area, and then continued tracking by the self-window capture method. The self-window capture method assumes that the projection of the moving target on the photosensitive unit array does not exceed one pixel displacement between two consecutive frames, and expands the periphery of the target in the previous frame image by one pixel, and then passes through it as a self-window. Perform 'AND' logic operation with the image of the next frame to obtain the target in the image of the next frame. The process of the method is simple, adopts a simple parallel operation mode, and is suitable for the realization of a vision system chip. However, this method requires a significant difference between the brightness of the target and the brightness of the background, so it is not suitable for complex backgrounds. In the process of information processing, the target area information is used, so it is not suitable for non-rigid body targets. There is room for improvement in the applicability and reliability of tracking. In addition, addition operations are used to determine the details of the collision and separation between the target and the background object, which requires a large-scale logic circuit.

Contents of the invention

The invention proposes an information processing method for high-speed moving target tracking with simple flow, small calculation amount and parallel operation. While ensuring the reliability of target tracking, the processing speed is greatly improved, and the scale of the circuit is reduced at the same time. It can realize real-time tracking of high-speed moving targets under complex target and background conditions.

Technical scheme of the present invention is described as follows:

Firstly, images are collected in real time by an image acquisition device at a certain period.

Second, confirm the target to be tracked in the captured image.

Then, with the brightness of the target as a reference, the brightness threshold window for image binarization is set. (It can also be a window based on other area information such as color, texture, etc. Here, only brightness is used as an example.)

Then, binarize each pixel of a frame of image collected, and start to track the target in the fixed field of view where the camera does not move, that is, extract the target and its position in the image. If it is found that the target will leave the field of view, use the current target position to rotate the camera to re-align it with the target, and perform field of view rotation processing; if the target is lost, perform target loss processing.

Finally, after the field of view rotation processing or target loss processing is completed, the target is still tracked in a fixed field of view where the camera does not move.

The description of the technical solution is as follows:

1) Image acquisition

The method is to use a CMOS or CCD photosensitive unit array to collect grayscale or color images at a certain period;

2) Target designation

The method is to obtain the target to be tracked manually or automatically before starting to track;

3) Binarization threshold window specification

The method is that one or more regional features of the target need to be selected, and a threshold window for image binarization is generated accordingly. Taking brightness as an example, the brightness value of any point is taken from the specified target, or the average or median value of the target brightness is taken Recorded as I ₀ , the brightness window is set as {I ₀ -I, I ₀ +I}, I is the half window size, it depends on the situation, and I ₀ can be used according to the target tracking process at the beginning of the fixed field of view each time. The luminance characteristic is determined, and may not change after the first determination.

4) Image binarization

The method is to convert the image into a binary image through a threshold window. Taking the brightness as an example, the pixels whose brightness values are within the window are recorded as 1 or 0, and those located outside the window are recorded as 0 or 1.

5) Target tracking when the field of view is fixed

The method is to keep the camera still and track the target's motion characteristics as the main criterion; at the initial frame of tracking, the target on the binary image has been obtained, or can be obtained through target search. The target region is then subtracted from the binary image and the remaining region is stored as the background. In each frame of the tracking process, when the image binarization is completed, the processing steps are: (1) use the stored background to perform background subtraction on the newly obtained binarized image; (2) perform background subtraction on the background subtracted image The image is cleaned up, the purpose of which is to eliminate the interference caused by some subtle changes in the background; (3) then use a method similar to the self-window capture method proposed by the University of Tokyo to perform target capture on the image after image cleaning; (4) There are three situations for judging the target after capture: (a) if the target is lost, the target tracking process in this fixed field of view will be ended, and the target loss processing will be carried out; (b) the target will leave the field of view, then the fixed field of view will be ended. During the target tracking process in the field, the camera is notified to rotate, and the field of view is rotated; (c) if the above two situations do not occur, the target position is output, and the following target tracking process in the fixed field of view is continued. (5) The next step is performed according to the value of a state flag state, state=0 represents that no target separation event has occurred, and state=1 represents that a target separation event has just occurred. (a) state=0, judging whether a target separation event occurs in the current frame. If there is no target separation event, the current frame is processed, and the next frame is started to be processed; if a target separation event occurs, any part of the target after separation is stored, and state=1 is set, and the next frame is started to be processed; (b) state=1, first use the self-window method to capture part of the previously stored target in the current frame, and get the other part from the whole target minus the captured part. Then judge the motion characteristics of the two parts of the separated target: (i) when any part meets the confirmation condition of the moving target, this part is confirmed as the target, and whether the other part satisfies the background condition is judged , if satisfied, add it to the background, and finally set state=0 to start processing the next frame. (ii) If the two parts do not satisfy the confirmation condition of being a moving target, then end the current frame and start processing the next frame. (iii) If it is found that any part of the target has completely disappeared, then confirm the remaining part as the target, let state=0, and start the next frame.

The following is a specific description of the main operations in the above-mentioned target tracking method when the field of view is fixed:

Target seek operation.

The process is: the first step is to use a certain point as the seed; the second step is to expand from the seed until it intersects with the original binary image or exceeds a limited search radius; the third step is if there is an intersection then Do region growth on the intersection area and confirm the obtained area as the target, and the search ends. If the limited search radius is exceeded, the search will fail.

Image Cleanup Operations

Due to the subtle changes in the background, after background subtraction, some fine backgrounds will remain in the non-target area. Image cleaning is to remove these fine backgrounds. The starting point of this operation is to use the significant difference in size between the target and these fine backgrounds to remove the fine background. Image cleaning can also remove the effects of image noise, but not the large new background regions that result from large changes in the background.

class self window snapping operation

The process is the same as the self-window capture proposed by the University of Tokyo, except that the movement of the target between two frames can exceed one pixel. After the AND operation of the self-window and the image is completed, the region is grown to obtain the complete target.

target separation event

There are two possibilities for separation. One is that the target itself is separated; the other is that the target overlaps with the newly added large area in the background and then separates. Every time a target separation event occurs, the target must be reconfirmed.

Target Separation Event Judgment Operation

It is required to judge whether the target is separated into two parts and get one of them. Its process is: first, take a point randomly on the target. Second, region growing is performed using this point as a seed. Then, after the region growing is completed, the region obtained by region growth is subtracted from the original target region. Finally, it is judged whether the remaining area is an empty area, if it is, it is not separated, otherwise the target is separated, and the part of the separated target is obtained by growing the area before.

Region growing operation

The growth area is denoted as K, and the growth reference image is denoted as T. The process is: first step, expand the boundary of K by one pixel to get K'. The second step is to find the intersection K" of K' and T. The third step is to compare K and K", if K" is more than K, let K=K", and repeat the first step. Otherwise the growing process ends. If the direction of growth is restricted, the process is altered.

Motion Feature Judgment Operation

The condition that an area belongs to the moving target or the condition that belongs to the background can use the change of the boundary position of the area, the movement of the center position or the center of gravity position as the basis for setting.

6) Target position output

The method is to obtain a reference point that can reflect the position of the target, which can be the center of gravity of the target, the center, or even any point on the target.

7) Field rotation processing

Do some processing to keep tracking continuous after the camera turns. The method is that when the camera rotates, based on the position of the target before the rotation, the camera is rotated so that the position of the target is at the center of the field of view. The subsequent processing includes moving the target area in the target image before rotation to the center of the image. Considering the error in the rotation process of the camera, the target area is enlarged by several pixels. After the camera rotation is completed, the obtained first frame binary image and the target image are intersected to obtain the rotated initial target, and then a new round of target tracking process is started when the field of view is fixed.

8) Mark loss processing

This feature attempts to retrieve a target after it has been lost. There are two common cases of target loss. One is that the target enters a large background with similar brightness to the target, and the other is that it is blocked by a foreground with different brightness. This can be obtained by judging whether there is an object of similar brightness near the target position at the moment before the target is lost on the original binarized image. For the first case, the target search is carried out immediately on the original binarized image with the last position before the loss of the target as the search seed. For the second case, after waiting for a certain amount of time, start the target search on the original binarized image with the last position before the target is lost as the search seed. If it is found, it will start a new round of target tracking process when the field of view is fixed. If the search fails, it will search again in the next frame.

The high-speed target tracking method, this high-speed target tracking method based on binarized images, (1) the target can be manually specified or automatically acquired, and the target object can be a non-rigid body, requiring it to move at a high speed relative to the background;

(2) Binarize the input image according to the regional features of brightness, color, or texture;

(3) The calculation of the tracking process is carried out on the binary image, and the tracking process is composed of three aspects: tracking when the field of view is fixed, processing of field of view rotation, and target loss processing;

(4) The output target position is the coordinates of a point on the image that can represent the position of the target on the image.

The binarization requires a feature threshold window to distinguish the corresponding features on the image that fall within the window and the area outside the window. The window is composed of the window midpoint and half the window width. The selection of the window midpoint depends on the corresponding area of the target. Features, taking the brightness as an example, the average value of the target brightness can be selected as the window midpoint.

Operations on binary images are completed through logical operations such as AND, OR, and NOT, excluding arithmetic operations such as addition, subtraction, multiplication, and division.

Operations on binary images are performed in parallel.

Binarization is based on the brightness, color, or texture features of the region. During the tracking process, only the movement of the binarized region is concerned, so the image does not require high resolution.

The tracking process when the field of view is fixed includes target search, background storage, background subtraction, image cleaning after background subtraction, target capture, target loss judgment, target separation event judgment, target reconfirmation after target separation, and update background operations.

The field of view rotation condition occurs when the target boundary reaches the field of view boundary. The camera rotates according to the target position at that time. After the rotation, the target position is at the center of the field of view. After the rotation, the enlarged area of the previously stored target area is used as the window to capture the target at the center of the field of view. .

After the target is lost, according to the judgment of whether it is mixed into the background area or blocked by the foreground, the target is searched near the last position before the target is lost at different times.

The tracking process restarts when the field of view is fixed after the field of view rotation processing and target loss processing are complete.

The present invention has following characteristics and effects:

1. All image processing processes are carried out entirely on binarized image data, and the entire operation can basically be completed only by using basic logic operations such as AND, OR, and NOT. Therefore, both the processing process and the calculation process are very simple, and the processing speed is very high.

2. Because only the regional features of the target and the motion features of the target are used, the image is not required to have a high resolution, even a low resolution like 64×64 is fine.

3. The calculation process has two-dimensional parallelism, which is suitable for parallel processing systems.

4. Suitable for on-chip integration to complete real-time tracking of high-speed moving targets.

5. The whole method is relatively complete, suitable for backgrounds with a certain degree of complexity, and the target can also be a non-rigid body, so the method of the present invention has better applicability and reliability.

6. Regarding the consideration of camera rotation, it also solves the problem that the tracking based on the background subtraction method is limited to a fixed scene.

Description of drawings

Fig. 1 is a flow chart of the high-speed target tracking method of the present invention.

Figure 2 is a flow chart of the target tracking process when the field of view is fixed.

FIG. 3 is a flowchart of field of view rotation processing.

FIG. 4 is a flowchart of target loss processing.

Fig. 5 is a system diagram of the high-speed target tracking circuit of the present invention.

Detailed ways

A specific embodiment of the present invention-a high-speed moving target tracking system will be described in detail below in conjunction with the accompanying drawings.

1. As shown in Figure 1, the method flow of the high-speed moving target tracking system has 1) image acquisition, 2) target designation, 3) binarization threshold window designation, 4) image binarization, and 5) target when the field of view is fixed There are eight steps of tracking, 6) target position output, 7) field of view rotation processing and 8) target loss processing. Its working process is as follows:

The image acquisition function continuously acquires grayscale images with a certain period.

Before tracking starts: the first step is to specify the target to be tracked. Can be specified manually or automatically. A simple manual designation method is: manually point the camera at the target so that the midpoint of the field of view is on the target. The second step is to set the brightness threshold window for binarization. The brightness value of a pixel on a point in the image is taken as I ₀ , and the brightness window is defined as {I ₀ -I, I ₀ +I}.

Start tracking right away after completing the steps above. First, the grayscale image is binarized. Then start the target tracking process when the field of view is fixed, and output the coordinates of the point that can represent the position of the target on the image every time the signal processing of a frame of image is completed, and start the target tracking process when the field of view is fixed for the next frame of image. If the target is found to be lost during the process, the target loss processing will be performed. After the target is found, the target tracking process will restart when the field of view is fixed in the next frame. If the target is found to be out of the camera's field of view during the process, the camera will be notified to rotate so that the target is located in the center of the field of view, and the field of view will be rotated. After the completion of the next frame, the target tracking process will restart when the field of view is fixed.

2. When the field of view is fixed as shown in Figure 2, the processing flow of the target tracking process for each frame of binary image information is as follows:

(A) The first frame of image processing needs to go through the process of: target search→obtain background→end every time the fixed field of view target tracking process is started (or restarted). The system processes information in two situations according to the past processing status. The first case is that the system has just started the tracking process, we define it as start=1; the second case is restarting the tracking process after the field of view rotation processing or target loss processing, we define it as start=0. When start=1, the system first obtains the target through target search, and then obtains the background information. When start=0, the target area has been obtained in the previous processing, so it is enough to directly obtain the background.

(B) The processing of a certain frame image in the tracking process when the field of view is fixed, it needs to go through: background subtraction → image cleaning → target capture → target loss judgment → field of view rotation judgment → position extraction → state value judgment process. When it is judged that the target is lost or the field of view is rotated and the target tracking process leaves the fixed field of view, the corresponding target loss processing or field of view rotation processing is performed. (1) When state=0, the system determines that there is no target in the separation state in the previous frame image, and at the same time judges whether the target is separated in the current frame. If no target separation is found, the processing of the current frame ends directly; if the target separation is found, a part of the target after separation is stored first, and after state=1 is defined, the processing of the current frame ends. (2) When state=1, the system determines that there is a separate target in the previous frame image, then reconfirms a certain part of the separated target as the target to continue tracking, and considers whether to update the background. The first thing to do is to update the two separated parts of the target, and then determine whether a part has disappeared in the current frame. If one part disappears, the other part is recognized as the target, and the processing of this frame ends. If both parts are there, it is judged whether they meet the motion target condition. If none are satisfied, the processing of the current frame ends, but the target cannot be confirmed yet. If a certain part satisfies the moving target condition, this part is confirmed as the target, and the other part is checked at the same time, if it is confirmed to be the background, the background is updated, and the processing of the current frame ends. In the above process, once the target is confirmed, state=0 must be set.

The specific implementation of some steps and calculation process in the target tracking process when the field of view is fixed is as follows:

否则

Definition If a binary image X, X(i, j) is all 0, that is, all X(i, j) are equal to or equal to 0, then

otherwise

●Expand a certain area X by one pixel

● Defined as expand(X), it implements:

X(i,j)=X(i+1,j)/X(i-1,j)/X(i,j+1)/X(i,j-1)/X(i,j).

●Regional growth

A＝C1，回到(1)；如果

结束。Let A be the area to be grown, and B be the growth reference area, which is defined as grow(A, B). (1) C1=expand(A); (2) C1=C1 &B; (3) C2=C1 &~A; (4) if

A＝C1, back to (1); if

Finish.

●Target search

C＝grow(C，B)，C是找到的目标，结束；或者r＞R(指定的搜索半径)，结束。否则回到(2)。Let A be the seed, and B be the image to be searched. (1) Let A(x, y)=1, other points are 0, r=0; (2) A=expand(A); (3) C=A & B, r=r+1; (4) if

C=grow(C, B), C is the found target, end; or r>R (specified search radius), end. Otherwise go back to (2).

● get background

Let T be the target and G be the original binary image. Background BG = ~T & G.

●Background subtraction

Let BG be the background, and G be the binary image to be subtracted. Then do the operation G=~BG & G.

●Image cleaning

There are many ways to make corresponding calculations based on the fact that there is a significant difference between the residual fine background and the target size in the binary image after background subtraction. An example is as follows: Let X be the image to be cleaned. (1)X(i,j)=(X(i+1,j)/X(i-1,j))& X(i,j) (2)X(i,j)=(X(i , j+1)/X(i,j-1))& X(i,j) (3)X(i,j)=(X(i+1,j)/X(i-1,j) ) & X(i, j).

●Target acquisition

Let T0 be the target stored in the previous frame, G be the image to be captured, and T be the captured target. (1) T = expand(T0) (2) T = T & G (3) T = grow (T, G).

●Target lost

Let T be the target image, then

● Field of view rotation

When the boundary of the target is found to reach the boundary of the field of view, it needs to be rotated.

●Location Extraction

Can extract the center of gravity of the target, the center or any point representing the position of the target.

●Target separation judgment

则没有分离；否则分离，且A是分离后的一部分。Let T be the target image. T(x,y)=1 is any point on the target. (1) Let A(x,y)=1, and other pixels are all zero. (2) A=grow(A, T). (3) A=~A & T. (4) if

Then there is no separation; otherwise separation, and A is part of the separation.

●Judgment is the condition of moving target

Counting from the moment of separation, a certain part of the separated target has moved a distance of L pixels, and the difference between the moving direction and the moving direction before the separation of the target is smaller than that of the other part.

●Conditions for judging that it is a background object

Counting from the moment of separation to re-identification of the object, the boundary or center of the part not identified as the object moves by less than M pixels.

●Add area X to background BG

BG=BG|X.

3. Regarding the rotation of the field of view and the processing of the rotation of the field of view, as shown in Figure 3, the system calculates the angle and direction of the camera rotation according to the position of the target in the field of view before the rotation, so that the target is located in the center of the field of view after the rotation, and the system performs The translation processing of the target area in the target image T stored before the camera turns makes the position of the target area at the center of the image. Considering the error in the rotation process of the camera, execute T=expand(T) several times, so that the self-window used to capture the target after rotation is actually obtained. On the first frame of binary image G obtained after the camera rotates, T=T & G is processed. Then, restart the tracking process while fixing the field of view.

是前景遮挡，否则是混入相似亮度背景。对前者，等待若干时间后以p的位置为种子在新获得的二值图像上搜索目标。对后者，以同样方式立刻搜索。找到目标则处理过程完成，否则在接着的帧重新查找。完成后，跟踪继续，但重新开始固定视场时跟踪过程。4. About the target loss processing, as shown in Figure 4, the system first judges whether it is blocked by the foreground or mixed into the background area with similar brightness. Suppose the last position of the target on the image before the target is lost is p, and the original binary image G is obtained after the target is lost. Create a binary image P, let the point p be 1, and the rest of the points be 0. Execute P=expand(P) several times, then do P=P & G. if

It is foreground occlusion, otherwise it is mixed with background of similar brightness. For the former, after waiting for some time, use the position of p as the seed to search for the target on the newly obtained binary image. For the latter, search immediately in the same way. If the target is found, the processing is complete, otherwise, search again in the next frame. When complete, tracking continues, but restarts the tracking process when the field of view is fixed.

5. Embodiments of the present invention As shown in FIG. 5(1), the method of the present invention can be implemented through a parallel hardware structure. Including image acquisition and binary image processing module, control module, position output module, camera interface four parts. Wherein, the control module is interconnected with the image acquisition and binarized image processing module and the camera interface, and the position output module is respectively connected with the camera interface, image acquisition and binarized image processing module.

The image acquisition and binarization image processing module is an M×N pixel unit array, which has both image acquisition function and binary image processing function. As shown in FIG. 5(2), each pixel unit includes a photosensitive unit, a binarization processing unit, a storage unit, an operation unit and connection lines with other nearest neighbor pixel units. The photosensitive unit receives the light signal and converts it into an electrical signal, and obtains a grayscale image of a frame of M×N pixels after each acquisition cycle. The binarization processing unit in each pixel unit can convert the grayscale image into a binary image, and the binary image data of each pixel is stored in the storage unit of the pixel, and the storage unit is also used to store the intermediate results of the tracking process . The memory in a single pixel unit is n bits, and n can be n=1, 2, 3, . . . . The two-dimensional parallel processing structure composed of computing units in all pixel units completes single-pixel computing and neighborhood computing for binary images.

According to the high-speed target tracking method, the control module issues instructions according to the current situation of the tracking process to control the binary image processing module to perform corresponding processing to complete the tracking operation process.

The position output module is responsible for obtaining the position coordinates of a target and outputting it.

The camera interface is responsible for communicating with the mechanical part that controls the camera. It can notify the time, angle and direction of the camera's rotation, and can also know the state of the camera.

Claims (10)

1. A high-speed target tracking method, its composition steps have following eight: 1) Image acquisition; 2) target designation; 3) binarization threshold window designation; 4) Image binarization; 5) Target tracking when the field of view is fixed; 6) Target position output; 7) Field of view rotation processing; 8) Target loss processing; Wherein, the specific steps of the above-mentioned composition steps are as follows: 1) Image acquisition, using a CMOS or CCD photosensitive unit array to acquire grayscale or color images at a certain period; 2) Target designation, before starting to track, obtain the target to be tracked manually or automatically; 3) Specifying the binarization threshold window, one or more regional features of the target need to be selected, and a threshold window for image binarization is generated accordingly; the brightness value of any point is taken from the specified target, or the average or middle of the target brightness is taken The value is recorded as I ₀ , the brightness window is set as {I ₀ -I, I ₀ +I}, I is the half window size, it depends on the situation, and I ₀ can be used according to the current target when the target tracking process starts to fix the field of view each time. Determined by the brightness characteristics of , it can also not change after the first decision; 4) Image binarization, the image is converted into a binarized image through a threshold window; pixels whose brightness values are located in the window are recorded as 1 or 0, and those located outside the window are recorded as 0 or 1; 5) When the field of view is fixed, the target is tracked, and the camera is kept still, and the movement characteristics of the target are used as the main criterion for tracking; 6) Output the target position to obtain a reference point that can reflect the position of the target, which can be the center of gravity of the target, the center, or even any point on the target; 7) Field of view rotation processing, in order to maintain the continuity of tracking after the camera rotates, the method is that when the camera rotates, the camera is rotated based on the position of the target before the rotation so that the target position is in the center of the field of view, and the subsequent processing includes , move the target area in the target image before rotation to the center of the image, and expand the target area by several pixels in consideration of the error in the camera rotation process. Find the intersection, so that the initial target after rotation is obtained, and then start a new round of target tracking process when the field of view is fixed; 8) Target loss processing. There are two cases of target loss. One is that the target enters a large background with similar brightness to the target, and the other is that it is blocked by the foreground with different brightness. This can be determined by judging the target loss on the original binarized image. Whether there is an object of similar brightness near the target position at the previous moment is obtained. For the first case, start the target search immediately on the original binarized image with the last position before the target is lost as the search seed; for the second case, wait for a certain amount of time Then start the target search on the original binarized image with the last position before the target is lost as the search seed. If the target is found, a new round of target tracking process in a fixed field of view will start. If the search fails, search again in the next frame. 2. the high-speed target tracking method according to claim 1, is characterized in that following process: at first specify the target to be tracked by manual or automatic method, secondly, take the area feature of target as reference setting corresponding area feature threshold value window, use The threshold window binarizes the input image, and then tracks the target by processing each frame of the binarized image. When the target is found to be lost, perform target loss processing; when the target is found to be out of the camera field of view, turn the camera to align The target is processed after the field of view is rotated, and finally, the position of the target in each frame of image is output. 3. The high-speed target tracking method according to claim 1 or 2, characterized in that, said binarization requires a feature threshold window to distinguish the corresponding feature on the image falling in the area of the window and outside the window, the window consists of The midpoint of the window is composed of half the width of the window. The selection of the midpoint of the window depends on the corresponding regional characteristics of the target. When the characteristic of the region is brightness, the average value of the target brightness is selected as the midpoint of the window. 4. The high-speed target tracking method according to claim 1 or 2, characterized in that the operation on the binary image is completed by AND, OR, and non-logic operations. 5. The high-speed target tracking method according to claim 1 or 2, characterized in that the operation on the binary image is a parallel operation. 6. The high-speed target tracking method according to claim 1 or 2, wherein the binarization is based on the brightness, color or texture features of the region, and only cares about the motion of the region after binarization in the tracking process, so no Images are required to be high resolution. 7. The high-speed target tracking method according to claim 1 or 2, wherein the tracking process includes target search, background storage, background subtraction, image cleaning after background subtraction, target capture, and target loss judgment when the field of view is fixed , Target separation event judgment, target reconfirmation after target separation, update background operation. 8. The high-speed target tracking method according to claim 1 or 2, wherein the field of view rotation condition occurs when the target boundary reaches the field of view boundary, the camera rotates according to the target position at that time, and the target position is at the center of the field of view after the rotation , after turning, capture the target in the center of the field of view with the enlarged area of the previously stored target area as the window. 9. The high-speed target tracking method according to claim 1 or 2, characterized in that after the target is lost, the target is searched near the last position before the target is lost according to the judgment of whether it is mixed into the background area or blocked by the foreground at different times. 10. The high-speed target tracking method according to claim 1 or 2, characterized in that, after the field of view rotation processing and target loss processing are completed, the tracking process is restarted when the field of view is fixed.

Images