KR20160057503A - Violence Detection System And Method Based On Multiple Time Differences Behavior Recognition - Google Patents

Abstract

The present invention relates to a violence detection system for detecting violent behavior by reflecting time differences between behavior. based on multiple time difference behavior recognition and detection method thereof. The violence detection system of the present invention comprises: an image obtainment unit (110) for obtaining an image photographed by a camera; a person tracking unit (120) for detecting and tracking a person by setting the person as an object based on an image processing technology through the obtained image; a multi-period movement analysis unit (130) for analyzing and extracting movements of the person tracked during multiple divided periods including a short-term period, a mid-term period, and a long-term period; a mulitilayered violent behavior determination unit (140) for recognizing violent behavior in a mulitilayered structure based on data analyzed by the multi-period movement analysis unit (130), and for detecting violent behavior by integrating and analyzing a recognized result for the mulitilayered violent behavior; and an alarm generation unit (150) for generating an alarm by step according to detection of a violent behavior. Therefore the violence detection system according to the present invention can enhance accuracy for violent behavior detection and maximize a recognizing rate of violent behavior.

Description

TECHNICAL FIELD [0001] The present invention relates to a violent detection system and a detection method based on multi-paralytic behavior recognition,

The present invention relates to a violent detection system and a detection method based on multi-paralytic behavior recognition, and more particularly, to an image detection method and apparatus for detecting a violent act of a person, And more particularly, to a violent detection system and a detection method based on multiple parallax behavior recognition.

Many surveillance cameras have been installed in residential areas, inside and outside of sanitary parts, roads and public facilities, and the number of installed cameras has been increasing in recent years. The surveillance cameras primarily exist for monitoring security security situations and illegal activities.

However, there is a problem that the cognitive ability is inversely proportional to the increase in the numerical limit and monitoring time of the monitoring personnel. In order to overcome this, systems are applied which are applied to the situation and behavior recognition method through image analysis.

However, among the conventional methods, there is no method that takes into account the temporal difference in behavior. That is, all actions are recognized using the same parallax information. However, there are various disparities in behavior. For example, short-term violence is the act of punching, and longer-term violence is a combination of various actions. The surveillance system that does not take into account the time difference of such behavior has a problem that the recognition performance is inevitably reduced.

Korean Patent Registration No. 10-1394270 (May 31, 2014), image detection system and method

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and an object of the present invention is to provide a multi-layer structure-based action recognition technique using multiple parallaxes including short, The present invention provides a surveillance system and a surveillance method for designated objects and areas that can maximize the recognition rate by increasing the accuracy of detection of violent actions by using the detection of violent actions and considering the temporal difference of behaviors.

According to another aspect of the present invention, there is provided a violence detection system for detecting a violent behavior by reflecting a time lag of a behavior, the system comprising: An image acquisition unit 110; A person tracking unit 120 for detecting and tracking a person as an object based on the image processing technique through the obtained image; A multi-period motion analyzer 130 for analyzing and extracting motion information of a person being tracked for a plurality of periods including short-term, medium-term, and long-term; A multi-layered violent behavior determination unit 140 for recognizing a violent behavior of a multi-layered structure based on analysis data of the multi-period motion analysis unit 130 and detecting a violent behavior by convergence analysis of recognition results of the multiple actions; And an alarm generating unit 150 for generating an alarm in stages according to whether or not the violent action is detected.

Here, the human tracking unit 120 may include an object extraction unit 121 for detecting an object to be traced by a person in the image image, A motion extracting unit 122 for extracting motion information of an individual tracked object based on an optical stream and extracting group motion information for an object tracked by the object tracking unit 122, (123).

The multi-period motion analyzer 130 generates short-term motion data, medium-term motion data, and long-term motion data using object tracking and motion data from the human tracking unit 120, The determination unit 140 may include a short-term action determination unit 141 for recognizing an abusive action using the short-term motion data and providing a recognition result for the action and providing the recognition result, A long term action determination unit 142 for recognizing the violent action using the long term movement data and providing the result of recognition of the action as data, (143) for collecting data from each of the determination units (141 to 143) It may include a determination unit (144).

In addition, the violent behavior fusion determiner 144 may determine a probabilistic analysis method for the final violent behavior that is affected by the average of the multi-layer violent behavior probabilities and the previous final probabilities according to Equation 1 below, The violent behavior can be detected using the method of analyzing the violent behavior probability using the recognition result of each layer using the distance between the nodes of the SGT model according to Equation (2).

[Equation 1]

&Quot; (2) "

According to another aspect of the present invention, there is provided a violence detection method for detecting a violent behavior by reflecting temporal differences in behavior, the method comprising the steps of: An image acquiring step S210 of acquiring the image; A person tracking step (S220) of detecting and tracking a person as an object on the basis of the image processing technique through the obtained image; A multi-period motion analysis step (S230) for analyzing and extracting motion information of a person being tracked for a plurality of periods including a short-term, a middle-term, and a long-term; A multi-layer violent action determination step (S240) of recognizing a violent behavior of the multi-layer structure based on the analysis data of the multi-period motion analysis step (S230) and detecting a violent behavior by convergence analysis of recognition results of the multiple actions; And an alarm generating step (S250) of generating an alarm in stages according to whether or not the violent action is detected.

Herein, the human tracking step S220 may include an object extracting step S221 for detecting an object to be tracked in the video image, an image extracting step S221 for extracting an object detected in the object extracting step S221, An object tracking step (S222) for tracking a motion and a trajectory according to a change; and a motion information extracting step for extracting motion information of an individual track object based on an optical track, and extracting group motion information on an object tracked in the object tracking step (S222) And a motion extraction step (S223).

In addition, the multi-period motion analysis step S230 may generate short-term motion data, medium-term motion data, and long-term motion data using object tracking and motion data from the human tracking step S220, The determination step S240 may include a short-term action determination step S241 of recognizing an assaulting behavior using the short-term motion data and providing a recognition result of the act as data, and a short-term action determination step S241 of using the medium- A long-term action determination step (S242) of recognizing a violent action using the long-term motion data and providing a recognition result for the action, (S243), and the respective data are collected from the respective determination steps (S241 to S243) And a violent action fusion determination step (S244) for detecting the upper level.

In addition, the above-described violent behavior fusion determination step S244 may include a probabilistic analysis method for the final violent behavior that is influenced by the average of the multi-layer violent behavior probabilities and the previous final probabilities according to Equation (1) The violent behavior can be detected using the method of analyzing the violent behavior probability using the recognition result of each layer using the distance between the nodes of the SGT model according to Equation (2).

[Equation 1]

&Quot; (2) "

According to the monitoring system and the monitoring method for a designated object and area according to the present invention, a multi-layer structure-based action recognition technique using multiple parallax including short-term, medium-term, and long-term and a fusion technique using a plurality of SGT- By detecting violent behavior, it is possible to maximize the recognition rate by increasing the accuracy of detection of violent behavior and considering the time difference of behavior.

1 is a block diagram illustrating a functional configuration of a monitoring system for a designated object and area according to a preferred embodiment of the present invention;
2 is a block diagram and a flow chart showing a functional configuration and an operation principle of an image acquisition unit according to a preferred embodiment of the present invention,
FIG. 3 is a block diagram illustrating an operation principle in which data of a multi-period motion analyzer according to a preferred embodiment of the present invention is transmitted to each behavior determiner;
FIG. 4 is a flow chart showing the operation principle of the multi-layer violent behavior determiner according to the preferred embodiment of the present invention,
FIG. 5 is a schematic view for explaining an operation principle of a violent action fusion determination unit according to a preferred embodiment of the present invention;
FIG. 6 is a table showing quantitative / qualitative unit states of individuals and groups according to a preferred embodiment of the present invention;
FIG. 7 is a schematic view for explaining an arrangement for generating an FMTL-based symbolic model by a multi-layer violent behavior determination unit according to a preferred embodiment of the present invention;
FIG. 8 is a schematic diagram for explaining a configuration in which a multi-layered violent action determination unit recognizes a behavior based on an SGT model-based symbol lock state transition according to a preferred embodiment of the present invention;
FIG. 9 is an explanatory diagram illustrating an SGT modeling method for recognizing a group behavior according to a preferred embodiment of the present invention; FIG.
10 is a flowchart illustrating a procedure of a violent detection method based on multiple parallax behavior recognition according to a preferred embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, terms and words used in the present specification and claims should not be construed as limited to ordinary or dictionary terms, and the inventor should appropriately interpret the concepts of the terms appropriately The present invention should be construed in accordance with the meaning and concept consistent with the technical idea of the present invention.

Therefore, the embodiments described in this specification and the configurations shown in the drawings are merely the most preferred embodiments of the present invention and do not represent all the technical ideas of the present invention. Therefore, It is to be understood that equivalents and modifications are possible.

First, the configuration and functions of a violent detection system based on multiple parallax behavior recognition according to a preferred embodiment of the present invention will be described.

The multi-paralytic-action-based violence detection system according to the preferred embodiment of the present invention is a multi-paralytic structure-based action recognition technique using multiple parallaxes including short-term, medium-term and long-term and a plurality of SGT- FIG. 1 is a block diagram of a violent detection system according to an embodiment of the present invention. FIG. 1 is a block diagram illustrating a violent detection system according to an embodiment of the present invention. A multi-layer motion analyzer 130, a multi-layer violent action determiner 140, and an alarm generator 150. The multi-

The image acquiring unit 110 acquires a video image required for image analysis, and receives an image photographed from a camera (not shown) disposed at a surveillance place where the violent detection system according to the preferred embodiment of the present invention is installed And acquires a video image, which is basic data for image analysis. Here, the photographed image image is stored in a separate memory or database provided in the system, and can be stored in accordance with a time series standard.

The person tracking unit 120 is a unit for detecting and tracking a person as an object on the basis of an image processing technique in the obtained image image. As shown in FIG. 2, An object tracking unit 122 for tracking a motion and a trajectory according to a change in the image of the object detected by the object extracting unit 122, And a motion extracting unit 123 extracting motion information of the individual tracked objects based on the optical streams, and extracting group motion information.

2, the object detection unit 121 firstly blurs an image using a Gaussian filter to reduce the influence of noise, and then uses a multi-Gaussian model to generate a background And updates the Gaussian model by using Backbround Subtraction. It confirms whether the background initialization is completed, and uses the input during N frames for background stabilization.

Next, the ratio of the number of foreground pixels of (t-1) to the number of foreground pixels of (t) is extracted, and the foreground change ratio and the threshold are compared to filter out sudden changes of foreground.

In addition, Morphology operation is performed to remove Sort noise in the foreground, and the remaining foreground is removed on the basis of size, and then the foreground is converted on the basis of pixels on a pixel basis through labeling.

Then, when the asymmetric foreground and the foreground are asymmetric by using the Pearson asymmetry coefficient, the foreground region information is supplemented after the region is removed, and the foreground region including the same region is regarded as one region.

Next, the position between the Blob of (t-1) and the Blob of (t) is compared with the overlap region, and the information (ID) is inherited between consecutive objects and the object condition such as size, The blob is filtered based on the object, the blob that can be defined as the object is promoted as an object, and the promoted object is registered as a candidate for tracking.

Referring to FIG. 2, the object tracking unit 122 determines whether there is an object currently being tracked. The object tracking unit 122 converts the input image into a color domain in order to extract features of the object to be tracked. Estimate the next position of the object.

Then, the degree of position overlap between the next target position predicted by the particle filter and the candidate object detected by the object detection unit 121 is checked. If there is the same candidate object as the next expected position of the tracking object, .

Also, the information of the object being tracked (feature feature, positional information, etc.) is updated, and unnecessary tracking objects are removed. At this time, the object continuously moving must have a foreground continuously inside it, and the tracking object which does not include the foreground among the moving objects is removed using the foreground.

Then, the candidate object is registered as a tracking object and the trajectory information of the object being tracked is generated and updated. The proximity of the position between the objects being tracked is determined, and the objects in the distance are classified into the same group and the foreground region information is used And the region information of the group is supplemented.

Referring to FIG. 2, the motion extracting unit 123 first determines whether there is a tracking object, and if there is a tracking object, extracts a motion magnitude and a direction of a pixel unit through an optical flow as basic information for analyzing the motion of the object .

Also, the object motion information such as the moving speed, the motion size, and the motion direction is extracted using the optical stream corresponding to the individual tracking object area, and the moving speed, the motion size, the moving direction, etc. The group motion information is extracted.

The multi-period motion analyzer 130 is configured to analyze and extract human motion information tracked during a multi-period including a short-term, a middle-term, and a long-term. As shown in FIG. 3, Short-term motion data, long-term motion data, and long-term motion data using object tracking and motion data from the object tracking unit 120.

6, the quantitative data is extracted from the tracker, and the extracted data is accumulated for a short term, a medium term, and a long term (for example, 0.2 second, 1 second, and 2 seconds) Respectively.

In addition, the current behavior and situation are recognized through the FMTL and SGT having the same average value, and the short term, medium term, and long term time periods can be continuously fixedly used. Alternatively, , It is also possible to select and use a time term randomly within the Min-Max interval for the middle and long term.

The multi-layer violent behavior determination unit 140 recognizes a violent behavior of the multi-layered structure based on the analysis data of the multi-period motion analysis unit 130 and detects a violent behavior by performing a fusion analysis on the recognition results of the multi- 1, 3 and 4, a short-term action determining unit 141 for recognizing an assaulting action using the short-term motion data and providing a recognition result for the action and providing the result, A medium-term action determiner 142 for recognizing the violent action using medium-term motion data and providing a recognition result for the action and providing the result of the recognition; and a medium-term action detector 142 for recognizing the violent action using the long- A long-term action determination unit 143 that provides results of the data, and a data acquisition unit 143 that collects the data from each of the determination units 141 to 143, The shipment includes the violent action fusion determination unit 144.

Here, FIG. 4 shows a basic flow chart of the SGT, and FIG. 9 shows an SGT model for recognizing a group action (violence).

More specifically, the SGT uses the current unit state of the object (group, person) and the node at the previous time (situation and behavior, e.g., each node = situation / action such as motion, Judge the current time situation.

In addition, the current node check (the result of recognition at the previous time) of FIG. 4 is confirmed, and when the object is detected for the first time, the node starts from the root node. Can be shifted / moved.

In addition, basically, check the prediction first (Root Node other), check if the specification is possible at the transition node after the transition if the prediction is possible, then confirm the specification until the specification is no longer possible, Output as a result.

The embodiment shows black arrows going down in FIG. 9, and the prediction shows a red arrow moving in the horizontal axis. For example, if at time t the object has no unit state velocity through Tracker → FTML and it was in motion at the previous t-1, then this object compares the unit state of the predictable node stop node to its unit state, And if it is the same, it transits to the node (stop).

Then, it is determined whether the transition from the stop to the materializable node, the severe motion node, or the unit state of the node is the same, and if there is an inconsistency, the situation is terminated. In the same case, the node is moved, Find the nodes and compare them. If it is impossible to predict, it is confirmed that the current node is an End-Sit. (Situation). If it is correct, it transits to an upper node. After moving to an upper node, it repeats prediction / specification again.

FIG. 5 graphically shows that the result of the recognition of the short / medium / long term is fused and the result of the present fusion recognition is affected by the result of recognition of the fusion at the previous time.

The multi-layer violent behavior determination unit 140 recognizes a violent behavior of the multi-layered structure based on the analysis data of the multi-period motion analysis unit 130 and detects a violent behavior by performing a fusion analysis on the recognition results of the multi- , A probabilistic analysis method for the final violent behavior that is affected by the average of the multi-layer violent behavior probabilities and the previous final probabilities according to the following equation (1), and the probabilistic analysis method for the SGT model Violent behavior can be detected by using the probability analysis method of violence using the recognition result of each layer using the distance between nodes.

[Equation 1]

Here, the e _t is that the fusion results that the multiple time difference of the current frame (t time) final force results (probability indicating that the violence of the current screen, the quantitative _{data, 0.0 ~ 1.0), e t} -1 is the previous frame The recognition result shows that N _layers are the number of multiple cognitive layers, ie, 3, P _{layer, t} is the distance (0.0-1.0) of the result of Equation 2, indicating how close the current node is to the target (violence) node. P _{short, t} represents short-term cognitive layer result, P _{mid, t} represents mid-term cognitive layer result, P _{long, t} represents long-term cognitive layer result respectively.

&Quot; (2) "

[Mathematical Expression 2] is a formula for changing the perceived result (qualitative data) for each layer into probabilistic data using the inter-node distance, where dist _{layer and current} are the perceived result / And dist _max denotes the distance between the node (the violence: 1vs1, 1vs, many, many vs many) and the distant node (the example of Fig. 9, the abduction node).

In other words, dist _max is 4 and dist _{layer and current} are different values for each recognition node. For example, when the current node moves strongly, it is 1 and when it is 1vs1, it is 0. Since dist _max uses the same SGT model regardless of the layer (single / medium / long-term sensing layer), it has the same value and has the same value.

9 is an SGT model for recognizing violence in a violent detection system according to a preferred embodiment of the present invention. This SGT model is a SGT model for recognizing violations in a violent detection system according to a preferred embodiment of the present invention. .

FIG. 6 shows quantitative / qualitative data on an object used in a violent detection system according to a preferred embodiment of the present invention.

The alarm generating unit 150 generates an alarm at each step according to whether or not the violent action is detected. When the violent action is determined, the alarm generating unit 150 performs an alarm operation in a predetermined manner so that the user can immediately recognize the violent action.

Here, the step-by-step alarm indicates that 100% certain violence has occurred when 1.0, for example, using the result of the current et, and if it is 0.5, it means that an alarm is issued to generate an alarm requiring the confirmation of the monitoring personnel . There is no definite way to do this, and the user can set the result of the et to generate a phased alarm.

Next, a description will be made of a violent detection method based on recognition of multiple parallaxes according to a preferred embodiment of the present invention.

The violent detection method according to a preferred embodiment of the present invention detects a violent action using a multi-layer structure-based action recognition technology using multiple parallaxes including short-term, medium-term and long-term and a fusion technique of a plurality of SGT- FIG. 10 is a flowchart illustrating a method of detecting violence according to an exemplary embodiment of the present invention. FIG. 10 is a flowchart illustrating a method of detecting violence according to an exemplary embodiment of the present invention. Referring to FIG. 10, A person tracking step S220 for detecting and tracking a person as an object on the basis of the image processing technology through the acquired image image, and a human tracking step S220 for detecting a person who is tracked for multiple periods including short term, middle term and long term A multi-period motion analysis step (S230) for analyzing and extracting motion information, and a multi-period motion analysis step (S230) A multi-layered violent action determination step (S240) of recognizing a violent behavior of the multi-layered structure based on the data of the plurality of layers and detecting a violent action by convergence analysis of recognition results of the plural actions; And generating an alarm (S250).

Herein, the human tracking step S220 may include an object extracting step S221 for detecting an object to be tracked in the video image, an image extracting step S221 for extracting an object detected in the object extracting step S221, An object tracking step (S222) for tracking a motion and a trajectory according to a change; and a motion information extracting step for extracting motion information of an individual track object based on an optical track, and extracting group motion information on an object tracked in the object tracking step (S222) And a motion extraction step (S223).

Also, the multi-period motion analysis step S230 generates short-term motion data, medium-term motion data, and long-term motion data using the object tracking and motion data from the human tracking step S220.

In addition, the multi-layer violent act determination step (S240) may include a short-term action determination step (S241) of recognizing an assault behavior using the short-term motion data and providing a recognition result for the action, (S242) for recognizing a violent act by using the long-term motion data and providing a recognition result of the act as data, and a middle-term action determination step (S242) for recognizing the violent action using the long- A long-term action determination step (S243), and a judging step (S241 to S243), and collecting and integrating the data to detect a violent act (S244).

The above-described violent behavior fusion determination step S244 may include a probabilistic analysis method for the final violence behavior that is influenced by an average of the multi-layer violent behavior probabilities and a previous final probability according to the following [Equation 1] (2), the violent behavior is detected using the method of analyzing the violent behavior probability using the recognition results of each floor using the distances between the nodes of the SGT model.

[Equation 1]

&Quot; (2) "

According to the multi-parallax-aware recognition-based violence detection system and detection method according to the preferred embodiment of the present invention, a multi-layered structure-based action recognition technique using multiple parallaxes including short-term, SGT - Based Behavior Recognition Result By detecting violent actions using convergence technology, it is possible to maximize the recognition rate by increasing the accuracy of detection of violent actions and considering the time difference of behavior.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. It is to be understood that various modifications and changes may be made without departing from the scope of the appended claims.

110 ... image acquiring unit 120 ... human tracking unit
121 ... object detection unit 122 ... object tracking unit
123 ... motion extracting unit 130 ... multi-period motion analyzing unit
140 ... Multi-layer violent act determiner 141 ... Short-
142 ... Intermediate-term action determining unit 143 ... Long-term action determining unit
144 ... violent action fusion judgment unit 150 ... alarm generating unit

Claims (8)

A violence detection system for detecting a violent act in response to a time difference in behavior,
An image acquiring unit 110 for acquiring a video image photographed from a camera;
A person tracking unit 120 for detecting and tracking a person as an object within the acquired image image;
A multi-period motion analyzer 130 for analyzing and extracting motion information of a person being tracked for a plurality of periods including short-term, medium-term, and long-term;
A multi-layered violent behavior determination unit 140 for recognizing a violent behavior of a multi-layered structure based on analysis data of the multi-period motion analysis unit 130 and detecting a violent behavior by convergence analysis of recognition results of the multiple actions; And
And an alarm generating unit (150) for generating an alarm in stages according to whether or not the violent action is detected.
The method according to claim 1,
The person tracing unit 120,
An object extracting unit 121 for detecting an object to be tracked in the image,
An object tracking unit 122 for tracking a motion and a trajectory according to a change in the image of the object detected by the object extracting unit,
And a motion extracting unit (123) for extracting motion information of an individual tracked object based on an optical track, and extracting group motion information, on the object tracked by the object tracking unit (122).
3. The method of claim 2,
The multi-period motion analyzer 130 generates short-term motion data, medium-term motion data, and long-term motion data using object tracking and motion data from the human tracking unit 120,
The multi-layer violent action determiner 140 includes a short-term action determiner 141 for recognizing an assault behavior using the short-term motion data and providing recognition results for the action, A medium-term action determination unit 142 for recognizing the violent action and providing a result of recognition of the action, and providing the result of recognition of the violent action using the long- And a violent action convergence judging unit (144) for collecting the data from the long term action judging unit (143) and each judging unit (141 to 143) and analyzing the result of the collective analysis to detect the violent act.
The method of claim 3,
The violent behavior fusion determination unit 144,
A probabilistic analysis method for the final violent behavior affected by the average of the probability of multi-layered violence action and the previous final probability according to Equation 1 below and the distance between nodes of the SGT model according to the following Equation (2) And detecting the violent action using the analysis result of the violent action using the recognition result of each layer used.
[Equation 1]

&Quot; (2) "

A violence detection method for detecting a violent behavior reflecting a time lag difference in behavior,
An image acquiring step (S210) of acquiring a video image photographed from a camera;
A person tracking step S220 of detecting and tracking a person as an object in the acquired image image;
A multi-period motion analysis step (S230) for analyzing and extracting motion information of a person being tracked for a plurality of periods including a short-term, a middle-term, and a long-term;
A multi-layer violent action determination step (S240) of recognizing a violent behavior of the multi-layer structure based on the analysis data of the multi-period motion analysis step (S230) and detecting a violent behavior by convergence analysis of recognition results of the multiple actions; And
And an alarm generating step (S250) of generating an alarm in stages according to whether or not the violent action is detected.
6. The method of claim 5,
The human tracking step S220 includes:
An object extraction step (S221) of detecting an object to be traced by a person in the image,
An object tracing step (S222) of tracing a motion and a trajectory according to a change in the image of the object detected in the object extracting step (S221)
And a motion extracting step (S223) of extracting motion information of an individual tracked object based on an optical stream, and extracting group motion information, on the object tracked in the object tracking step (S222).
The method according to claim 6,
The multi-period motion analysis step S230 generates short-term motion data, medium-term motion data, and long-term motion data using object tracking and motion data from the human tracking step S220,
The multi-layer violent action determination step (S240) may include a short-term action determination step (S241) of recognizing an assault behavior using the short-term motion data and providing recognition results of the action by data, A medium-term action determination step (S242) of recognizing the violent action and providing a result of recognition of the action, and providing the result of recognition of the violent action using the long-term motion data; (S244) of collecting data from the long-term action determination step (S243) and each of the determination steps (S241 to S243), and performing an integrated analysis to detect a violent action.
8. The method of claim 7,
In the violation action fusion determination step S244,
A probabilistic analysis method for the final violent behavior affected by the average of the probability of multi-layered violence action and the previous final probability according to Equation 1 below and the distance between nodes of the SGT model according to the following Equation (2) And detecting a violent action by using a method of analyzing a violent action probability using recognition results of each layer used.
[Equation 1]

&Quot; (2) "

Images