CN100568266C - An Abnormal Behavior Detection Method Based on Local Statistical Feature Analysis of Sports Field - Google Patents

Abstract

The present invention relates to a kind of abnormal behaviour detection method based on the sports ground partial statistics characteristic analysis.This system mainly comprises the sports ground analysis of video image, and partial statistics characteristic extracts, based on the statistical learning and the mode identification technology of sample.At first, extract the feature of object in the image, calculate its motion state, form a sports ground by basic motion analysis technique.On this basis, local movable information is carried out statistical nature extract, obtain the local motion feature of sports ground.Space distribution relation table with these motion features is shown as global structure information at last, adopts the method based on statistical learning, discerns the type of behavior.This algorithm improves the efficient and the robustness of algorithm by directly carrying out behavior identification based on the analysis of movable information.

Description

An Abnormal Behavior Detection Method Based on Local Statistical Feature Analysis of Sports Field

technical field

The invention relates to computer monitoring technology based on moving images, pattern recognition technology based on statistical learning and feature analysis technology based on local sports field. The invention is a visual monitoring content analysis method, which belongs to the fields of computer vision and intelligent information processing.

Background technique

As society pays more attention to public safety issues, real-time monitoring has been more and more widely used. The current monitoring problem is that it is difficult to process a large amount of monitoring information in a timely and effective manner. Using computer vision technology to analyze and understand people's movements, and provide records and alarms will help improve the level of security monitoring in public places. Using computers to assist in the recognition of human actions and events has become a hot issue in the field of computer vision.

The intelligent analysis technology of visual surveillance is a hot and difficult issue in the field of computer vision, involving cutting-edge topics such as image processing and machine learning. In recent years, there have been a lot of academic related research, including the research on the national 863,973 intelligent monitoring project. Abnormal behavior detection refers to first analyzing and modeling the custom normal action behavior and other behavior data, and then judging whether the behavior is abnormal according to the test behavior and the similar procedures of the two. The current research mainly focuses on detecting basic events such as the appearance of people and crossing boundaries by detecting whether the movement area and speed of the human body meet the predetermined conditions and other restrictions. On this basis, companies such as ObjectVideo in the United States have developed corresponding products and obtained certain applications.

However, there are still some problems in these researches and applications: First, the current research is mostly based on the detection and tracking information of the foreground area, and then analyzes whether the predetermined conditions are met, so that the problem is limited to the detection and tracking of objects, and seldom involves the detection and tracking of objects. The understanding of object movement and its laws has certain limitations in function. Secondly, part of the visual understanding method uses a mechanical human body model method to restore the state of human joint motion by matching video images, and then understand the motion of the human body. This process increases the intermediate links, making the model more complex and difficult to achieve accurately.

Contents of the invention

The present invention extracts local motion features through the analysis of the human sports field, and synthesizes the features and distributions of different regions to obtain the description of individual motions, and finally performs machine learning on a large amount of normal and abnormal motion data to establish a human motion feature model, Realize real-time detection of human abnormal behavior.

The present invention mainly relates to the fields of computer vision and pattern recognition. The basic information of the human sports field is obtained through video images. On this basis, the complex human motion is decomposed into simple local motions based on the characteristics of motion complexity. Through positioning and analysis The local sports field extracts local motion features; then integrates the motion features and distribution characteristics of different regions of the human body to obtain a description of individual motion; finally, performs machine learning on a large amount of normal and abnormal motion data to establish a human motion feature model to realize abnormal human behavior real-time detection.

The whole process of the method is as follows: First, several groups of predefined different types of actions are collected through video as training samples; then, according to the detection results of the human body motion area, the local area is divided; the optical flow field of the movement is established, and the local motion field is positioned and analyzed. method to obtain the description of local motion; extract regional features and integrate them into the spatial distribution relationship of feature pairs, and express them as overall features; Uncertain actions are subjected to feature extraction in the above steps, and the probability of belonging to a certain type of action is judged by the Bayesian method, and finally the action type with the highest probability is selected as the recognition result.

(1) Based on the feature analysis of the local motion field.

Firstly, the basic optical flow data of the sports field is provided through the analysis of the video image, and the given image I(x, y, t) represents the brightness value I at the image coordinate (x, y) at any time t. Then, according to the basic optical flow formula (1), on the assumption that the luminous characteristics of the human body do not change, we can obtain the motion velocity V of the feature point at (x, y).

$\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; /</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; /</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; xu</span>}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; /</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; yv</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Where u, v are image sampling coordinate axes.

The motion feature L={V ₁ , V ₂ } of the optical flow is obtained by performing Gaussian clustering on the optical flow field vector, and expressed as two main independent motion components V ₁ , V ₂ to simplify the expression of the overall optical flow field.

The spatial distribution of the optical flow is achieved through the calculation of moments, and the basic moment features can be obtained according to formula (2), where the zero-order moment features M _0,0 represent the area of the motion area, and the first-order moment features M _0,1 and M _{1, 0} respectively represent the distribution of the center of gravity in the direction of the two coordinate axes, and higher-order features represent the distribution information of a higher frequency range, so that the distribution characteristics of the motion area can be represented by a series of moments and their transformations J={M ₀₀ , M ₀₁ , M ₁₀ ,...}.

M _pq =∑ _x ∑ _y ||V||(x, y)x ^p y ^q (2)

Due to the complexity of human motion, global motion and its distribution analysis can only obtain the overall motion trend, and its role is very limited; in addition, due to the structural characteristics of the human body, there are obvious global uncertainties and continuity of local area motion , we localize the above optical flow analysis and motion region analysis, that is, perform feature computation within a given region.

(2) Global human motion feature description method

In addition to the description of local features, global motion features are more important to describe the overall features through local features and their distribution. The specific methods are as follows:

Firstly, the distribution state of the motion area is analyzed through the difference image between the background and the current frame, and the motion difference image can be directly obtained according to the formula (3).

D(x,y,t)=I(x,y,t)-I(x,y,t ₀ ) (3)

Among them, I(x, y, t ₀ ) is the background image; then calculate the boundary of this motion area to determine the position of the human body, and combine the structural characteristics of the human body to pre-divide the human body motion area into independent r parts, respectively denoting The movement situation of the area where the local limbs of the person is located; thus, in each area, it is necessary to independently calculate the optical flow field, as well as the optical flow motion feature L _i and motion distribution feature J _i in the area i, where i∈N, i≤r;

In order to describe the relative positions of different feature regions, we define a feature pair T _q =<J ₁ , J _q >, which represents the vector formed by the line connecting the feature J ₁ and the central position of J _q . The final global feature of human motion can be encoded as:

g=[L ₁ , ..., L _r , J ₁ , ..., J _r , T ₂ , ..., T _r ].

(3) Action sequence recognition based on Bayesian

First define the target set H={h ₀ , h ₁ …h _n } for action sequence recognition, and collect a certain amount of samples for each action in each action set G={G ₀ , G ₁ …G _n }, and establish The corresponding SVR (Support Vector Regression) model. Then given a new action sequence g _i , firstly, the instant probability P(g _i |h _m ) belonging to a certain target can be obtained through the SVR model, and by comparing the membership probabilities of all targets, the one with the highest probability is the instant probability of the model Motion recognition results.

Since the immediate result cannot reflect the entire movement process, in order to reduce the _error rate, it is necessary to calculate an action sequence g _t , g _(t-1) , ... the probability P(h _l |g _t , g _(t-1) , ...).

According to the continuity of the state, we can regard the motion change as a Markov process, and think that its state is only related to the last state and the current action, then

P(h _l |g _t , g _(t-1) ,...)＝P(h _l |h _(l-1) , g _t )

In addition, assuming that the t-1 state and the action at time t are independent, then

P(h _l |h _(l-1) ，g _t )＝P(h _l |h _(l-1) )P(h _l |g _t )

According to Bayes formula:

P(h _l |g _t )＝P(h _l )P(g _t |h _l )/∑ _k P(g _t |h _k )P(h _k )

Given the state transition matrix P(h _l |h _(l-1) ) and the prior probability P(h _l ), the recognition result of the entire sequence can be obtained according to the SVR model. Among them, the state transition matrix and prior probability are obtained according to the frequency statistics of various actions in the training samples, and can also be carried out by using a uniform distribution strategy.

Description of drawings:

Fig. 1 is the system operation process of the present invention.

Fig. 2 is the algorithm flow of the present invention.

Fig. 3 is the schematic diagram of motion field analysis of the present invention, and wherein square box is the local area that divides, and red circle represents the detected local motion characteristic center, and red line represents its speed, and the blue connection line between the circle represents the characteristic vector of global characteristic.

Detailed ways:

Embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings.

Fig. 1 is a flow chart of system operation, and Fig. 2 is a flow chart of feature extraction algorithm, and the inventive method is according to Fig. 1 flow chart, and system operation comprises following specific steps:

1. Sample video data collection: This system uses machine learning principles to model and recognize human motion, so a large number of motion samples are required for related learning and verification. Taking five specific actions as examples, we collected several video sequences, and took some of them as training sets for learning, and some of them as test sets for model verification, and constructed video data samples.

2. Image analysis and feature extraction: the specific extraction process is carried out as shown in Figure 2, which is introduced as follows.

1) Obtain video data: including training video and video collected in real time during detection.

2) Motion field calculation: Given a video image I(x, y, t), according to the basic optical flow formula (1), calculate the motion speed of all feature points, which contains size and direction information.

3) Human body detection and region division: firstly, the human body movement region is obtained by the method of difference between the current image I(x, y, t) and the background image I(x, y, t ₀ ). Calculate the center and rectangular boundary information of the area, divide the area by quadtree, and obtain four local initial areas of motion.

4) Local feature extraction: local feature extraction includes two aspects of features, one is vector principal component analysis L, for the velocity vectors of all feature points, two main velocity components are obtained by clustering; Calculate the moment in the local area of the velocity, and obtain the moment feature J including the center of gravity in the feature area.

5) Global feature analysis: using the master-slave feature mode method to describe. First, we select a main feature and match it with other features to form three feature pairs T, which are expressed in the form of vectors, including the distance between features and angle information. The global feature g of the motion is obtained by sequentially encoding the structural information between the features and the local features.

6) Model learning and recognition: Finally, all training and learning are based on such global motion features.

3. Behavior modeling: First, a certain amount of samples are collected for each action, and through image analysis and feature extraction, the global features of human motion are obtained, and a corresponding SVR model is established. At the same time, the state transition matrix P(h _l |h _(l-1) ) and the prior probability P(h _l ) are calculated through the change of the state sequence.

4. Establish a model library: So far the entire offline training process is completed, and the trained SVR model, state transition matrix P(h _l |h _(l-1) ) and prior probability P(h _l ) are saved in the database.

5. Real-time data acquisition: In the application process, the above SVR model and related parameters must be initialized first, and the motion video sequence is collected in real time through the camera for analysis.

6. Use the same method as 2 for feature extraction.

7. Behavior recognition: Given a new action sequence g _t , g _t-1 , ..., the probability P( _{h l |} g _t , g _t-1 , .. ....). We regard it as a Markov process, and think that its state is only related to the last state and current action, then the Bayesian formula:

P(h _l |g _t )＝P(h _l )P(g _t |h _l )/∑ _k P(g _t |h _k )P(h _k )

Given the state transition matrix P(h _l |h _(l-1) ) and the prior probability P(h _l ), the similarity of the entire sequence to a certain action can be obtained according to the SVR model.

8. Result analysis: Finally, by comparing the similarity of each behavior, select the most similar behavior that meets the threshold (Y=0.5) as the recognition result, otherwise it is considered as not belonging to the above training actions, that is, it is judged as abnormal behavior.

Claims (1)

1. A method for abnormal behavior detection based on local statistical feature analysis of sports field, characterized in that, the method comprises the following steps: (1) Sample video data collection: several groups of different types of actions predefined through video collection, as training samples; (2) Image analysis and feature extraction, including the following steps: i) Obtain video data: including training video and video collected in real time during detection; ii) Calculation of sports field: Firstly, the basic optical flow data of the sports field is provided through the analysis of the video image, and the given image I(x, y, t) represents the brightness value I at the image coordinate (x, y) at any time t; Then according to the basic optical flow formula (1), on the assumption that the luminous characteristics of the human body do not change, we can get the motion speed V of the feature point at (x, y): $\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; /</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; /</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; xu</span>}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; /</span><span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; yv</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$ Where u, v are image sampling coordinate axes; iii) Human body detection and area division: Firstly, the distribution state of the motion area is analyzed through the difference image between the background and the current frame, and the motion difference image can be directly obtained according to formula (2): D(x,y,t)=I(x,y,t)-I(x,y,t ₀ ) (2) Among them, I(x, y, t ₀ ) is the background image; then calculate the boundary of this motion area to determine the position of the human body, and combine the structural characteristics of the human body to pre-divide the human body motion area into independent r parts, respectively denoting The movement situation of the area where the local limbs of the person is located; thus, in each area, it is necessary to independently calculate the optical flow field, as well as the optical flow motion feature L _i and motion distribution feature J _i in the area i, where i∈N, i≤r; iv) Local feature extraction: The motion feature L={V ₁ , V ₂ } of the optical flow is obtained by Gaussian clustering of the optical flow field vector, expressed as two main independent motion components V ₁ , V ₂ to simplify the whole The expression of the optical flow field; the spatial distribution of the optical flow is realized by the calculation of the moments, and the basic moment features can be obtained according to the formula (3), where the zero-order moment feature M _{0, 0} represents the area of the motion area, and the first-order moment feature M _{0 , 1} and M _{1, 0} respectively represent the distribution of the center of gravity in the direction of the two coordinate axes, and higher-order features represent the distribution information of a higher frequency range, so that the distribution characteristics of the motion area can be represented by a series of moments and their transformations J={M ₀₀ , M ₀₁ , M ₁₀ ,...}; M _pq =∑ _x ∑ _y ||V||(x, y)x ^p y ^q (3) v) Global feature analysis: use the method based on the master-slave feature mode to describe, in order to describe the relative positions of different feature areas, define a feature pair T _q =<J ₁ , J _q >, which means the connection line between the feature J ₁ and the central position of J _q The resulting vector, the final global feature of human motion can be coded as: g=[L ₁ ,...,L _r , J ₁ ,...,J _r , T ₂ ,...,T _r ]; vi) Model learning and recognition: Finally, all training and learning are based on such global motion features; (3) Behavior modeling, including the following steps: First define the target set H={h ₀ , h ₁ …h _n } for action sequence recognition, and collect a certain amount of samples for each action in each action set G={G ₀ , G ₁ …G _n }, and establish The corresponding support vector regression model; then given a new action sequence g _t , firstly, the instant probability P(g _t |h _m ) belonging to a certain target can be obtained through the support vector regression model, and by comparing the membership probabilities of all targets , the one with the highest probability is the instant motion recognition result of the model; Since _the immediate results cannot reflect the entire movement process, in order to reduce the error rate, it is necessary to calculate the _probability P(h _l |g _t , g _{( t- 1)} ,...); According to the continuity of the state, we can regard the motion change as a Markov process, and think that its state is only related to the last state and the current action, then P(h _l |g _t , g _(t-1) ,...)＝P(h _l |h _(l-1) , g _t ) In addition, assuming that the t-1 state and the action at time t are independent, then P(h _l |h _(l-1) ，g _t )＝P(h _l |h _(l-1) )P(h _l |g _t ) According to Bayes formula: P(h _l |g _t )＝P(h _l )P(g _t |h _l )/∑ _k P(g _t |h _k )P(h _k ) Given the state transition matrix P(h _l |h _(l-1) ) and prior probability P(h _l ), the recognition result of the entire sequence can be obtained according to the support vector regression model; (4) Establish a model library: save the trained support vector regression model, state transition matrix P(h _l |h _(l-1) ) and prior probability P(h _l ) into the database; (5) Real-time data collection: real-time collection of motion video sequences through the camera for analysis; (6) adopt the method identical with step (2) to carry out feature extraction; (7) Behavior recognition: Given a new action sequence, the Bayesian formula: P(h _l |g _t )＝P(h _l )P(g _t |h _l )/∑ _k P(g _t |h _k )P(h _k ) Given the state transition matrix P(h _l |h _(l-1) ) and the prior probability P(h _l ), the similarity of the entire sequence to a certain action can be obtained according to the support vector regression model; (8) Result analysis: Finally, by comparing the similarity of each behavior, select the most similar behavior that meets the threshold requirements as the recognition result, otherwise it is considered that it does not belong to the above training actions, that is, it is judged as abnormal behavior.

Images