CN105046285B - A kind of abnormal behaviour discrimination method based on kinematic constraint - Google Patents

Abstract

The invention relates to a method for identifying abnormal behavior based on motion constraints, which belongs to the field of video monitoring, and includes a training process, which includes the following steps: 1. Using a fast calibration method to calibrate the input first group of video frame images; 2. The calibrated input video frame image is used for background modeling, and the optical flow feature is extracted from the frame image by the optical flow method; 3. The optical flow feature in the foreground is analyzed and selected by the pedestrian’s motion constraints; 4. The SVM is used to analyze the The selected optical flow features are learned to obtain the parameters of the SVM classifier model that can be used for abnormal behavior identification; it also includes the detection process, the input of the detection process is the second group of videos, and the SVM parameters used are obtained during the training process and used for Detect and judge model parameters. The invention optimizes the calculation of optical flow features, eliminates noise optical flow features, and makes the optical flow features on pedestrians more prominent. And it can make judgments on abnormal behaviors such as running.

Description

An Abnormal Behavior Identification Method Based on Motion Constraints

technical field

The invention relates to an abnormal behavior identification method based on motion constraints.

Background technique

When the monitoring environment is somewhat crowded but not serious, the abnormal behavior identification method based on support vector machine (SVM) can be improved based on the previous work. For example, the robust tracking method is used in the middle, and the mutual occlusion between pedestrians is not serious. Tracking about 20 people in the real-time environment, a method for tracking the defective human body image is proposed in the paper.

However, in a particularly crowded environment, when the degree of mutual occlusion between pedestrians exceeds a certain limit, and it is impossible to obtain complete or clear pedestrian outline features, the above methods will have difficulties in understanding crowd behavior, and new methods are needed. to solve this problem. To this end, the concept of motion constraints is proposed and combined with the learning of optical flow features to realize abnormal behavior identification and monitoring in crowded environments.

Contents of the invention

The technical scheme that the present invention solves the problems of the technologies described above is as follows:

A method for identifying abnormal behavior based on motion constraints, including a training process, the training process includes the following steps:

(1), input the first group of videos, and preprocess the input video frame images, including calibration and denoising by using a fast calibration method;

(2), carry out background modeling to the calibrated input video frame image, then use the optical flow method to extract the optical flow feature from the frame image;

(3) Analyze and select the optical flow features in the foreground with the pedestrian's motion constraints;

(4) Use SVM to learn the selected optical flow features to obtain an SVM classifier that can be used for abnormal behavior identification.

(5) Write down the parameters of the SVM classifier model.

On the basis of the above technical solutions, the present invention can also be improved as follows.

Further, the extraction of the optical flow feature from the frame image adopts the Lucas-Kanade algorithm.

Further, the motion constraint model is established by Bayesian formula, the formula is:

P(u,v|I(x,y,t))=αP(u,v)ΠP(I(x _i ,y _i ,t _i )|u,v)

Among them, P(u,v) is the prior probability, which represents the probability of the two-dimensional motion speed of the object, α is the proportional factor, P(u,v|I(x,y,t)) is the space (x,y) The probability of its two-dimensional motion velocity under the condition of the gray value at time t, P(I(x,y,t)|u,v) is the probability of the gray change of the object when it is moving, that is, the posterior probability.

Further, the analysis and selection of the optical flow features in the foreground with the pedestrian's motion constraints in the step 3 specifically includes the following steps:

(1), from the foreground image after the background detection, select the optical flow feature from the image area of the pedestrian;

(2) Select the optical flow features that meet the threshold requirement through the motion constraint model, and the threshold is the threshold used for screening the posterior probability P(I(x,y,t)|u,v).

Further, the range of α is between 0 and 1,

Further, the threshold is optimized during the iterative solution process of the SVM modeling process to obtain an optimal value.

Further, the present invention also includes a detection process, the steps of which are:

(1), input the second group of videos, and preprocess the input video frame images, including calibration and denoising by using a fast calibration method;

(2), carry out background modeling to the calibrated input video frame image, then use the optical flow method to extract the optical flow feature from the frame image;

(3) Using the motion constraints of pedestrians, analyze and select the optical flow features in the foreground according to the threshold obtained in the training stage;

(4) Use the SVM classifier obtained in the training process to judge the selected optical flow features. If the abnormal behavior (knowing that the input is an abnormal behavior) is judged correctly, it can be proved that the classifier is reliable.

The beneficial effects of the present invention are: the present invention combines optical flow feature selection with motion constraints to obtain an SVM classifier that can be used for abnormal behavior judgment in the training process, and the second group with known behavior categories is input in the detection process The video is used to test the trained model to verify the correctness and reliability of the model, and it can also play a role in further generalizing the model. The present invention uses pedestrian motion constraints to optimize optical flow feature selection for the first time, effectively eliminates noise optical flow features and enhances flow features of interest, and can make accurate judgments on abnormal behaviors such as running.

Description of drawings

Fig. 1 is a schematic diagram of the method of the present invention;

Fig. 2 is the schematic diagram utilizing Bayesian modeling;

Fig. 3 is a comparison chart of calculation results of optical flow features using motion constraints and calculation results of optical flow features not using motion constraints.

Detailed ways

The principles and features of the present invention are described below in conjunction with the accompanying drawings, and the examples given are only used to explain the present invention, and are not intended to limit the scope of the present invention.

An abnormal behavior identification system based on motion constraints includes a preprocessing module, a background modeling module, an optical flow extraction module, a flow field selection module, a motion constraint module and an SVM classifier module.

The preprocessing module is used to preprocess the input video frame image, including calibration, denoising, etc.;

The background modeling module is used to carry out background modeling to the preprocessed input video frame image;

The optical flow extraction module is used to extract optical flow features from frame images after background modeling;

The motion constraint module is used to use the motion constraints of pedestrians to filter out qualified optical flow features according to the threshold,

The SVM classifier module is used to learn the selected optical flow features. In the training phase, the optimal model parameters and thresholds are obtained through the iterative and optimized solution of the SVM algorithm; in the testing phase, the model obtained by training is used to test the known samples to check the accuracy and reliability of the model, whether it can Pedestrian behavior (including abnormal behavior) to make accurate judgments.

Example 1

As shown in Figure 1, an abnormal behavior identification method based on motion constraints, including a training process and a detection process;

The training process includes the following steps:

(1), input the first group of videos, and preprocess the input video frame images, including calibration and denoising by using a fast calibration method;

(2), carry out background modeling to the calibrated input video frame image, then use the optical flow method to extract the optical flow feature from the frame image;

(3) Using the motion constraints of pedestrians, analyze and select the optical flow features in the foreground according to a certain threshold;

(4) Use SVM to learn the selected optical flow features to obtain an SVM classifier that can be used for abnormal behavior identification.

(5) Write down the parameters of the SVM classifier model.

1. Optical flow feature definition

Optical flow is widely used in computer vision tasks, such as face recognition, gait modeling, object tracking, etc. The two optical flow algorithms currently used are the Horn-Schunck algorithm and the Lucas-Kanade algorithm. The Horn-Schunck algorithm has smoother optical flow, global information and accurate temporal differentiation, but is relatively slow and has rough boundary contours. The Lucas-Kanade algorithm has more boundary errors, but the algorithm is relatively simple and can be calculated quickly. Considering the calculation speed, we use the Lucas-Kanade algorithm.

When motion constraints are not considered, the steps of obtaining optical flow features include optical flow calculation and optical flow extraction steps. The optical flow feature vector obtained at this time is a vector that contains the motion information of the feature points in the frame image, and does not contain any prior knowledge of pedestrian motion constraints.

The generation of optical flow feature vector is divided into two steps:

The first step is to find the feature vector points of the human body image block, such as the angle vertex between the head and the left shoulder of a certain pedestrian.

In the second step, the feature vector points of the previous frame image corresponding to the found feature vector points, for example, the vertex of the angle between the head and the left shoulder of the same pedestrian.

2. Optical flow feature extraction

All optical flows in the selected foreground image can be written as P=[p(1),p(2),...,p(i)], i=1..N, where p(i)=[Lx ,Ly,Mx,My,α], the number of optical flow is N (can be calculated from the optical flow algorithm), Lx and Ly represent the abscissa value and ordinate value of an optical flow position in the image block, Mx and My Respectively represent the abscissa value and the ordinate value of the corresponding optical flow amplitude (equivalent to the velocity value), and α represents the direction angle value. Since the number of optical flows in each image block is different, it is necessary to standardize each image block, and the processed optical flow features can be written as Q=[p(x ₁ ),p(x ₂ ), ...,p(xi ₎ ], where p( _xi ) is selected from p(i). The input parameters can finally be described by formula (2.1):

(2.1) I=[p(x ₁ ),p(x ₂ ),...,p(x _i )], i=1..N

After selecting the optical flow features as the input of the classifier, the problem can be reduced to a classification problem of two or more classes. Since SVM is used as a classifier for pedestrian motion behavior, the selected input parameters are the position, direction and velocity of optical flow features in the image block, etc.

3. The introduction of motion constraints

When the motion constraints are not considered, the optical flow features of pedestrian behavior are directly used as the input of the SVM to train the classifier. Although optical flow features can play a role in behavior recognition, due to the low robustness of the optical flow method in complex environments, motion constraints need to be introduced.

After analyzing the optical flow characteristics of pedestrians in crowded conditions, it is found that the movement pattern of pedestrians conforms to the movement rules related to people, and is significantly different from other moving objects such as driving vehicles (the movement of vehicles is a a rigid body motion, and the speed and trajectory are different from pedestrians); while the motion pattern of pedestrians, especially the speed characteristics of walking under normal circumstances, is very regular. The pedestrian motion constraints are modeled using Bayesian criteria, as shown in Figure 2.

The modeling process has two steps:

The first step is to ensure that optical flow features are only selected from the image area of pedestrians (in a crowded environment, it is difficult to judge whether a pixel is located in the image area of a person, a solution to this problem is to use a robust A fast and fast algorithm for detecting people, or selecting points from the foreground image after background detection, the latter is the method used in this method), the proportion of the selected optical flow features to the total number of optical flow features is P(A) ;

The second step is to filter out the optical flow features that meet the threshold requirements through the motion constraint model, so as to make the calculation of the optical flow features more accurate, so as to judge whether the selected pixels meet the human body motion constraints. The ratio of optical flow features is P(B), and during the training process, P(A) and P(B) will reach a dynamic balance;

In the second step, the method adopts Bayesian method (Bayesian) to model the motion constraints. The purpose of Bayesian methods for human motion analysis is to compute the posterior probability in a given image. The posterior probability is obtained by calculating the gray level I(x,y,t) at (x,y) at time t and the two-dimensional motion velocity (u,v) of the object.

(3.1) P(u,v|I(x,y,t))=αP(u,v)P(I(x,y,t)|u,v)

In formula (3.1), if it is assumed that the images are observed from different positions and the time is continuous and independent, then there are:

(3.2) P(u,v|I(x,y,t))=αP(u,v)ΠP(I(x _i ,y _i ,t _i )|u,v)

Among them, P(u,v) is the prior probability, which represents the probability of the two-dimensional motion speed of the object, α is the proportional factor, P(u,v|I(x,y,t)) is the space (x,y) The two-dimensional motion speed probability under the condition of the gray value at time t can be obtained in the optical flow calculation, P(I(x,y,t)|u,v) is the gray value of the object when it is moving Probability, that is, the posterior probability, and then filter out the optical flow features that meet the conditions according to the threshold value P(I(x,y,t)|u,v), and the threshold value is any value in the initial stage;

After filtering out the optical flow features that meet the threshold requirements through the motion constraint model, the selected optical flow features are used as the input of the SVM algorithm and learned. In the iterative solution of the SVM algorithm, the threshold and SVM model parameters will be continuously optimized, and finally the best Optimal threshold and SVM model parameters.

Method of the present invention also comprises detection process, and described detection process comprises the following steps:

(1), input the second group of videos, and preprocess the input video frame images, including calibration and denoising by using a fast calibration method;

(2), carry out background modeling to the calibrated input video frame image, then use the optical flow method to extract the optical flow feature from the frame image;

(3) Using the motion constraints of pedestrians, analyze and select the optical flow features in the foreground according to the threshold obtained in the training stage;

(4) The SVM classifier obtained in the training process is used to judge the selected optical flow features. If the abnormal behavior (knowing that the input is an abnormal behavior) is judged correctly, it can be proved that the classifier is reliable.

4. Experimental results

This experiment demonstrates the results of using motion constraints for optical flow feature computation and a class of abnormal behavior monitoring in crowded environments. In the experiment, the resolution of the surveillance video is 320×240, and the average duration is 30 minutes. For each 30-minute image, the first 10 minutes are used for model training, and the last 20 minutes are used for model testing.

Computational results of optical flow features considering motion constraints.

Figure 3 illustrates the calculation results of an optical flow feature using motion constraints. The left figure shows the calculation results without motion constraints, and the right figure shows the calculation results with motion constraints. It can be seen from Figure 3 that the motion constraint optimizes the calculation of optical flow features, eliminates the noise optical flow features, and the optical flow features on pedestrians are more significant. At the same time, it can also be seen from the figure that there are obvious differences between the motion constraints of the car and the motion constraints of pedestrians.

The above descriptions are only preferred implementation cases of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within range.

Claims (5)

1. a kind of abnormal behaviour discrimination method based on kinematic constraint, which is characterized in that including training process and detection process, institute Training process is stated to include the following steps：

(1), first group of video is inputted, frame image is demarcated using quick calibrating method；

(2), background modeling is carried out to calibrated input video frame image, then extracts light stream from frame image with optical flow method Feature；

(3), the Optical-flow Feature in foreground is analyzed and is chosen with the kinematic constraint of pedestrian；

(4), selected Optical-flow Feature is learnt with SVM, obtains the SVM classifier that can be used for abnormal behaviour identification；

(5), the parameter of SVM classifier model is write down；

The kinematic constraint is modeled by Bayesian formula, and the Bayesian formula is:

P (u, v | I (x, y, t))=α P (u, v) Π P (I (x_i,y_i,t_i)|u,v)

Wherein, P (u, v) is prior probability, represents the two dimensional motion speed probability of object, and α is the direct proportion factor, P (u, v | I (x, Y, t)) it is its two dimensional motion speed probability under the conditions of gray value when space (x, y) is in moment t, P (I (x, y, t) | u, v) be The probability of the grey scale change of object during exercise, i.e. posterior probability；

The Optical-flow Feature in foreground is analyzed and being chosen with the kinematic constraint of pedestrian in the step (3), specifically includes Following steps：

(1), from the foreground image after background detection, Optical-flow Feature is selected from the image-region of pedestrian；

(2), go out to meet the Optical-flow Feature of threshold requirement by kinematic constraint model discrimination, the threshold value is to posterior probability P (I (x, y, t) | u, v) carries out the threshold value of Select to use.

2. utilizing the abnormal behaviour discrimination method described in claim 1 based on kinematic constraint, which is characterized in that described from frame figure Extract that Optical-flow Feature takes as in is Lucas-Kanade algorithms.

3. utilizing the abnormal behaviour discrimination method described in claim 1 based on kinematic constraint, which is characterized in that the model of the α It encloses between zero and one.

4. utilizing the abnormal behaviour discrimination method described in claim 1 based on kinematic constraint, which is characterized in that the threshold value is It is optimized during the iterative solution of SVM modeling process, obtains optimal value.

5. utilizing the abnormal behaviour discrimination method described in claim 1 based on kinematic constraint, which is characterized in that described to detect The step of journey is：

(1), second group of video is inputted, input video frame image is demarcated using quick calibrating method；

(2), background modeling is carried out to calibrated input video frame image, while light stream is extracted from frame image with optical flow method Feature；

(3), with the kinematic constraint of pedestrian, the threshold value obtained according to the training stage, to the Optical-flow Feature in foreground carry out analysis and It chooses；

(4), the SVM model parameters for obtaining training process input SVM models, are carried out to selected Optical-flow Feature with SVM models Judge.