KR101472674B1 - Method and apparatus for video surveillance based on detecting abnormal behavior using extraction of trajectories from crowd in images - Google Patents

Abstract

The video surveillance apparatus according to embodiments of the present invention includes a foreground extractor for extracting a foreground from image frames, a length detector for detecting length information, coordinate information, and direction information based on feature points extracted from the foreground using a Kanade-Lucas- For each of the extracted feature point trajectories, feature point trajectories for determining crowd behavior are selected based on length similarity, coordinate similarity, and direction similarity with respect to other feature point trajectories And a trajectory classifying unit for classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory.

Description

TECHNICAL FIELD [0001] The present invention relates to a video surveillance method and an image surveillance method based on abnormal behavior detection using crowd trajectory extraction,

The present invention relates to computer vision technology, and more particularly, to video surveillance technology.

In the case of a large-scale CCTV system, since dozens to hundreds of images must be monitored by the management personnel, it is often necessary to take a large number of personnel and miss important situations due to concentration, fatigue and arbitrary judgment of the monitoring personnel , There are many problems in storing images of a huge amount of time.

Therefore, in order to monitor public places, CCTV camera shoots public places and automatically analyzes the images to extract unspecified number of objects and analyze the movement. If abnormal movement is detected, it automatically warns the manager or other related automation There is a growing demand for an intelligent video surveillance system that transmits information to the system.

Conventional intelligent video surveillance systems first extract individual objects from an image, track the extracted individual objects to obtain a trajectory, and analyze the trajectory to estimate the motion of the individual object. It is then analyzed whether the estimated motion of the object is normal or not.

Such conventional systems require a great deal of computation and errors while extracting, tracking, trajectory analysis, and motion estimation steps of an object.

For example, it is not easy to extract and analyze objects with conventional intelligent video surveillance systems when monitoring crowded crowds such as squares, parks, train stations, playgrounds.

An object of the present invention is to provide a video surveillance method and video surveillance apparatus based on abnormal behavior detection using computationally efficient mass trajectory extraction.

According to an aspect of the present invention,

A foreground extractor for extracting a foreground from image frames;

A feature point trajectory extractor for extracting feature point trajectories represented by length information, coordinate information, and direction information based on feature points extracted from the foreground using a Kanade-Lucas-Tomaasi algorithm;

A trajectory selecting unit for selecting, for each of the extracted feature point trajectories, feature point trajectories for determining a crowd behavior based on length similarity, coordinate similarity, and direction similarity with other feature point trajectories; And

And a trajectory classifying unit for classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory.

According to one embodiment, the locus classifying section may operate to classify the normal locus or the abnormal locus based on the length information and the direction information of each of the selected minutiae loci.

According to one embodiment, the locus classifying unit may linearly classify, for each selected feature point locus,

The abnormal exponent Tg is compared with a predetermined threshold value and classified into a normal trajectory or an abnormal trajectory,

Here, the weight α is 0??? 1, T _length is the length information of the trajectory, and T _direction is the direction information of the trajectory.

According to one embodiment, the locus classifier comprises:

And to classify each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory based on learning using the example patterns.

According to one embodiment, the video surveillance apparatus includes:

And an abnormal situation determination unit for determining whether or not the abnormal situation exists based on the classified abnormal trajectories.

According to one embodiment,

The abnormal condition determination unit may be operable to determine whether the abnormal condition is present according to the ratio of the abnormal trajectories with respect to the selected minutia traces.

According to one embodiment, the foreground extractor may operate to extract the foreground by subtracting the background from the image using an adaptive background model based on the Gaussian blend model.

According to one embodiment, the locus selector comprises:

The length similarity degree, the coordinate similarity degree, and the direction similarity degree with respect to the first trajectory in the first frame among the two consecutive frames, and the second trajectories in the second frame are respectively calculated, and the calculated length similarity degree, coordinate similarity degree, If the number of second trajectories higher than predetermined threshold values is greater than a predetermined number, selecting the first trajectory, or otherwise discarding the first trajectory.

According to an embodiment, the trajectory selector may calculate coordinate similarities with respect to feature point trajectories selected based on the length similarity,

And calculate the direction similarity to target feature point trajectories based on the coordinate similarity.

According to another aspect of the present invention,

Extracting a foreground from image frames;

Extracting feature point trajectories represented by length information, coordinate information, and orientation information based on feature points extracted from the foreground using a Kanade-Lucas-Tomaasi algorithm;

Selecting feature point trajectories for determining crowd behavior based on length similarity, coordinate similarity, and direction similarity with other feature point trajectories, for each of the extracted feature point trajectories; And

And classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory.

According to an embodiment, the step of classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory includes:

And sorting the normal trajectory or the abnormal trajectory based on the length information and the direction information of each of the selected feature point trajectories.

According to an embodiment, the step of classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory includes:

The method according to claim 1,

And comparing the abnormal index Tg with a predetermined threshold value to classify the abnormality index Tg into a normal trajectory or an abnormal trajectory,

Here, the weight α is 0??? 1, T _length is the length information of the trajectory, and T _direction is the direction information of the trajectory.

According to an embodiment, the step of classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory includes:

And classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory based on learning using the example patterns.

According to an embodiment, it may further comprise determining whether the abnormal situation is based on the classified abnormal trajectories.

According to one embodiment, the step of determining whether an abnormal situation is based on the classified abnormal trajectories,

And determining whether the abnormal condition is present according to the ratio of the abnormal trajectories to the selected feature point trajectories.

According to one embodiment, extracting the foreground from the image frames comprises:

And extracting the foreground by subtracting the background from the image using the adaptive background model based on the Gaussian mixture model.

According to one embodiment, the step of selecting the feature point trajectories comprises:

Calculating a length similarity degree, a coordinate similarity degree and a direction similarity degree with the second trajectories in the second frame with respect to the first trajectory in the first frame among the two consecutive frames; And

Selecting the first trajectory if there are a plurality of second trajectories having the calculated length similarity degree, coordinate similarity degree, and direction similarity degree higher than predetermined threshold values, respectively, or discarding the first trajectory otherwise have.

According to one embodiment, the step of selecting the feature point trajectories comprises:

Calculating coordinate similarities with respect to feature point trajectories selected based on length similarity; And

And calculating the direction similarity degree with respect to the feature point trajectories selected based on the coordinate similarity.

According to the video surveillance method and video surveillance apparatus based on abnormal behavior detection using the crowd trajectory extraction according to the present invention, when sudden movements are identified while certain movements are observed in a dense crowd, if an abnormal behavior is determined, can do.

According to the video surveillance method and video surveillance apparatus based on abnormal behavior detection using the crowd trajectory extraction of the present invention, compared with the conventional method of analyzing the trajectory obtained by tracking the object and judging whether the behavior is abnormal, Instead, it collectively analyzes the trajectories of feature points in the foreground, and notifies the administrator only when an abnormal trajectory is detected, or traces only an object showing an abnormal trajectory through a separate image recognition system. Finally, an abnormal situation actually occurs Complexity and accuracy can be satisfied by allowing the administrator to perform the determination as to whether or not to perform the determination, or to perform the determination of the situation only in an abnormal situation even if the determination is dependent on the machine.

1 is a block diagram illustrating an image surveillance apparatus capable of detecting an abnormal behavior using a crowd locus extraction according to an embodiment of the present invention.
2 is a flowchart illustrating an image surveillance method based on abnormal behavior detection using crowd trajectory extraction according to an embodiment of the present invention.

For the embodiments of the invention disclosed herein, specific structural and functional descriptions are set forth for the purpose of describing an embodiment of the invention only, and it is to be understood that the embodiments of the invention may be practiced in various forms, The present invention should not be construed as limited to the embodiments described in Figs.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The same reference numerals are used for the same constituent elements in the drawings and redundant explanations for the same constituent elements are omitted.

1 is a block diagram illustrating an image surveillance apparatus capable of detecting an abnormal behavior using a crowd locus extraction according to an embodiment of the present invention.

Prior to describing the features of the present invention, the conventional method will be described first. In the method adopted in the existing intelligent video surveillance apparatus, individual objects are first extracted from the image. Individual objects are tracked individually and trajectories of individual objects obtained through tracking are obtained. If the trajectory is specified according to the time, the trajectory is analyzed to determine whether it is normal or abnormal. If it is determined that the abnormal behavior is abnormal, the administrator is notified. Since the intelligent video surveillance device detects all the behaviors of individual objects and then determines whether or not they are normal, the amount of computation is large and complex, and the accuracy is low.

On the other hand, the abnormal behavior detection method and the video surveillance apparatus using the crowd trajectory extraction of the present invention collectively analyze the trajectories of the feature points of the entire foreground separated from the background, instead of tracking individual objects, for example, locally It is determined that there is a possibility that some kind of abnormal situation occurs at the position when the trajectory of the vehicle suddenly changes in the tendency of the trajectory or the trajectory specified in advance as abnormal is displayed and it is notified to the manager or the additional trajectory is related to the abnormal trajectory Image analysis is performed only on the part.

As described above, the abnormal behavior detection method and the image sensing apparatus of the present invention are limited to the role of determining the possibility of an abnormal situation and notifying the administrator thereof, and it is left to the administrator to judge whether an abnormal situation actually occurs or not, Even if the situation is directly determined without manager intervention, the image analysis is performed only in the portion related to the abnormal trajectory, so that the accuracy can be secured while greatly reducing the calculation amount.

1, the image sensing apparatus 10 includes a foreground extracting unit 11, a feature point trajectory extracting unit 12, a trajectory selecting unit 13, a trajectory classifying unit 14, and an abnormality determining unit 15 .

First, the foreground image extracting unit 11 removes a background from image frames provided from a camera or other image source and extracts a foreground.

The foreground is a concept that includes all the partial images except for the background in the image frame, and in the present invention, the foreground is collectively handled without distinguishing them as individual objects.

In order to remove the background from the consecutive image frames, for example, a difference image technique which removes the unchanged part as a background in successive images, a background modeling technique by accumulating or learning image frames for a predetermined time, And adaptive background modeling techniques such as Gaussian mixture model.

Since the conventional video surveillance system extracts the foreground in order to identify objects, the sophistication of the background model greatly affects the overall performance. However, the video surveillance system of the present invention collectively uses the extracted foreground as a whole, There is an advantage that background modeling and accurate object extraction are not necessarily required.

The foreground data extracted by the foreground extracting unit 11 may miss a part of pixels of moving objects existing in the image or may partially include pixels belonging to the background in reality. Can be effectively eliminated while the computation steps are being performed.

Preferably, the foreground extracting unit 11 extracts the foreground using an adaptive background model based on the Gaussian mixture model.

The adaptive background model based on the Gaussian mixture model can separate the background without being influenced by weather, shadows, changes in light source, appearance and disappearance of objects in a short period of time.

The foreground data extracted by the foreground extracting unit 11 is provided to the feature point trajectory extracting unit 12.

The feature point trajectory extracting unit 12 extracts feature point trajectories represented by length information, coordinate information, and direction information based on feature points extracted from foreground data using the Kanade-Lucas-Tomaasi algorithm do.

The feature points are reference points for extracting a significant locus between foreground data of consecutive image frames. In general, a feature point for tracking an object moving in an image is a pixel having a feature that is distinguished from neighboring pixels.

Common techniques for extracting feature points include Scale Invariant Feature Transforms (SIFT), Harris corner detectors, Forstner detectors, Kanade-Lucas-Tomasi (KLT) It is known.

Among them, the Kanade-Lucas-Thoma algorithm is based on a feature point tracking theory using a gradient descent method that repeatedly aligns image patches in successive image frames, Algorithms that add the notion of estimating operators by affine transforms that can clearly define the relationships, detect edges in consecutive image frames (or stereo vision images) It is proposed that image tracking (or matching) between two images can be effectively performed with high correlation corners as feature points.

In particular, the KLT algorithm implicitly assumes brightness constancy, temporal persistence, and spatial coherence. Brightness consistency is the assumption that the pixels belonging to one object in the scene do not change, and time persistence is the assumption that the motion of the image piece changes in time with time, and spatial coherence assumes that neighboring points belonging to the same surface have similar motion to be.

Accordingly, the locus data between the minutiae points of two temporally successive frames extracted by the KLT algorithm can be expressed with length information, coordinate information, and direction information.

The minutiae locus extracting unit 12 extracts minutiae points from the foreground data extracted from the consecutive image frames using the KLT algorithm and extracts minutiae points Respectively.

In the present invention, a trajectory is a straight line connecting two feature points that match a subject in the same real world in two consecutive image frames. Throughout this specification, a minutiae trajectory can also be referred to simply as a trajectory. The trajectory in a certain frame means a trajectory from a feature point belonging to the frame to a feature point belonging to the next frame.

Since the feature points extracted by the KLT algorithm are feature points having a correlation between two consecutive frames in essence, a trajectory between two feature points having correlation in the foreground data of two consecutive frames is generated .

These feature point trajectories substantially represent the overall flow of objects moving over the background in the image. For example, assuming that a large number of people appear, but the approximate direction in which people move is to monitor a public place, such as a subway station culled in several directions, the foreground data will contain a large number of human objects, The number of feature points extracted by the feature point trajectory extracting unit 12 and the number of traces of the feature point will also be large.

However, as will be discussed subsequently, no matter how many feature point trajectories, except for feature trajectories that are noises, they can be grouped into feature point trajectories and other feature point trajectories that coincide with several directions in which people are primarily moving.

To this end, the trajectory selecting unit 13 uses the fact that the trajectory selecting unit 13 has length information, coordinate information, and direction information of the trajectory data provided by the KLT algorithm, For each of them, feature point trajectories for determining crowd behavior are selected based on length similarity, coordinate similarity, and direction similarity with other feature point trajectories.

For example, a feature point trajectory in which all three similarities with other feature point trajectories are higher than predetermined similarity threshold values or at least one similarity degree is higher than a predetermined similarity threshold value is selected.

Specifically, first, the trajectory selector 13 calculates the length similarity with the trajectories t _j in the second frame with respect to a certain trajectory t _i in the first one of two consecutive frames, If there are more than a predetermined number of trajectories t _j higher than the threshold value, the corresponding trajectory t _i is selected, otherwise, the trajectory t _i is discarded.

For example, the trajectory selection unit 13 has a length of similarity is high trajectory t _j for the trajectory t _i of the current frame to find among the trajectory in the next frame or, conversely long degree of similarity is high trajectory to the trajectory t _i of the present frame a t _j can be found from among the trajectory in the previous frame.

The length similarity can be measured according to how close the length information of the two trajectories t _i , t _j is.

If the number of trajectories t _j whose length similarity is higher than a predetermined threshold among the trajectories in the second frame is less than a predetermined number with respect to a certain trajectory t _i in the first frame among the two consecutive frames, , The trajectory t _i in the first frame means a trajectory that exists only in the frame among several tens of frames per second that constitute the moving image, and thus it can be said that the trajectory is not a trajectory reflecting the movement of real objects. Therefore, such a trajectory can be removed as foreground artifacts inevitably arising only in the frame in the foreground extraction and trajectory extraction processes.

Next, the trajectory selecting unit 13 calculates the coordinate similarity with the trajectories t _j in the second frame with respect to a certain trajectory t _i in the first one of two consecutive frames, and if the calculated coordinate similarity satisfies a predetermined threshold If there are more than a predetermined number of trajectories t _j higher than the value, the corresponding trajectory t _i is selected, otherwise the trajectory t _i is discarded.

For example, the trajectory selection unit 13 is the current coordinate degree of similarity is high trajectory t _j for the trajectory t _i of a frame found in the sign in the next frame, or, on the contrary coordinate degree of similarity is high trajectory to the trajectory t _i of the present frame a t _j can be found from among the trajectory in the previous frame.

The coordinate similarity can be measured depending on how close the coordinate information of the two trajectories t _i and t _j is, for example, whether it is within a circle of a predetermined radius or within a predetermined range with respect to each coordinate axis.

At this time, according to the embodiment, the coordinate similarity sorting operation may be performed only on the trajectories t _i that have successfully passed the length similarity sorting operation.

If the number of trajectories t _{j having} a coordinate similarity degree higher than a predetermined threshold value is less than a predetermined number in the trajectories in the second frame with respect to any trajectory t _i in the first frame of the two consecutive frames, , Then this trajectory may be artifact, or it may be removed as a weak trajectory that is not suitable for behavioral determination.

Finally, the trajectory selector 13 calculates the direction similarity to the trajectories t _j in the second frame with respect to a certain trajectory t _i in the first one of two consecutive frames, If there are more than a predetermined number of trajectories t _j higher than the value, the corresponding trajectory t _i is selected, otherwise the trajectory t _i is discarded.

For example, the trajectory selection unit 13 is present the trajectory t _i a high trajectory t _j direction similarity for the frames found in the sign in the next frame, or, on the contrary Current high trajectory direction similarity to the trajectory t _i of the frame a t _j can be found from among the trajectory in the previous frame.

In this case, according to the embodiment, the direction similarity selection operation may be performed only on the trajectories t _i that have successfully passed the length similarity selection operation and the coordinate similarity selection operation.

The direction similarity can be measured according to how close the directional information of the two trajectories t _i , t _j is, for example, the degree of cosine similarity.

If the number of trajectories t _{j having} a direction similarity degree higher than a predetermined threshold is less than a predetermined number in the trajectories in the second frame with respect to a certain trajectory t _i in the first frame among the two consecutive frames, , This trajectory may also be regarded as an artifact or, if not, a weak trajectory unsuitable for behavioral determination.

The trajectory selection unit 13 can greatly reduce the number of trajectories to be determined for determining normal behavior and abnormal behavior through the calculation of such length similarity, coordinate similarity, and direction similarity selection.

For example, the feature trajectories were extracted by KLT algorithm for foreground extracted by Gaussian mixture modeling from images of scenes where dozens of people were moving in one direction along a park walkway, and 10,000 trajectories were extracted Approximately 500 trajectories were eliminated, 9,500 trajectories were selected, and about 740 trajectories were selected through coordinate similarity selection. 460 trajectories were selected through direction similarity selection.

Finally, the selected feature point trajectories can be regarded as robust trajectories that can determine the normality of behavior while reflecting the movement of real objects.

The locus classifying unit (14) classifies each of the selected feature point loci into a normal locus or an abnormal locus.

According to the embodiment, the locus classifying section 14 can classify the normal locus or the abnormal locus based on the length information and direction information of each of the selected minutiae loci.

Specifically, the trajectory classification unit 14 can classify the abnormal index Tg calculated linearly for each trajectory into a normal trajectory or an abnormal trajectory by comparing the abnormal trajectory Tg with a predetermined threshold as shown in the following Equation (1).

Wherein the weight α is 0≤α≤1 and, T is the length information _length, T _direction of the locus is the direction information of the trajectory, the length or the direction of the trajectory may be selectively adjusted by the effect of an abnormal index by weight α.

For example, if the length of certain trajectories is significantly longer than the other trajectories and the speed estimated by the length of such trajectories is much greater than the predetermined rate, or vice versa, the length of certain trajectories is significantly shorter than the other trajectories If the velocity estimated by the length of such a trajectory is significantly smaller than the predetermined velocity, this means that the movement of the objects corresponding to such trajectory is not normally and much faster or much slower. Thus, the abnormal exponent Tg of such trajectories will exceed the threshold and is classified as an abnormal trajectory.

In another embodiment, the locus classifying section 14 may classify each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory based on learning using the example patterns.

On the other hand, according to the embodiment, since a small number of abnormal trajectories does not necessarily indicate an abnormal situation, the video surveillance apparatus 10 further includes an abnormal situation determination section 15 for determining whether or not the abnormal situation is based on the classified abnormal trajectories .

Specifically, the abnormal situation determining unit 15 can determine whether or not the abnormal state is present according to the ratio of the abnormal trajectories to the selected feature point trajectories. For example, if the trajectories classified in the abnormal trajectory among the trajectories exceed 20%, it can be determined that the abnormal condition is present.

For example, suppose that there are about 100 people appearing in a video where people usually move at a certain speed and direction. If two or three of them suddenly run, suddenly stop, or suddenly move in a different direction, a small number of abnormal trajectories will be extracted by such actions, and most people are moving in a calm manner. It is difficult to assume that other urgent situations have occurred.

On the other hand, if more than 20 out of 100 people suddenly run, suddenly stop, suddenly move in a different direction, it is assumed that a large number of abnormal trajectories will be extracted by such actions and a serious situation other than usual occurs in the place can do.

Furthermore, the abnormal situation determination unit 15 may issue an abnormal situation warning to the manager, or may call a more accurate image recognition calculation function only in the image region in which the abnormal trajectories are concentrated.

2 is a flowchart illustrating an image surveillance method based on abnormal behavior detection using crowd trajectory extraction according to an embodiment of the present invention.

Referring to FIG. 2, in the video surveillance method of the present invention, in step S21, a foreground is extracted from image frames provided from a camera or other image source.

Preferably, the foreground can be extracted using an adaptive background model based on a Gaussian mixture model.

Subsequently, in step S22, feature point trajectories extracted as length information, coordinate information, and orientation information are extracted based on the feature points extracted from the foreground data using the Kanade-Lucas-Toomasi algorithm.

In step S23, feature point trajectories for determining the crowd behavior are selected on the basis of the length similarity degree, coordinate similarity degree, and direction similarity degree with respect to other feature point trajectories, for each extracted feature point trajectory. Unselected trajectories are eliminated.

Specifically, with respect to any trajectory t _i in the first frame of two consecutive frames, trajectory t _j whose length similarity is higher than a predetermined threshold can be selected among the trajectories in the second frame.

Next, with respect to a trajectory t _i passing through, for example, length similarity selection in the first frame of two consecutive frames, a trajectory t _j whose coordinate similarity is higher than a predetermined threshold in the trajectories in the second frame is Can be selected.

Finally, with respect to a certain trajectory t _i that has passed through, for example, length similarity selection and coordinate similarity selection in the first of two consecutive frames, the direction similarity in the trajectories in the second frame is higher than a predetermined threshold a trajectory t _j may be selected.

In step S24, each of the selected feature point trajectories is classified into a normal trajectory or an abnormal trajectory.

According to the embodiment, each of the selected feature point trajectories can be classified into a normal trajectory or an abnormal trajectory based on the respective length information and direction information.

Specifically, each of the selected feature point trajectories can be classified into a normal trajectory or an abnormal trajectory by comparing the abnormal index Tg linearly calculated for each trajectory with a predetermined threshold value as shown in Equation (1).

In another embodiment, each of the selected feature point trajectories may be classified into a normal trajectory or an abnormal trajectory based on learning using the example patterns.

On the other hand, according to the embodiment, in an additional step S25, it is possible to judge whether or not the abnormal situation is based on the classified abnormal trajectories.

Specifically, the abnormality may be determined depending on the ratio of the abnormal trajectory among the selected feature point trajectories.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, Modification is possible. Accordingly, the spirit of the present invention should be understood only in accordance with the following claims, and all of the equivalent or equivalent variations will fall within the scope of the present invention.

Further, the apparatus according to the present invention can be implemented as a computer-readable code on a computer-readable recording medium. A computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer system is stored. Examples of the recording medium include a ROM, a RAM, an optical disk, a magnetic tape, a floppy disk, a hard disk, a nonvolatile memory, and the like, and a carrier wave (for example, transmission via the Internet). The computer-readable recording medium may also be distributed over a networked computer system so that computer readable code can be stored and executed in a distributed manner.

10 image sensing device
11 Foreground Extractor
12 feature point trajectory extracting section
13 Trajectory selector
14 Trajectory classification section
15 Abnormal situation judgment section

Claims (20)

A foreground extractor for extracting a foreground from image frames;
A feature point trajectory extractor for extracting feature point trajectories represented by length information, coordinate information, and direction information based on feature points extracted from the foreground using a Kanade-Lucas-Tomaasi algorithm;
Each of the extracted feature point trajectories is compared with other feature point trajectories on the basis of each of the length similarity degree, coordinate similarity degree and direction similarity degree to select the feature point trajectories for determining the crowd behavior, A locus selector; And
And a trajectory classifying unit for classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory. The video surveillance apparatus according to claim 1, wherein the trail classifying section operates to classify the trajectory classifying section into a normal trajectory or an abnormal trajectory based on length information and direction information of each of the selected trajectory of feature points. [Claim 2] The method according to claim 2, wherein the locus classifying unit linearly calculates, for each of the selected feature point trajectories,

The abnormal exponent Tg is compared with a predetermined threshold value and classified into a normal trajectory or an abnormal trajectory,
Wherein the weight α is 0 ≦ α ≦ 1, T _length is the length information of the trajectory, and T _direction is the direction information of the trajectory. The apparatus according to claim 1, wherein the locus classifying section
And classifies the selected feature point trajectory into a normal trajectory or an abnormal trajectory on the basis of learning using the example patterns. The video surveillance apparatus according to claim 1, further comprising an abnormal situation determination section for determining whether or not an abnormal situation exists based on the classified abnormal trajectories. The video surveillance apparatus according to claim 5, wherein the abnormal situation determination unit is operable to determine whether or not an abnormal situation exists according to a ratio of abnormal trajectories to the selected feature point trajectories. The video surveillance apparatus of claim 1, wherein the foreground extractor is operable to extract a foreground by subtracting a background from an image using an adaptive background model based on a Gaussian mixture model. [2] The apparatus according to claim 1,
The length similarity degree, the coordinate similarity degree, and the direction similarity degree with respect to the first trajectory in the first frame among the two consecutive frames, and the second trajectories in the second frame are respectively calculated, and the calculated length similarity degree, coordinate similarity degree, Wherein the controller is operative to select the first trajectory if the second trajectories are higher than predetermined threshold values, and to discard the first trajectory otherwise. The apparatus according to claim 8, wherein the locus selector is configured to calculate coordinate similarities based on feature point trajectories selected based on length similarity,
And calculates the direction similarity degree of the feature point trajectories selected based on the coordinate similarity. delete Extracting a foreground from image frames;
Extracting feature point trajectories represented by length information, coordinate information, and orientation information based on feature points extracted from the foreground using a Kanade-Lucas-Tomaasi algorithm;
Each of the extracted feature point trajectories is compared with other feature point trajectories on the basis of each of the length similarity degree, coordinate similarity degree and direction similarity degree to select the feature point trajectories for determining the crowd behavior, step; And
And classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory. 12. The method of claim 11, wherein classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory comprises:
And classifying the image into a normal trajectory or an abnormal trajectory based on the length information and the direction information of each of the selected feature point trajectories. [12] The method of claim 12, wherein classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory comprises:
The method according to claim 1,

And comparing the abnormal index Tg with a predetermined threshold value to classify the abnormality index Tg into a normal trajectory or an abnormal trajectory,
Wherein the weight α is 0 ≦ α ≦ 1, T _length is the length information of the trajectory, and T _direction is the direction information of the trajectory. 12. The method of claim 11, wherein classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory comprises:
And classifying each of the selected feature point trajectories into a normal trajectory or an abnormal trajectory based on learning using the example patterns. 12. The method of claim 11, further comprising the step of determining whether an abnormal situation exists based on the classified abnormal trajectories. 16. The method of claim 15, wherein the step of determining an abnormal condition based on the classified abnormal trajectories comprises:
And determining whether the abnormal condition is present according to the ratio of the abnormal trajectories to the selected feature point trajectories. The method of claim 11, wherein extracting the foreground from the image frames comprises:
And extracting the foreground by subtracting the background from the image using the adaptive background model based on the Gaussian mixture model. 12. The method of claim 11, wherein the step of selecting the feature point trajectories comprises:
Calculating a length similarity degree, a coordinate similarity degree and a direction similarity degree with the second trajectories in the second frame with respect to the first trajectory in the first frame among the two consecutive frames; And
Selecting the first trajectory if the calculated length similarity degree, coordinate similarity degree, and direction similarity degree are equal to or greater than a predetermined number, and if not, discarding the first trajectory Wherein the video monitoring method comprises the steps of: 19. The method of claim 18, wherein selecting the feature point trajectories comprises:
Calculating coordinate similarities with respect to feature point trajectories selected based on length similarity; And
And calculating direction similarities with respect to the feature point trajectories selected based on the coordinate similarity. A recording medium on which a program recorded on a computer for implementing a video surveillance method according to any one of claims 11 to 19 can be read by a computer.

Images