TW201501080A - Method and system for object detection and tracking - Google Patents

Abstract

A method and system for object detection and tracking are provided. In the present method, a frame in a video is regarded as a first frame, and a display area of an object in the first frame is detected by scanning the first frame. Next, the object is tracked according to a ratio of the display area to the first frame and at least one reference frame in the video, wherein the at least one reference frame follows the first frame. Besides, a scale and rotation invariant feature of the object being tracked is recorded. After that, whether the object appears in a specific frame is determined by virtue of the scale and rotation invariant feature of the object.

Description

Target detection and tracking method and system

The present invention relates to a method and system for detecting and tracking a target, and more particularly to a method and system for instantaneous detection and tracking of one or more targets in a complex environment.

In recent years, visual monitoring technology has been continuously refined and widely used in various fields. Take the Intelligent Transportation System (ITS) as an example. It uses visual monitoring technology for ranging and dynamic object detection to provide pre-warning drivers' improper driving behavior and adaptive speed control. Prevent accidents and ensure that driving speed does not exceed safe range.

The existing ranging technology is based on a hardware-based approach such as radar or laser. However, radar-based technology has accurate depth information and long detection distance, but the expensive hardware cost leads to its popularization. Not easy. The existing dynamic object detection technology is prone to high false positive rate in an environment with complex background. At present, most of the techniques for detecting and tracking objects need to achieve good results in a relatively simple background environment, and it must be costly in order to achieve a robust detection effect. A large amount of computing costs, it is difficult to take into account system performance and accuracy.

In view of this, the present invention provides a target detection and tracking method and system, which can provide a fast and stable target detection and tracking mechanism in various environments such as large difference in brightness, frequent dynamics, and high background complexity. .

The target detection and tracking method of the present invention comprises the steps of: first capturing a frame in the movie as the first frame, and scanning the first frame to detect the display area of the target in the first frame. Then, the object is tracked according to the area ratio of the display area occupying the first frame, and at least one reference frame following the first frame in the movie, and the anti-scale and anti-rotation features of the tracked object are recorded. Subsequently, the anti-scale and anti-rotation features are utilized to identify whether the object is present in a particular frame.

In an embodiment of the invention, the target detection and tracking method further comprises: obtaining a plurality of training positive samples and a plurality of training negative samples, extracting texture features of each training positive sample and each training negative sample, and using each training positive The sample and the texture features of each training negative sample, and a classifier fusion algorithm to train the target classifier.

In an embodiment of the invention, the step of scanning the first frame to detect the display area of the target in the first frame comprises: forming a cascade classifier with n target classifiers, wherein Is a positive integer. The first frame is scanned by n times using a cascaded classifier to detect the display area of the object in the first frame. Each of the scanning procedures includes: in the first frame Move the detection window in the direction to scan the first frame completely and then reduce the size of the first frame.

In an embodiment of the invention, the step of tracking the target according to the area ratio of the display area occupying the first frame and the at least one reference frame following the first frame in the movie includes: determining whether the area ratio exceeds default value. If the area ratio does not exceed the preset value, at least one feature point about the target object is extracted from the display area, and the target object is tracked by using the feature point extracted from the display area and the reference frame. If the area ratio exceeds the preset value, extracting at least one feature point about the target object from the plurality of sub-areas in the display area, and using the feature point extracted from the sub-area and the reference frame, Tracking.

In an embodiment of the invention, the sub-region is an upper left corner region and a lower right corner region of the display region.

In an embodiment of the invention, the sub-regions do not overlap or partially overlap each other.

In an embodiment of the invention, the reference frame comprises: a second frame adjacent to the first frame, and a third frame adjacent to the second frame. And according to the ratio of the area occupied by the display area to the first frame, and the reference frame following the first frame in the movie, the step of tracking the target includes: extracting at least one feature point about the target from the first frame, according to The Kalman filter filters the estimated range in the second frame and performs an optical flow algorithm to estimate the prediction of the above feature points from the first frame in the second frame. Fan Mobile information within the perimeter. Then, at least one feature point about the target object is taken from the second frame, the estimated range in the third frame is calculated according to the Kalman filter, and an optical flow tracking algorithm is performed to estimate the above-mentioned from the second frame. The movement information of the feature point within the estimated range of the third frame.

In an embodiment of the invention, the target detection and tracking method further comprises: using a Speeded Up Robust Features (SURF) algorithm to establish anti-scale and anti-rotation features.

In an embodiment of the invention, the specific frame is any frame that is temporally after all reference frames, and the step of using the anti-scaling and anti-rotation features to identify whether the object appears in a particular frame includes: from a specific frame 撷The feature to be identified is taken out, compared to the feature to be recognized and the anti-scale and anti-rotation features. When the feature to be identified conforms to the anti-scale and anti-rotation features, it is determined that the object appears in a particular frame.

The target detection and tracking system of the present invention comprises: a storage unit and a processing unit coupled to each other. The storage unit records a plurality of modules, and the processing unit accesses and executes the above modules in the storage unit. The above modules include: a detection module, a tracking module, and an identification module. The detection module captures a frame in the movie as the first frame, and scans the first frame to detect the display area of the target in the first frame. The tracking module tracks the target according to the area ratio of the display area occupying the first frame, and at least one reference frame following the first frame in the movie, and records the anti-scale and anti-rotation features of the tracked object. The recognition module utilizes anti-scale and anti-rotation features to identify whether the object appears in a particular frame.

Based on the above, the target detection and tracking method and system of the present invention utilizes Machine learning-based algorithms and cascaded architectures are used to instantly detect objects in complex dynamic scenes, and then use the optical flow tracking algorithm and the Kalman filter tracking mechanism to track the target. The target detection and tracking method and system of the present invention can be widely applied to various environments and provide fast and robust detection and tracking effects.

The above described features and advantages of the invention will be apparent from the following description.

100‧‧‧Target Detection and Tracking System

110‧‧‧ storage unit

111‧‧‧ training module

113‧‧‧Detection module

115‧‧‧Tracking module

117‧‧‧ Identification Module

120‧‧‧Processing unit

S210~S230‧‧‧ steps of the target detection and tracking method according to an embodiment of the present invention

310‧‧‧ training positive sample

320‧‧‧ training negative samples

330‧‧‧Character extraction unit

340‧‧‧ training unit

400‧‧‧ Cascade Classifier

C ₁ , C ₂ , C ₃ , C ₄ ‧ ‧ target classifier

F ₁ ‧‧‧first frame

NDR ₁ , NDR ₂ , NDR ₃ , NDR ₄ ‧‧‧ Non-target area

DR‧‧‧ display area

W‧‧‧Detection window

60‧‧‧ Targets

63‧‧‧Display area

63’‧‧‧Reduced area

65, 67‧‧‧ sub-areas

+‧‧‧Feature points

R ₁ , R ₂ , R ₃ ‧‧‧ Estimated range

O ₁ , O ₂ , O ₃ ‧‧‧ optical flow computing area

S810~S850‧‧‧ The steps of identifying whether the target object appears in a specific frame according to an embodiment of the present invention

FIG. 1 is a block diagram of a target detection and tracking system according to an embodiment of the invention.

2 is a flow chart of a method for detecting and tracking a target according to an embodiment of the invention.

FIG. 3 is a schematic diagram of a training target classifier according to an embodiment of the invention.

4 is a schematic diagram of a cascaded classifier according to an embodiment of the invention.

FIG. 5 is a schematic diagram of a cascaded classifier performing a plurality of scanning procedures according to an embodiment of the invention.

6A-6C are schematic diagrams showing feature points of a captured object according to several embodiments of the present invention.

FIG. 7 is a diagram of a target in an adjacent frame according to an embodiment of the invention. Schematic diagram of the object tracking.

FIG. 8 is a flow chart of identifying whether an object appears on a particular frame, according to an embodiment of the invention.

FIG. 1 is a block diagram of a target detection and tracking system according to an embodiment of the invention. Referring to FIG. 1, the target detection and tracking system 100 of the present embodiment can be implemented in a computer system, a workstation, a server, or any electronic device having computing and processing capabilities. The target detection and tracking system 100 includes: a storage unit 110 and a processing unit 120 coupled to each other, and its functions are as follows: the storage unit 110 is, for example, a random access memory (RAM), a read-only memory ( Read-Only Memory (ROM), Flash memory, hard disk or other similar device or a combination of these devices for recording a plurality of modules executable by the processing unit 120, which can be loaded The processing unit 120 performs a detection and tracking function of the target.

The processing unit 120 can be a central processing unit (CPU), or other programmable general purpose or special purpose microprocessor (Microprocessor), digital signal processor (DSP), programmable Controllers, Application Specific Integrated Circuits (ASICs), Programmable Logic Devices (PLDs), or other similar devices or combinations of these devices. The processing unit 120 can access and execute each module recorded in the storage unit 110 to make the target The detection and tracking system 100 performs instant detection and tracking of one or more objects.

The modules recorded in the storage unit 110 include: a training module 111, a detection module 113, a tracking module 115, and an identification module 117. The modules can be computer programs and implement objects under the execution of the processing unit 120. The detection and tracking function; the following embodiment illustrates the detailed operation of the target detection and tracking system 100.

2 is a flow chart of a method for detecting and tracking a target according to an embodiment of the present invention. Please refer to FIG. 1 and FIG. 2 at the same time. In this embodiment, the target detection and tracking system 100 can receive a movie taken by a webcam (or other film capture device) in various environments (eg, highways, suburbs, or urban streets, but not limited to). (video), and detect and track specific objects in the movie. The movie includes a plurality of static frames, and the target may be a vehicle, a human face, or other kinds of objects.

First, in step S210, the detecting module 113 captures a frame in the movie as the first frame, and scans the first frame to detect the display area of the target in the first frame.

In detail, in order to enable the detection module 113 to detect whether the target appears on a certain frame and the display area of the target in the frame, the training module 111 is used in this embodiment. A large number of training positive and negative samples will be obtained during the offline training phase, and a target classifier that can distinguish between the target and the background is trained through a machine learning method. FIG. 3 is a schematic diagram of a training target classifier according to an embodiment of the invention. As shown in FIG. 3, the training module 111 obtains a plurality of training positive samples 310 and a plurality of training negative samples 320. Wherein, the training positive sample 310 is a package A picture containing the same kind of object as the target, and the training negative sample 320 is a picture that does not contain the same kind of object as the target. The training module 111 utilizes the feature extraction unit 330 to retrieve the texture features of all the training positive samples 310 and the training negative samples 320, respectively. For example, the feature extraction unit 330 can quickly acquire the distinct texture features in each training positive sample 310 and the training negative sample 320 in combination with the integral image technique, such as a Haar-like feature. However, the invention is not limited to this. Then, the training module 111 performs a classifier fusion algorithm on the texture features of each training positive sample and each training negative sample by using the training unit 340, thereby training a target classifier. For example, the training unit 340 trains an adaptive Boosting algorithm to distinguish between a positive sample (ie, an object of the same kind as the target) and a negative sample (ie, a different type from the target). The object classifier of the object, which is composed of a plurality of weak classifiers, which are basic elements capable of distinguishing the object to be detected.

In the online detection phase, in order to balance the detection speed and accuracy, the detection module 113 forms a cascade classifier with n (n is a positive integer) target classifier, and utilizes a cascade classifier. The first frame is subjected to n scanning procedures to detect the display area of the object in the first frame. Since the cascaded classifier can eliminate the area of the first frame that cannot be the target as early as possible, the detection speed can be greatly improved. 4 is a schematic diagram of a cascaded classifier according to an embodiment of the invention. Referring to FIG. 4, in this example, the detection module 113 is assumed to be four target classifiers C ₁ to C _{4 .} A cascade type classifier 400 is formed. The first frame F ₁ passes through the layer detection of the target classifiers C ₁ to C ₄ , and the target classifier of each level excludes the non-target region that does not belong to the target from the first frame F ₁ (ie The non-target regions NDR ₁ to NDR ₄ ), and the regions passing through all the target classifiers C ₁ to C ₄ are the display regions DR of the target.

FIG. 5 is a schematic diagram of the cascaded classifier 400 performing four scanning procedures. Refer to FIG. 5, in order to identify the object in each scanning procedure, the cascade classifier 400 will move in the particular detection window in the direction F, the first frame of _a display area in _a first frame of the F W, in a full scan of the first frame F ₁ , the specific direction is, for example, scanning the first frame F ₁ from left to right and then from top to bottom. Subsequently, the cascade classifier 400 the first frame of reduced size F _1, then the next scan procedure, whereby the first can find Movies frame F _1, the object having a different size of the display area. The scanning method shown in FIG. 5 is also called an image pyramid type scanning.

As described above, the detection module 113 can detect the display area of all the objects belonging to a specific category in the first frame. However, for convenience of explanation, it is assumed that the first frame has only one target, so the detection The test module 113 detects only the display area of one object.

After the detection module 113 detects the display area of the target object, the tracking module 115 of the embodiment uses a tracking mechanism combining a Kalman filter and an optical flow algorithm. A fast and stable target tracking effect is achieved. In detail, as shown in step S220, the tracking module 115 tracks the target according to the area ratio of the display area occupying the first frame, and at least one reference frame following the first frame in the movie, and records the target being tracked. Resistance Zoom and anti-rotation features. In this embodiment, the tracking module 115 performs an optical flow tracking algorithm only on a certain portion (or portions) of the frame without tracking the entire frame. For example, after the detection module 113 detects the display area, the tracking module 115 uses only the optical flow tracking algorithm to calculate the display area (or one or more smaller areas in the display area). For the feature points extracted from parts of the current frame, in the next frame, find the moving position of the feature points to track the target. The following describes how the tracking module 115 captures feature points for the optical flow tracking algorithm in several examples.

In general, because the film captures the moving state of the target in the environment, the same target is not the same size in different frames. When the target to be tracked is very close to the webcam that shoots the movie, the proportion of the area occupied by the display area of the target in the frame is quite large. Since the tracking module 115 needs to perform arithmetic processing on a large area, it may result in a low computational efficiency. In order to increase the processing speed, the tracking module 115 uses only the information of several local regions of the target to perform tracking processing for the object that is too large in the frame, so that the operation time cost of the excessive target in the frame can be reduced.

Specifically, in this embodiment, the tracking module 115 determines whether the display area of the target occupies an area ratio of the first frame exceeds a preset value, and the preset value is related to the resolution of the movie and the size of the frame. The appropriate values are determined experimentally. If the area ratio does not exceed the preset value, it means that the target object is not too large compared to the first frame, so the tracking module 115 treats the target object as a whole, and extracts at least the target object from the display area. a feature point, and use the self-display The above feature points of the display area and the reference frame track the target. If the area ratio exceeds the preset value, indicating that the target occupies a large area in the first frame, the tracking module 115 divides the target into several parts for tracking. Based on this, the tracking module 115 defines a plurality of sub-regions in the display area, and extracts at least one feature point about the target object from the sub-regions, and then uses the feature points and reference frames extracted from the sub-regions. Track the target.

6A to 6C are schematic diagrams showing feature points of a captured object according to several embodiments of the present invention. In FIG. 6A, assuming that the object 60 (vehicle) F ₁ in the first frame of the display area 63 occupies the area ratio of the first frame F ₁ Movies does not exceed a preset value. In one embodiment, the tracking module 115 extracts one or more feature points about the target directly from the display area 63. In another embodiment, since the edge of the display area 63 may cover a background that does not belong to the target, in order to prevent the captured feature point from belonging to the background area, the tracking module 115 may obtain the display area 63 to be centered and retracted. A reduced area 63', and from this reduced area 63', extracts one or more feature points (indicated by the symbol "+" in Fig. 6A) with respect to the object 60. However, as long as the area occupied by the first 63 Frame F ₁ area ratio does not exceed a preset value, the tracking module 115 will display the target was viewed as a whole track.

In the example shown in FIG 6B, the target 60 is assumed to occupy the first frame of the area ratio F ₁ exceeds a predetermined value in the display region of the first frame F ₁ in the Movies 63, the tracking module 115 to the object 60 is divided into a number of for tracking part, first define two sub-regions in the first frame of ₁₆₅ F and 67, respectively, and retrieving a plurality of characteristic points on a target or object 60 (in FIG. 6B in the sub-regions 67 and 65 , indicated by the symbol "+"). In the present embodiment, the sub-region 65 is the upper left corner region of the display region 63, and the sub-region 67 is the lower right corner region of the display region 63, in which case the two sub-regions partially overlap. It should be particularly noted that the size of the sub-areas 65 and 67 may be a preset fixed size or may be dynamically adjusted according to the size of the display area 63. Since the edge of the display area 63 may contain a background that does not belong to the target, in order to prevent the captured feature points from being located in the background, the positions of the sub-areas 65 and 67 are not completely close to the edge of the display area 63, and It is contracted to the center of the display area 63 by a specific distance.

In the embodiment illustrated in FIG. 6C, assume that the target object 60 in the first frame F ₁ Movies display area 63 occupies the area ratio of the first frame of F ₁ exceeds a preset value, the tracking module 115 of the first frame F ₁ in Movies Defining two sub-regions 65 and 67, and extracting one or more feature points about the object 60 in the sub-regions 65 and 67, respectively (in FIG. 6C, denoted by the symbol "+"), and then the target The object 60 is divided into several parts for tracking. As shown in FIG. 6C, in this example, the sub-regions 65 and 67 do not overlap each other.

In the embodiment of FIGS. 6B and 6C, although the two sub-regions are defined as the upper left corner region and the lower right corner region of the display region 63, the present invention is not limited thereto. In other words, in other embodiments, the tracking module 115 may also define more than two sub-regions, and the location of each sub-region is not limited to the upper left corner region or the lower right corner region of the display region.

On the other hand, the traditional optical flow tracking algorithm tracks the movement of the same set of feature points after capturing the feature points. However, since this embodiment is applied to a complex environment with a large amount of change in brightness and darkness, in order to adapt to the scene change, the set of feature points captured by the tracking module 115 each time is only used to perform an optical flow tracking process. Detailed The tracking module 115 will capture a set of feature points for each frame, and for this set of feature points, only one optical flow tracking procedure will be performed, and then a new set of features will be retrieved again in the new frame. Point, because each set of feature points is used for a short time, this can avoid the traditional optical flow tracking algorithm must maintain the limit of consistent brightness. In this embodiment, the tracking module 115 uses a Kalman filter to estimate the range in which the feature points may appear in the next frame to balance the tracking speed and accuracy.

Specifically, it is assumed that the reference frame used to track the target includes: a second frame adjacent to the first frame, and a third frame adjacent to the second frame. First, the tracking module 115 extracts at least one feature point about the target object from the first frame (hereinafter, the feature points extracted from the first frame are referred to as a first group of feature points), and the first group of feature points is Extract from the display area of the target in the first frame. Next, the tracking module 115 calculates an estimated range in the second frame according to the Kalman filter, and the estimated range is an area in which the first set of feature points may be located in the second frame. Subsequently, the tracking module 115 performs an optical flow tracking algorithm to actually estimate the movement information of the first set of feature points within the estimated range of the second frame. After completing the optical flow tracking process between the first and second frames, the tracking module 115 discards the first set of feature points, and re-takes at least one feature point about the target from the second frame (eg, from the target) Taking feature points in a display area of the second frame, or extracting feature points from one or more small areas smaller than the display area, the following feature points taken from the second frame are referred to as a second group Feature point), according to the Kalman filter, calculate the estimated range in the third frame (ie, the region where the second set of feature points may be located in the third frame), and perform an optical flow tracking algorithm to actually estimate The second set of feature points in the third frame of the estimated range The movement information within the circumference, thus completing the optical flow tracking procedure between the second and third frames, and so on.

FIG. 7 is a schematic diagram of the tracking module 115 using three adjacent frames in the movie to track the target. The estimated ranges R ₁ , R ₂ , and R ₃ are calculated according to the Kalman filter, which is an area that may be obtained from the feature points of the previous frame. The smaller optical flow computing regions O ₁ , O ₂ , and O ₃ are the actual locations of the feature points taken from the previous frame. Since the target moves in the environment, the size of the target in each frame is inconsistent, so as shown in FIG. 7, the estimated ranges R ₁ , R ₂ , R ₃ and the optical flow computing regions O ₁ , O ₂ , O ₃ The size will change. However, the interaction of the optical flow tracking algorithm and the Kalman filter can avoid the result of tracking errors caused by the above situation. It must be specially stated that although FIG. 7 illustrates the manner of tracking the target object by three adjacent frames, the number of frames is not limited to three, as long as the optical flow tracking algorithm and the Kalman filter are used for tracking. And only one optical flow tracking procedure is taken from a set of feature points of a frame, which belongs to the category of the tracking module 115 of the present invention.

In this embodiment, once tracking of an object is started, the tracking module 115 facilitates the use of the Speeded Up Robust Features (SURF) algorithm to establish anti-scale and anti-rotation features of the target, and It is stored. The anti-scale and anti-rotation feature allows the target detection and tracking system 100 to track the target for a long time, regardless of whether the target once disappeared into the movie, as long as the target appears again in the movie, the target detection and tracking system 100 The anti-scale and anti-rotation features can be used to confirm the identity of the target. As shown in step S230 of FIG. 2 The identification module 117 uses the anti-scale and anti-rotation features to identify whether the object appears in a particular frame. For example, a particular frame refers to any frame that appears temporally after the reference frame.

FIG. 8 is a flow chart of identifying whether an object appears on a particular frame, according to an embodiment of the invention. Referring to FIG. 8, first, as shown in step S810, the recognition module 117 extracts the feature to be identified from the specific frame. Next, in step S820, the recognition module 117 compares the feature to be recognized with the previously stored anti-scale and anti-rotation features. For example, the identification module 117 uses the Kth Nearest Neighbor (KNN) algorithm to perform the comparison action. Next, as shown in step S830, the recognition module 117 determines whether the feature to be identified conforms to the anti-scale and anti-rotation features according to the comparison result. If the feature to be recognized does not conform to the anti-scale and anti-rotation features, as shown in step S840, the recognition module 117 determines that the target does not appear in the specific frame. On the other hand, if the feature to be recognized meets the anti-scale and anti-rotation features, the identification module 117 determines that the target appears in a specific frame as shown in step S850.

In the above embodiment, although the detection and tracking are performed for one target, it must be particularly noted that the target detection and tracking system 100 can perform synchronous detection and tracking for a plurality of objects in the movie. Whenever the detection module 113 detects the display area of the target in the new frame, the tracking module 115 first determines whether the target is an object that has been previously tracked. Next, tracking is performed by combining the optical flow tracking algorithm and the tracking mechanism of the Kalman filter, and the tracking result (for example, the position and size of the object) is stored.

In summary, the target detection and tracking method and system of the present invention, Using the classifier fusion algorithm and the texture features extracted from a large number of training positive and negative samples, the target classifier is trained to distinguish the target and background regions in the frame, so that even in the dynamic environment, it can still detect Measure the location of the target. For the detected objects, the tracking mechanism combined with the Kalman filter and the optical flow tracking algorithm is further utilized to achieve fast and stable tracking effect. Once the target is tracked, the anti-scale and anti-rotation features of the target are instantly created and stored, so that even if the target temporarily disappears into the movie, the anti-scaling can be utilized when the target appears again in the movie. The anti-rotation feature identifies the identity of the target. The target detection and tracking method and system of the present invention can achieve fast and accurate tracking effects in a dynamic environment.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention, and any one of ordinary skill in the art can make some changes and refinements without departing from the spirit and scope of the present invention. The scope of the invention is defined by the scope of the appended claims.

S210~S230‧‧‧ steps of the target detection and tracking method according to an embodiment of the present invention

Claims (10)

A method for detecting and tracking a target includes: capturing a frame in a movie as a first frame, and scanning the first frame to detect a display of a target in the first frame. a region; according to the display area occupying an area ratio of the first frame, and at least one reference frame in the movie following the first frame to track the target, and recording the primary target scaling of the target being tracked And anti-rotation features; and using the anti-scale and anti-rotation features to identify whether the object appears in a particular frame. The method of claim 1, wherein before the image frame in the movie is captured as the first frame, the method further comprises: obtaining a plurality of training positive samples and a plurality of training negative samples; Each of the training positive samples and a texture feature of each of the training negative samples; and using the training positive samples and the texture features of the training negative samples, and a classifier fusion algorithm to train a target a classifier, and scanning the first frame to detect the display area of the target in the first frame, comprising: forming a cascade classifier by n target classifiers, wherein n is a positive integer; and using the cascaded classifier to perform n scanning procedures on the first frame to detect the display area of the target in the first frame, wherein each of the n scanning programs includes In the first frame, a detection window is moved according to a specific direction to completely scan the first frame and then reduce the size of the first frame. The method of claim 1, wherein the target area is performed according to the area ratio of the display area occupying the first frame, and the at least one reference frame of the movie subsequent to the first frame. The step of tracking includes: determining whether the area ratio exceeds a preset value; if not, extracting at least one feature point about the target from the display area, and using the at least one feature extracted from the display area And tracking the target with the at least one reference frame; and if so, extracting at least one feature point about the target from the plurality of sub-regions in the display area, and extracting from the plurality of feature points The at least one feature point of the area and the at least one reference frame are tracked, wherein the sub-areas are an upper left corner area and a lower right corner area of the display area, and the sub-areas do not overlap each other or Partial overlap. The method of claim 1, wherein the at least one reference frame comprises: a second frame adjacent to the first frame, and a third frame adjacent to the second frame, and according to the The ratio of the area occupied by the display area to the first frame, and the at least one reference frame following the first frame in the movie, the step of tracking the target, comprising: extracting from the first frame At least one feature point of the target; calculating a predicted range in the second frame according to a Kalman filter; performing an optical flow algorithm to estimate the extracted from the first The at least one feature point of a frame, the mobile capital within the estimated range of the second frame Extracting at least one feature point about the target from the second frame; calculating an estimated range in the third frame according to the Kalman filter; and performing the optical flow tracking algorithm to estimate 撷The at least one feature point taken from the second frame, the movement information within the estimated range of the third frame. The method of claim 1, wherein the specific frame is any frame that is temporally after the at least one reference frame, and the anti-scaling and anti-rotation feature utilizes an accelerated robust feature capture (Speeded Up Robust Features). , the SURF) algorithm is established, and the step of identifying whether the object appears in the specific frame by using the anti-scaling and anti-rotation feature comprises: taking out a feature to be identified from the specific frame; And the anti-scaling and anti-rotation feature; and determining that the object appears in the particular frame when the feature to be recognized conforms to the anti-scale and anti-rotation feature. A target detection and tracking system includes: a storage unit that records a plurality of modules; and a processing unit coupled to the storage unit to access and execute the modules in the storage unit, the modules The method includes: a detecting module, capturing a frame in a movie as a first frame, and scanning the first frame to detect a display area of a target in the first frame; a tracking module, according to an area ratio of the display area occupying the first frame, and at least one reference frame in the movie following the first frame to track the target, and recording the target being tracked A primary anti-scaling and anti-rotation feature; and an identification module that utilizes the anti-scale and anti-rotation features to identify whether the object appears in a particular frame. The target detection and tracking system as described in claim 6 wherein the modules further comprise: a training module, obtaining a plurality of training positive samples and a plurality of training negative samples, and taking each of the trainings. a positive sample and a texture feature of each of the training negative samples, and using the training positive samples and the texture features of each of the training negative samples, and a classifier fusion algorithm to train a target classifier, and The detecting module forms a first-level classifier by using the target classifiers, and uses the cascade classifier to perform n scanning processes on the first frame to detect that the target is in the first frame. The display area of the display area, wherein each of the n scanning programs includes: moving, in the first frame, a detection window according to a specific direction to completely scan the first frame and then reducing the size of the first frame, And n is a positive integer. The target detection and tracking system of claim 6, wherein the tracking module determines whether the area of the display area occupying the first frame exceeds a preset value, and if not, from the display area Extracting at least one feature about the target Pointing, and tracking the target object by using the at least one feature point extracted from the display area and the at least one reference frame, and if so, extracting the target object from the plurality of sub-areas in the display area At least one feature point, and tracking the target object by using the at least one feature point extracted from the sub-regions, wherein the sub-regions are an upper left corner region and a right region of the display region Lower corner regions, and the sub-regions do not overlap or partially overlap each other. The target detection and tracking system of claim 6, wherein the at least one reference frame comprises: a second frame adjacent to the first frame, and a third adjacent to the second frame a frame, and the tracking module extracts at least one feature point about the object from the first frame, calculates an estimated range in the second frame according to a Kalman filter, and performs an optical flow tracking calculation Estimating the movement information of the at least one feature point of the first frame from the estimated range of the second frame, and extracting at least one feature point about the object from the second frame, Calculating an estimated range in the third frame according to the Kalman filter, and performing the optical flow tracking algorithm to estimate the at least one feature point extracted from the second frame in the third frame Mobile information within the estimated range. The target detection and tracking system according to claim 6, wherein the specific frame is any frame that is temporally after the at least one reference frame, and the tracking module uses an accelerated robust feature capture algorithm to establish The anti-scaling and anti-rotation feature, and the identification module extracts a feature to be recognized from the specific frame, compares the feature to be recognized with the anti-scaling and anti-rotation feature, and when the feature to be recognized meets the anti-scaling and When the anti-rotation feature is anti-rotation, it is determined that the object appears in the specific frame.