CN103152558B - Based on the intrusion detection method of scene Recognition - Google Patents

Abstract

The invention provides a kind of intrusion detection method based on scene Recognition, belong to Intelligent Video Surveillance Technology field.The method efficiently solves the problem of the real-time intrusion detection in video area under dynamic background.The present invention includes following steps: initialization: whole video area is divided into N × N number of image block, and calculate average and the standard deviation of each image block.Input monitoring area video image: the image of input is the video image obtained by monitoring camera Real-time Collection, also can be decomposed into the image sequence that multiple frame forms, according to time sequencing input picture one by one by the video file gathered; Scene Recognition and invasion regional analysis and process: according to the existing normal mode of scene, first current scene is effectively identified and is mated, then by calculating the model deviation of each image block and comparing with threshold value, obtain invaded video area, thus realize the intrusion detection to monitor video.Be mainly used in intrusion detection.

Description

Based on the intrusion detection method of scene Recognition

Technical field

The invention belongs to Intelligent Video Surveillance Technology field.

Background technology

Intrusion detection is the important component part of intelligent safety and defence system, is widely used in the supervision to key area and protection, as place of military importance, and railway, museum, test site, hazardous area, security area etc.Compared with the sensor device that installation is special (as infrared ray, voice-operated device), it is large that the intrusion detection based on video image has detection coverage, installs simple, easy to maintenance, the features such as project cost is low, widely applicable, thus become the focus of current Research of Intrusion Detection Technology.

Intrusion detection based on video image utilizes the video image content of computer vision technique to monitoring scene to analyze, abnormal conditions in automatic test and monitoring picture, and alarm and provide useful information, thus security protection personnel can be more effectively reminded to process illegal invasion in time.At present, the intrusion detection method based on video image mainly contains based on gray scale comparison method, Background difference, frame-to-frame differences method and optical flow method, and these methods are all by detecting that from video sequence moving target realizes intrusion detection and warning function.Gray scale comparison method adopts and detects moving target to the gray-scale statistical value of background and target, but it is very responsive to the change of ambient light.Background difference is by calculating the difference of present incoming frame image and background image to extract moving target, but background image needs to refresh in real time, and its accuracy of detection depends on the reliability of background image to a great extent.Frame-to-frame differences method is subtracted each other adjacent two frames or multiframe, detects the moving target information retained.Although the method is little by surround lighting variable effect, when camera shake and cause corresponding " shake " of adjacent two frame background dots time, the method completely by background filtering, thus can not cause erroneous judgement; In addition, the method, for static or that movement velocity is excessively slow target, can not effectively detect.Optical flow method is by analyzing the sports ground of image slices vegetarian refreshments, and then extracts moving target, and the method is difficult to the target detection problems under process dynamic background equally.These methods all effectively cannot solve the intrusion detection problem under dynamic background above, as the camera of movement, and scene switching etc., and adaptive capacity is poor, autgmentability is not strong.

Given this, the present invention proposes a kind of intrusion detection method based on scene Recognition.Whole video scene is divided into multiple subregion by the method, i.e. image block, then the various normal modes (non-intrusive pattern) of scene are set up according to the average of each image block and standard deviation, then operationally the pattern of current scene is identified, mate with these normal modes by it, finally by calculating the deviation of each image block in present mode and Corresponding matching pattern and compare threshold, obtain invaded video area (being made up of multiple image block), thus realize intrusion detection in real time.

Summary of the invention

The object of this invention is to provide a kind of intrusion detection method based on scene Recognition, it can realize real-time video area intrusion detection effectively.Due to the input along with video flowing, this invasion region dynamically will be upgraded according to the invaded situation of each image block, and therefore this process automatically achieves the motion tracking to intrusion target.Therefore, method of the present invention may be used for being no matter static or dynamic background, the intrusion detection task of fixing camera or dollying head, and it is more accurate not only to detect, adapt to and extended capability strong, and structure is simple, is easy to realization.

The object of the invention is to be achieved through the following technical solutions: described technical scheme comprises the steps:

(1) initialization

Whole video area is divided into N × N number of image block, the actual area scope that the large I of N covers according to video is arranged, and if N=30. is according to the pixel brightness value of image, calculates average and the standard deviation of each image block.If average and the standard deviation of i-th image block are respectively μ _iand σ _i, then the pattern Z of whole scene is the vector of average that this N × N number of image block is corresponding and standard deviation composition, that is:

Z=（μ ₁,σ ₁,μ ₂,σ ₂,…,μ _i,σ _i,…,μ _N×N,σ _N×N）.

For dynamic scene, each image block will have different averages and standard deviation in varied situations, and therefore dynamic scene will have multiple different pattern.If Z _krepresent K pattern of scene, the average of i-th image block that this pattern is corresponding and standard deviation are expressed as μ _k,iand σ _k,i, 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 fixing camera, the change of scene is mainly from illumination variation (as daytime, night, light) and background motion (as trees wave); And for the camera of movement, the change of scene will be more, and namely scene will have more pattern.Before monitoring starts, obtain the pattern that scene is possible.Be specially, for fixing camera, under multiple moment and various weather condition, gather representative video image; For dollying head, then gather one group of video image when camera often moves certain angle (as often mobile 1 degree), this group image refers to the representative video image collected under multiple moment and various weather condition.According to the image obtained above, calculate each pattern of scene, these patterns are the normal mode of scene when not yet having invasion to occur.

(2) input monitoring area video image

Carry out the video image input of intrusion detection, the image of input is the video image obtained by monitoring camera Real-time Collection, also can be decomposed into the image sequence that multiple frame forms by the video file gathered, the image inputted one by one according to time sequencing.If input picture is empty, then whole Flow ends.

(3) scene Recognition

According to the method identical with initialization, calculate the scene mode Z of current time t monitoring image ^t, that is:

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

Wherein, with represent scene mode Z respectively ^tthe average of i-th image block and standard deviation.If represent current scene pattern Z ^twith the distance of scene K normal mode, then be calculated as:

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

Calculate and compare current scene pattern Z ^twith the distance of all normal modes of scene, if for the sequence number of scene normal mode corresponding to minimum range in all distances, that is:

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

Wherein, H is the set of all normal mode sequence numbers of scene.Therefore, the result of scene Recognition is, by of scene individual 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 current scene Z ^taffiliated pattern, wherein with represent scene mode respectively the average of i-th image block and standard deviation.

(4) regional analysis and process is invaded

According to current scene pattern Z ^twith the pattern belonging to it , calculate the model deviation that each image block is corresponding.If e _ibe the model deviation that i-th image block is corresponding, then e _ibe calculated as:

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

To all N × N number of image block, if its deviate is greater than threshold value θ _e, then this image block has been labeled as invasion, otherwise has been not invasion.θ _evalue can be selected according to application testing result as the case may be and be arranged.

Therefore, if having one and above image block to be marked as invasion, then think that whole scene is invaded state, otherwise think that scene is not invaded.For fixing camera situation, directly can highlight invasion region (being made up of multiple image block) when scene is invaded and carry out alarm; For dollying head, when scene is invaded, camera first should be stopped to move, then highlight invasion region and report to the police.Along with the input of video flowing, invasion region dynamically will be upgraded according to the invaded situation of each image block, and therefore this process automatically achieves the motion tracking to intrusion target.

If proceed intrusion detection, then jump to (2), otherwise whole Flow ends.

Method of the present invention through more than (1) ~ (4) process after, according to the existing normal mode of scene, first current scene is effectively identified and is mated, then by calculating the model deviation of each image block and comparing with threshold value, obtain invaded video area, thus realize the intrusion detection to monitor video.

The present invention's advantage compared with prior art and good effect: the method is by setting up corresponding pattern to video scene, then identification and the coupling of scene mode is operationally carried out, and then extract invaded video area, thus realize intrusion detection in real time and tracking.Therefore, method of the present invention can be used for being no matter static or dynamic background, the intrusion detection task of fixing camera or dollying head, and it is more accurate not only to detect, adapt to and extended capability strong, and structure is simple, is easy to realization.

Accompanying drawing explanation

Fig. 1 is that video area of the present invention is divided into image block schematic diagram

Fig. 2 is the technology of the present invention flow chart

Embodiment

With reference to the accompanying drawings the present invention is described further below:

Step described by the logical process of the technology of the present invention flow chart and summary of the invention can implement the present invention.Method of the present invention can be used for the various occasions of intrusion detection under video image.Before for intrusion detection, first by the scene image that process is not representative under different condition in the same time, set up the possible various normal mode of scene or video area is divided into image block; And video monitoring camera is installed in place, makes range of video can cover required guarded region; Then suitable transmission of video means are adopted, as wired or wireless mode, the video image of camera Real-time Collection is extracted in intrusion detection process, then to it, scene Recognition and coupling are carried out according to method of the present invention to these images, carry out Intrusion analysis finally by image block each in scene, obtain region intrusion detection result.According to method of the present invention, can be used for being no matter static or dynamic background, the intrusion detection task of fixing camera or dollying head, it is more accurate not only to detect, adapt to and extended capability strong, and structure is simple, is easy to realization.

Technical solution of the present invention comprises the steps:

(1) initialization

Whole video area is divided into N × N number of image block, the actual area scope that the large I of N covers according to video is arranged, and if N=30. is according to the pixel brightness value of image, calculates average and the standard deviation of each image block.If average and the standard deviation of i-th image block are respectively μ _iand σ _i, then the pattern Z of whole scene is the vector of average that this N × N number of image block is corresponding and standard deviation composition, that is:

Z=（μ ₁,σ ₁,μ ₂,σ ₂,…,μ _i,σ _i,…,μ _N×N,σ _N×N）.

For dynamic scene, each image block will have different averages and standard deviation in varied situations, and therefore dynamic scene will have multiple different pattern.If Z _krepresent K pattern of scene, the average of i-th image block that this pattern is corresponding and standard deviation are expressed as μ _k,iand σ _k,i, 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 fixing camera, the change of scene is mainly from illumination variation (as daytime, night, light) and background motion (as trees wave); And for the camera of movement, the change of scene will be more, and namely scene will have more pattern.Before monitoring starts, obtain the pattern that scene is possible.Be specially, for fixing camera, under multiple moment and various weather condition, gather representative video image; For dollying head, then gather one group of video image when camera often moves certain angle (as often mobile 1 degree), this group image refers to the representative video image collected under multiple moment and various weather condition.According to the image obtained above, calculate each pattern of scene, these patterns are the normal mode of scene when not yet having invasion to occur.

(2) input monitoring area video image

Carry out the video image input of intrusion detection, the image of input is the video image obtained by monitoring camera Real-time Collection, also can be decomposed into the image sequence that multiple frame forms by the video file gathered, the image inputted one by one according to time sequencing.If input picture is empty, then whole Flow ends.

(3) scene Recognition

According to the method identical with initialization, calculate the scene mode Z of current time t monitoring image ^t, that is:

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

Wherein, with represent scene mode Z respectively ^tthe average of i-th image block and standard deviation.If represent current scene pattern Z ^twith the distance of scene K normal mode, then be calculated as:

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

Calculate and compare current scene pattern Z ^twith the distance of all normal modes of scene, if for the sequence number of scene normal mode corresponding to minimum range in all distances, that is:

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

Wherein, H is the set of all normal mode sequence numbers of scene.Therefore, the result of scene Recognition is, by of scene individual 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 current scene Z ^taffiliated pattern, wherein with represent scene mode respectively the average of i-th image block and standard deviation.

(4) regional analysis and process is invaded

According to current scene pattern Z ^twith the pattern belonging to it calculate the model deviation that each image block is corresponding.If e _ibe the model deviation that i-th image block is corresponding, then e _ibe calculated as:

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

To all N × N number of image block, if its deviate is greater than threshold value θ _e, then this image block has been labeled as invasion, otherwise has been not invasion.θ _evalue can be selected according to application testing result as the case may be and be arranged.

Therefore, if having one and above image block to be marked as invasion, then think that whole scene is invaded state, otherwise think that scene is not invaded.For fixing camera situation, directly can highlight invasion region (being made up of multiple image block) when scene is invaded and carry out alarm; For dollying head, when scene is invaded, camera first should be stopped to move, then highlight invasion region and report to the police.Along with the input of video flowing, invasion region dynamically will be upgraded according to the invaded situation of each image block, and therefore this process automatically achieves the motion tracking to intrusion target.

If proceed intrusion detection, then jump to (2), otherwise whole Flow ends.

Method of the present invention through more than (1) ~ (4) process after, according to the existing normal mode of scene, first current scene is effectively identified and is mated, then by calculating the model deviation of each image block and comparing with threshold value, obtain invaded video area, thus realize the intrusion detection to monitor video.

The inventive method is by any computer programming language (as C language) programming realization, and the systems soft ware realized based on the inventive method can realize real-time region intrusion detection application in any PC or embedded system.

Claims (1)

1., based on an intrusion detection method for scene Recognition, comprise the steps:

(1) initialization

Whole video area is divided into N × N number of image block, according to the pixel brightness value of image, calculates average and the standard deviation of each image block, if the average of i-th image block and standard deviation are respectively μ _iand σ _i, then the pattern Z of whole scene is the vector of average that this N × N number of image block is corresponding and standard deviation composition, that is:

Z＝(μ ₁,σ ₁,μ ₂,σ ₂,…,μ _i,σ _i,…,μ _N×N,σ _N×N).

For dynamic scene, each image block will have different averages and standard deviation in varied situations, and therefore dynamic scene is by having multiple different pattern, if Z _krepresent K pattern of scene, the average of i-th image block that this pattern is corresponding and standard deviation are expressed as μ _k,iand σ _k,i, 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, obtain scene mode, concrete grammar is: 1) for fixing camera, under multiple moment and various weather condition, gather representative video image; 2) for dollying head, when camera often moves certain angle, then gather one group of video image, this image is the representative video image collected under multiple moment and various weather condition; According to the image obtained above, calculate each pattern of scene, this pattern is the normal mode of scene when not yet having invasion to occur;

(2) input monitoring area video image

Carry out the video image input of intrusion detection, the image of input is the video image obtained by monitoring camera Real-time Collection, also can be decomposed into the image sequence that multiple frame forms by the video file gathered, according to time sequencing input picture one by one; If input picture is empty, then whole Flow ends;

(3) scene Recognition

According to the method identical with initialization, calculate the scene mode Z of current time t monitoring image ^t, that is:

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

Wherein, with represent scene mode Z respectively ^tthe average of i-th image block and standard deviation; If represent current scene pattern Z ^twith the distance of scene K normal mode, then be calculated as:

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

Calculate and compare current scene pattern Z ^twith the distance of all normal modes of scene, if for the sequence number of scene normal mode corresponding to minimum range in all distances, that is:

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

Wherein, H is the set of all normal mode sequence numbers of scene; Therefore, the result of scene Recognition is, by of scene individual normal mode $Z_{\hat{K}}=({\mathrm{\μ}}_{\hat{k},1},{\mathrm{\σ}}_{\hat{k},1},{\mathrm{\μ}}_{\hat{k},2},{\mathrm{\σ}}_{\hat{k},2},...,{\mathrm{\μ}}_{\hat{k},i},{\mathrm{\σ}}_{\hat{k},i},...,{\mathrm{\μ}}_{\hat{k},N\×N},{\mathrm{\σ}}_{\hat{k},N\×N})$ As current scene Z ^taffiliated pattern, wherein with represent scene mode respectively the average of i-th image block and standard deviation;

(4) regional analysis and process is invaded

According to current scene pattern Z ^twith the pattern belonging to it calculate the model deviation that each image block is corresponding, if e _ibe the model deviation that i-th image block is corresponding, then e _ibe calculated as:

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

To all N × N number of image block, if its deviate is greater than threshold value θ _e, then this image block has been labeled as invasion, otherwise has been not invasion, θ _evalue can be selected according to application testing result as the case may be and be arranged;

Through more than (1) ~ (4) process after, according to the existing normal mode of scene, first current scene is effectively identified and is mated, then by calculating the model deviation of each image block and comparing with threshold value, obtain invaded video area, thus realize the intrusion detection to monitor video.