CN103152558A - Intrusion detection method based on scene recognition - Google Patents

Abstract

The invention provides an intrusion detection method based on scene recognition, and belongs to the technical field of intelligent video monitoring. According to the method, the problem of real-time intrusion detection of a video area under the dynamic background is effectively solved. The intrusion detection method based on the scene recognition comprises the following steps of initialization: dividing the whole video area into N*N image blocks, and calculating the mean value and the standard deviation of each image block; inputting a monitored area video image, wherein the input image is a video image which is collected by a surveillance camera in real time, and can be an image sequence consisting of a plurality of frames decomposed by a collected video file, and inputting the images one by one in sequence; and scene recognition and intrusion area analysis and processing: according to the existing normal mode of the scene, firstly, effectively recognizing and matching the current scene, and then, calculating and comparing the mode deviation of each image block with a threshold value to obtain the intruded video area so as to realize the intrusion detection of the surveillance video. The intrusion detection method based on the scene recognition is mainly used for the intrusion detection.

Description

Intrusion detection method based on scene recognition

Technical Field

The invention belongs to the technical field of intelligent video monitoring.

Background

Intrusion detection is an important component of an intelligent security system and is widely applied to monitoring and protection of key areas, such as military heavy areas, railways, museums, test fields, dangerous areas, warning areas and the like. Compared with the installation of special sensor equipment (such as infrared and sound control equipment), the intrusion detection based on the video image has the characteristics of large detection coverage, simplicity in installation, convenience in maintenance, low construction cost, wide application range and the like, so that the intrusion detection becomes a hotspot of the research of the current intrusion detection technology.

The intrusion detection based on the video images is to analyze the video image content of a monitoring scene by utilizing a computer vision technology, automatically detect abnormal conditions in a monitoring picture, alarm and provide useful information, and thus, security personnel can be effectively reminded to timely process illegal intrusion. At present, intrusion detection methods based on video images mainly include a gray scale comparison method, a background difference method, an inter-frame difference method and an optical flow method, and all of the methods realize intrusion detection and alarm functions by detecting moving objects from a video sequence. The gray scale comparison method detects a moving object by using a gray scale statistic value for a background and an object, but is very sensitive to changes of ambient light. The background difference method is used for extracting a moving target by calculating the difference value between a current input frame image and a background image, but the background image needs to be refreshed in real time, and the detection precision of the background difference method depends on the reliability of the background image to a great extent. The frame difference method is to subtract two or more adjacent frames and detect the reserved moving target information. Although the method is slightly influenced by the change of ambient light, when the shaking of the camera causes the corresponding shaking of background points of two adjacent frames, the method cannot completely filter the background, thereby causing misjudgment; in addition, this method cannot effectively detect objects that are stationary or moving at too low a speed. The optical flow method is to extract a moving target by analyzing the motion field of the image pixel point, and the method is difficult to process the target detection problem under the dynamic background. The methods can not effectively solve the intrusion detection problem under the dynamic background, such as a mobile camera, scene switching and the like, and have poor adaptability and low expansibility.

In view of this, the present invention provides an intrusion detection method based on scene recognition. The method comprises the steps of dividing the whole video scene into a plurality of sub-regions, namely image blocks, establishing various normal modes (non-invasive modes) of the scene according to the mean value and standard deviation of each image block, identifying the mode of the current scene during operation, namely matching the mode with the normal modes, and finally obtaining an invaded video region (consisting of a plurality of image blocks) by calculating the deviation of each image block in the current mode and the corresponding matching mode and comparing the threshold value, thereby realizing real-time invasion detection.

Disclosure of Invention

The invention aims to provide an intrusion detection method based on scene recognition, which can effectively realize real-time video area intrusion detection. The process automatically realizes the motion tracking of the invading target because the invading area is dynamically updated according to the invaded condition of each image block along with the input of the video stream. Therefore, the method can be used for the intrusion detection task of a fixed camera or a mobile camera regardless of a static or dynamic background, has more accurate detection and strong adaptability and expansibility, and has simple structure and easy realization.

The purpose of the invention is realized by the following technical scheme: the technical scheme comprises the following steps:

(1) initialization

Dividing the entire video area into N image blocks, the size of N being dependent onThe actual area range covered by the video is set, such as N =30. Let the mean and standard deviation of the ith image block be mu_iAnd σ_iThen, the mode Z of the entire scene is a vector formed by the mean and the standard deviation corresponding to the N × N image blocks, that is:

Z=（μ₁,σ₁,μ₂,σ₂,…,μ_i,σ_i,…,μ_N×N,σ_N×N）.

for a dynamic scene, the individual image blocks will have different mean and standard deviations under different circumstances, and thus the dynamic scene will have a plurality of different modes. Is provided with Z_kRepresenting the Kth mode of the scene, the mean value and the standard deviation of the ith image block corresponding to the mode are respectively represented as mu_k,iAnd σ_k,iAnd then:

Z_k=（μ_k,1,σ_k,1,μ_k,2,σ_k,2,…,μ_k,i,σ_k,i,…,μ_k,N×N,σ_k,N×N）.

for a fixed camera, scene changes mainly come from illumination changes (e.g., day, night, light) and background motion (e.g., tree sway); while for a moving camera the scene will change more, i.e. the scene will have more modes. Before monitoring begins, possible patterns of a scene are acquired. Specifically, for a fixed camera, representative video images are collected at a plurality of moments and under various weather conditions; for a mobile camera, when the camera moves a certain angle (for example, every 1 degree), a group of video images is acquired, and the group of images refers to representative video images acquired at multiple moments and under various weather conditions. And calculating all modes of the scene according to the obtained images, wherein the modes are normal modes of the scene when no intrusion occurs.

(2) Inputting monitoring area video image

And inputting a video image for intrusion detection, wherein the input image is a video image acquired by a monitoring camera in real time, or can be an image sequence formed by decomposing an acquired video file into a plurality of frames, and the images are input one by one according to a time sequence. If the input image is empty, the entire process is terminated.

(3) Scene recognition

The scene mode Z of the monitored image at the current time t is calculated in the same way as in the initialization^tNamely:

$Z^t=({\mathrm{\μ}}_1^t,{\mathrm{\σ}}_1^t,{\mathrm{\μ}}_2^t,{\mathrm{\σ}}_2^t,...,{\mathrm{\μ}}_i^t,{\mathrm{\σ}}_i^t,...,{\mathrm{\μ}}_{N\×N}^t,{\mathrm{\σ}}_{N\×N}^t),$

wherein,

and

respectively representing scene modes Z^tThe mean and standard deviation of the ith image block of (1). Is provided with

Representing the current scene mode Z^tThe distance from the Kth normal mode of the scene is

The calculation is as follows:

$d_K^t=\sqrt{\underset{i=1}{\overset{N\×N}{\mathrm{\Σ}}}({({\mathrm{\μ}}_i^t-{\mathrm{\μ}}_{k,i})}^2+{({\mathrm{\σ}}_i^t-{\mathrm{\σ}}_{k,i})}^2)}.$

calculating and comparing the current scene mode Z^tDistance from all normal modes of the scene, letThe sequence number of the scene normal mode corresponding to the minimum distance in all the distances is as follows:

$\hat{K}=\arg \underset{K\∈H}{\min }{d}_K^t$

wherein, H is a set of sequence numbers of all normal modes of the scene. Thus, the result of scene recognition is the first to be a sceneA normal mode $Z_{\hat{K}}=({\mathrm{\μ}}_{\hat{k},1},{\mathrm{\σ}}_{\hat{k},1}\text{,}{\mathrm{\μ}}_{\hat{k},2},{\mathrm{\σ}}_{\hat{k},2},...,{\mathrm{\μ}}_{\hat{k},i}\text{,}{\mathrm{\σ}}_{\hat{k},i},...,{\mathrm{\μ}}_{\hat{k},N\×N},{\mathrm{\σ}}_{\hat{k},N\×N})$ As the current scene Z^tA mode in which

And

respectively representing scene modes

The mean and standard deviation of the ith image block of (1).

(4) Intrusion zone analysis and processing

According to the current scene mode Z^tAnd the mode to which it belongs

And calculating the mode deviation corresponding to each image block. Let e_iE is the mode deviation corresponding to the ith image block_iThe calculation is as follows:

$e_i=\sqrt{{({\mathrm{\μ}}_i^t-{\mathrm{\μ}}_{\hat{k},i})}^2+{({\mathrm{\σ}}_i^t-{\mathrm{\σ}}_{\hat{k},i})}^2}.$

for all the N × N image blocks, if the deviation value is greater than the threshold value theta_eIf so, the image block is marked as invaded, otherwise, the image block is not invaded. Theta_eThe values may be selected and set as the application test results are specified.

Therefore, if one or more image blocks are marked to be invaded, the whole scene is considered to be invaded, otherwise, the scene is considered to be not invaded. For the situation of a fixed camera, when a scene is invaded, an invasion area (consisting of a plurality of image blocks) can be directly highlighted and alarm prompt is carried out; for the mobile camera, when the scene is invaded, the camera is firstly stopped to move, and then the invasion area is highlighted and an alarm is given. With the input of the video stream, the intrusion area is dynamically updated according to the intrusion condition of each image block, so the process automatically realizes the motion tracking of the intrusion object.

And (4) if the intrusion detection is continued, jumping to (2), otherwise, stopping the whole process.

After the processing of the steps (1) to (4), according to the existing normal mode of the scene, the current scene is firstly effectively identified and matched, and then the mode deviation of each image block is calculated and compared with the threshold value to obtain the invaded video area, so that the invasion detection of the monitoring video is realized.

Compared with the prior art, the invention has the advantages and positive effects that: the method establishes a corresponding mode for a video scene, and then identifies and matches the scene mode during operation, so as to extract the invaded video area, thereby realizing real-time invasion detection and tracking. Therefore, the method can be used for the intrusion detection task of a fixed camera or a mobile camera regardless of a static or dynamic background, has more accurate detection and strong adaptability and expansibility, and has simple structure and easy realization.

Drawings

FIG. 1 is a schematic diagram of the division of video area into image blocks according to the present invention

FIG. 2 is a technical flow chart of the present invention

Detailed Description

The invention is further described below with reference to the accompanying drawings:

the present invention can be implemented according to the logic procedures of the technical flow chart and the steps described in the summary of the invention. The method of the invention can be used in various occasions of intrusion detection under video images. Before the method is used for intrusion detection, various normal modes of possible scenes are established or a video area is divided into image blocks by processing representative scene images at different moments and under different conditions; the video monitoring camera is arranged at a proper position, so that the video range can cover a required monitoring area; then, a suitable video transmission means is adopted, such as a wired or wireless mode, video images collected by a camera in real time are extracted in the intrusion detection process, then scene identification and matching are carried out on the images according to the method of the invention, and finally, the regional intrusion detection result is obtained by carrying out intrusion analysis on each image block in the scene. The method can be used for the intrusion detection task of a fixed camera or a mobile camera regardless of a static or dynamic background, has more accurate detection and strong adaptability and expansibility, and has simple structure and easy realization.

The technical scheme of the invention comprises the following steps:

(1) initialization

The whole video area is divided into N × N image blocks, where the size of N may be set according to the actual area covered by the video, such as N =30. Let the mean and standard deviation of the ith image blockIs mu_iAnd σ_iThen, the mode Z of the entire scene is a vector formed by the mean and the standard deviation corresponding to the N × N image blocks, that is:

Z=（μ₁,σ₁,μ₂,σ₂,…,μ_i,σ_i,…,μ_N×N,σ_N×N）.

for a dynamic scene, the individual image blocks will have different mean and standard deviations under different circumstances, and thus the dynamic scene will have a plurality of different modes. Is provided with Z_kRepresenting the Kth mode of the scene, the mean value and the standard deviation of the ith image block corresponding to the mode are respectively represented as mu_k,iAnd σ_k,iAnd then:

Z_k=（μ_k,1,σ_k,1,μ_k,2,σ_k,2,…,μ_k,i,σ_k,i,…,μ_k,N×N,σ_k,N×N）.

for a fixed camera, scene changes mainly come from illumination changes (e.g., day, night, light) and background motion (e.g., tree sway); while for a moving camera the scene will change more, i.e. the scene will have more modes. Before monitoring begins, possible patterns of a scene are acquired. Specifically, for a fixed camera, representative video images are collected at a plurality of moments and under various weather conditions; for a mobile camera, when the camera moves a certain angle (for example, every 1 degree), a group of video images is acquired, and the group of images refers to representative video images acquired at multiple moments and under various weather conditions. And calculating all modes of the scene according to the obtained images, wherein the modes are normal modes of the scene when no intrusion occurs.

(2) Inputting monitoring area video image

And inputting a video image for intrusion detection, wherein the input image is a video image acquired by a monitoring camera in real time, or can be an image sequence formed by decomposing an acquired video file into a plurality of frames, and the images are input one by one according to a time sequence. If the input image is empty, the entire process is terminated.

(3) Scene recognition

The scene mode Z of the monitored image at the current time t is calculated in the same way as in the initialization^tNamely:

$Z^t=({\mathrm{\μ}}_1^t,{\mathrm{\σ}}_1^t,{\mathrm{\μ}}_2^t,{\mathrm{\σ}}_2^t,...,{\mathrm{\μ}}_i^t,{\mathrm{\σ}}_i^t,...,{\mathrm{\μ}}_{N\×N}^t,{\mathrm{\σ}}_{N\×N}^t),$

wherein,

and

respectively representing scene modes Z^tThe mean and standard deviation of the ith image block of (1). Is provided withRepresenting the current scene mode Z^tThe distance from the Kth normal mode of the scene is

The calculation is as follows:

$d_K^t=\sqrt{\underset{i=1}{\overset{N\×N}{\mathrm{\Σ}}}({({\mathrm{\μ}}_i^t-{\mathrm{\μ}}_{k,i})}^2+{({\mathrm{\σ}}_i^t-{\mathrm{\σ}}_{k,i})}^2)}.$

calculating and comparing the current scene mode Z^tDistance from all normal modes of the scene, let

The sequence number of the scene normal mode corresponding to the minimum distance in all the distances is as follows:

$\hat{K}=\arg \underset{K\∈H}{\min }{d}_K^t$

wherein, H is a set of sequence numbers of all normal modes of the scene. Thus, the result of scene recognition is the first to be a scene

A normal mode $Z_{\hat{K}}=({\mathrm{\μ}}_{\hat{k},1},{\mathrm{\σ}}_{\hat{k},1}\text{,}{\mathrm{\μ}}_{\hat{k},2},{\mathrm{\σ}}_{\hat{k},2},...,{\mathrm{\μ}}_{\hat{k},i}\text{,}{\mathrm{\σ}}_{\hat{k},i},...,{\mathrm{\μ}}_{\hat{k},N\×N},{\mathrm{\σ}}_{\hat{k},N\×N})$ As the current scene Z^tA mode in which

Andrespectively representing scene modes

The mean and standard deviation of the ith image block of (1).

(4) Intrusion zone analysis and processing

According to the current scene mode Z^tAnd the mode to which it belongsAnd calculating the mode deviation corresponding to each image block. Let e_iE is the mode deviation corresponding to the ith image block_iThe calculation is as follows:

$e_i=\sqrt{{({\mathrm{\μ}}_i^t-{\mathrm{\μ}}_{\hat{k},i})}^2+{({\mathrm{\σ}}_i^t-{\mathrm{\σ}}_{\hat{k},i})}^2}.$

for all the N × N image blocks, if the deviation value is greater than the threshold value theta_eThen mark the image block as invadedOtherwise, no intrusion. Theta_eThe values may be selected and set as the application test results are specified.

Therefore, if one or more image blocks are marked to be invaded, the whole scene is considered to be invaded, otherwise, the scene is considered to be not invaded. For the situation of a fixed camera, when a scene is invaded, an invasion area (consisting of a plurality of image blocks) can be directly highlighted and alarm prompt is carried out; for the mobile camera, when the scene is invaded, the camera is firstly stopped to move, and then the invasion area is highlighted and an alarm is given. With the input of the video stream, the intrusion area is dynamically updated according to the intrusion condition of each image block, so the process automatically realizes the motion tracking of the intrusion object.

And (4) if the intrusion detection is continued, jumping to (2), otherwise, stopping the whole process.

After the processing of the steps (1) to (4), according to the existing normal mode of the scene, the current scene is firstly effectively identified and matched, and then the mode deviation of each image block is calculated and compared with the threshold value to obtain the invaded video area, so that the invasion detection of the monitoring video is realized.

The method can be realized by programming in any computer programming language (such as C language), and the system software realized based on the method can realize real-time regional intrusion detection application in any PC or embedded system.

Claims (2)

1. An intrusion detection method based on scene recognition comprises the following steps:

(1) initialization

Dividing the whole video area into N × N image blocks, calculating the mean value and standard deviation of each image block according to the pixel brightness value of the image, and setting the mean value and standard deviation of the ith image block as mu_iAnd σ_iThen, the mode Z of the entire scene is a vector formed by the mean and the standard deviation corresponding to the N × N image blocks, that is:

Z=（μ₁,σ₁,μ₂,σ₂,…,μ_i,σ_i,…,μ_N×N,σ_N×N）.

for a dynamic scene, each image block will have different mean and standard deviation under different conditions, so the dynamic scene will have a plurality of different modes, let Z_kRepresenting the Kth mode of the scene, the mean value and the standard deviation of the ith image block corresponding to the mode are respectively represented as mu_k,iAnd σ_k,iAnd then:

Z_k=（μ_k,1,σ_k,1,μ_k,2,σ_k,2,…,μ_k,i,σ_k,i,…,μ_k,N×N,σ_k,N×N）.

before monitoring starts, acquiring possible modes of a scene, specifically, acquiring representative video images of a fixed camera at a plurality of moments and under various weather conditions;

(2) inputting monitoring area video image

Inputting a video image for intrusion detection, wherein the input image is a video image acquired by a monitoring camera in real time, or an image sequence formed by decomposing an acquired video file into a plurality of frames, and inputting the images one by one according to a time sequence; if the input image is empty, the whole process is stopped;

(3) scene recognition

The scene mode Z of the monitored image at the current time t is calculated in the same way as in the initialization^tNamely:

$Z^t=({\mathrm{\μ}}_1^t,{\mathrm{\σ}}_1^t,{\mathrm{\μ}}_2^t,{\mathrm{\σ}}_2^t,...,{\mathrm{\μ}}_i^t,{\mathrm{\σ}}_i^t,...,{\mathrm{\μ}}_{N\×N}^t,{\mathrm{\σ}}_{N\×N}^t),$

wherein,

andrespectively representing scene modes Z^tThe mean and standard deviation of the ith image block of (1); is provided with

Representing the current scene mode Z^tThe distance from the Kth normal mode of the scene is

The calculation is as follows:

$d_K^t=\sqrt{\underset{i=1}{\overset{N\×N}{\mathrm{\Σ}}}({({\mathrm{\μ}}_i^t-{\mathrm{\μ}}_{k,i})}^2+{({\mathrm{\σ}}_i^t-{\mathrm{\σ}}_{k,i})}^2)}.$

calculating and comparing the current scene mode Z^tDistance from all normal modes of the scene, let

The sequence number of the scene normal mode corresponding to the minimum distance in all the distances is as follows:

$\hat{K}=\arg \underset{K\∈H}{\min }{d}_K^t$

h is a set of sequence numbers of all normal modes of a scene; thus, the result of scene recognition is the first to be a scene

A normal mode $Z_{\hat{K}}=({\mathrm{\μ}}_{\hat{k},1},{\mathrm{\σ}}_{\hat{k},1}\text{,}{\mathrm{\μ}}_{\hat{k},2},{\mathrm{\σ}}_{\hat{k},2},...,{\mathrm{\μ}}_{\hat{k},i}\text{,}{\mathrm{\σ}}_{\hat{k},i},...,{\mathrm{\μ}}_{\hat{k},N\×N},{\mathrm{\σ}}_{\hat{k},N\×N})$ As the current scene Z^tA mode in which

And

respectively representing scene modes

The mean and standard deviation of the ith image block of (1);

(4) intrusion zone analysis and processing

According to the current scene mode Z^tAnd the mode to which it belongsCalculating the mode deviation corresponding to each image block, and setting e_iE is the mode deviation corresponding to the ith image block_iThe calculation is as follows:

$e_i=\sqrt{{({\mathrm{\μ}}_i^t-{\mathrm{\μ}}_{\hat{k},i})}^2+{({\mathrm{\σ}}_i^t-{\mathrm{\σ}}_{\hat{k},i})}^2}.$

for all the N × N image blocks, if the deviation value is greater than the threshold value theta_eIf so, marking the image block as invaded, otherwise, marking the image block as not invaded by theta_eThe value can be selected and set according to the application test result according to specific conditions;

after the processing of the steps (1) to (4), according to the existing normal mode of the scene, the current scene is firstly effectively identified and matched, and then the mode deviation of each image block is calculated and compared with the threshold value to obtain the invaded video area, so that the invasion detection of the monitoring video is realized.

2. The intrusion detection method based on scene recognition according to claim 1, wherein: for a mobile camera, when the camera moves a certain angle, a group of video images are collected, and the images are representative video images collected at a plurality of moments and under various weather conditions; and calculating each mode of the scene according to the obtained images, wherein the modes are normal modes of the scene when no intrusion occurs.

Images