CN102509073A

CN102509073A - Static target segmentation method based on Gauss background model

Info

Publication number: CN102509073A
Application number: CN2011103145761A
Authority: CN
Inventors: 厉鹏; 王宸昊; 王冲鶄; 刘允才
Original assignee: Shanghai Jiaotong University
Current assignee: Shanghai Jiaotong University
Priority date: 2011-10-17
Filing date: 2011-10-17
Publication date: 2012-06-20

Abstract

The invention relates to a static target segmentation method based on a Gauss background model. The method comprises the following steps of: firstly, building a Gauss mixed model according to a first frame or pre-loading background, and loading parameter thresholds of the required parts; when a new frame enters, scanning pixel by pixel and matching with a background model; if distribution which is the same with a previous frame is matched, adding one, otherwise returning to zero; then carrying out parameter updating according to a matching result, sorting updated distributions, and selecting a corresponding generated background model according to the thresholds; filtering a smaller foreground part by finding a connected domain, so as to eliminate interference of noise and complicated background, and carrying out parameter restoration pixel by pixel according to an accumulated value of a counter at the foreground part. By applying the method provided by the invention, the Gauss background model can effectively identify a static object which enters into a background for a long time, and sensitivity identification to luminosity and learning capability of the Gauss background model can be maintained.

Description

A kind of method of cutting apart based on the static object of Gaussian Background model

Technical field

The present invention relates to advanced the manufacturing and automatic field, in particular, relate to the method for the Gaussian Background model that can static foreground target segmented extraction be come out.

Background technology

In video monitoring,, usually adopt the mode that makes up background model in order to extract target object.And the most typically be exactly the Gaussian Background model algorithm (Chris Stauffer and W. Eric L. Grimson, " Learning Patterns of Activity Using Real-Time Tracking, " IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22; no. 8, pp. 747 – 757,2000. use method for real time tracking to carry out the pattern action learning); it can make up several Gaussian distribution for each pixel; mate judgement and whether belong to prospect, and brings in constant renewal in the parameter of Gaussian distribution, and then has solved the problems such as noise, illumination of background effectively.

After the Gaussian Background model proposed, it was widely used in the various video processnig algorithms.The Gaussian Background model is all very outstanding to the learning ability and the adaptive faculty of change of background, can rapidly moving object and background separation be come out immediately.Yet after target object stopped to keep static a period of time, the Gaussian Background model can be updated to this object on the background automatically, and then lose objects.This cannot accept in the monitoring field, and the pace of learning of for this reason frequently regulating the Gaussian Background model can not well address this problem.

Present a kind of method (Xiaodong Cai, Falah Ali, E. Stipidis, Background Modeling for Detecting Move-then-stop Arbitrary-long time Video Objects2009 10th WIAMIS; Detect that random time moves the background model that stops object in the video) propose to adopt the mode that stops to upgrade the Gaussian Background model according to threshold determination; Can discern the stationary object that is no more than certain hour, but unsatisfactory for the object effect that surpasses certain hour, and also filtering illumination and shade etc. change the influence that produces effectively to stop renewal itself.

Summary of the invention

The objective of the invention is to deficiency to prior art; A kind of method of cutting apart based on the static object of Gaussian Background model is proposed; Be the improvement algorithm of present Gaussian Background model, this method can solve the problem that stationary body is updated to background model and not influenced by illumination shade etc. simultaneously.

The present invention with Gaussian Background model and EM improve algorithm (P. Kaewtrakulpong, R. Bowden, An Improved Adaptive Background Mixture Model for Realtime Tracking with Shadow Detection, In Proc. 2nd European Workshop on Advanced Video Based Surveillance Systems, AVBS01; Sept. 2001, the improvement and the shadow Detection of real-time follow-up adaptive background mixture model) (the EM algorithm is Dempster, Laind for the basis; A kind of method of asking the parameter maximum likelihood to estimate that Rubin proposed in 1977, i.e. expectation-maximization algorithm), (full name is: Open Source Computer Vision Library based on the OpenCV code library; The computer vision of promptly increasing income storehouse; Constitute by a series of C functions and a small amount of C++ class, realized a lot of general-purpose algorithms of Flame Image Process and computer vision aspect), added a kind of parameter retrieving algorithm; Thereby improved the shortcoming of the incompatibility stationary body of model own, and can adapt to factors vary such as illumination effectively.

Technical scheme of the present invention: at first set up the Gaussian Background model, and IMU is crossed the EM algorithm Gaussian Background model parameter is upgraded.After upgrading end, extract the zone of removing area much smaller than the target area through connected domain; As foreground area, each pixel substitution condition judgment surpasses certain threshold value if the pixel of target area repeats to belong to unified Gaussian distribution number of times with remaining afterwards zone; Then this pixel Gaussian distribution is carried out the parameter reduction; The reduction continued is participated in the renewal of Gaussian Background model, and then can effectively be partitioned into the static object object, and can adapt to the variation of luminosity and shade.

Stationary body partitioning algorithm based on the Gaussian Background model of the present invention comprises following step:

1. first two field picture according to prestore background picture or video makes up the Gaussian Background model, is written into the corresponding initial parameter of model, storage allocation space, some structural parameters of initial model.

2. after second frame or other frames are written into, each pixel and background dot respective pixel are mated, carry out parameter update according to matching result.The parameter update formula selects basis whether to arrive the frame number decision of learning rate.The same distribution of each coupling then this multiplicity adds one, and then zero clearing does not match.

3. to the Gaussian distribution weights of each point and descending ordering of ratio of variance.According to coupling distribute corresponding weights and before weights with whether whether mate background greater than this point of threshold value decision, not matching then is the foreground point.

4. carry out connected domain scanning, filter much smaller than the foreground area of target object size.

5. foreground point and the shadow spots that obtains scanned, whether the multiplicity that detects each point surpasses threshold value, and surpass threshold value and then carry out the parameter reduction, and the multiplicity zero clearing.

The mode of reducing through the adding parameter among the present invention, thereby can avoid static prospect to be learnt to get into background.Owing to do not stop to upgrade, can adapt to luminosity and change and change comparatively complicated background model simultaneously.This not only can remedy the shortcoming that Gauss model own can't be discerned long-time static target object; And can keep Gaussian Background model itself to responsive identification of luminosity and learning ability, can be applicable to fields such as low cost is camera supervised, video behavior description, human motion analysis.

Description of drawings

Fig. 1 upgrades schematic flow sheet for adopting Gauss model of the present invention.

Embodiment

In order to understand technical scheme of the present invention better, following formula computation process combines accompanying drawing, does further to describe in detail.Specifically carry out as follows:

1. according to video first frame or be preloaded into background image and make up Gauss model; Each Color Channel of each pixel (RGB3 passage altogether) makes up K Gauss model distribution (K gets 3 ~ 5 usually); Be written into corresponding initiation parameter; Comprise learning rate win_size (being that initial back-ground model study needs frame number), matching threshold Std_threshold, weights threshold value bg_stdshold, variance initial value var_init, target Minimum Area threshold value Area_threshold, repeat counter counter threshold value counter_threshold, initialization each point Gaussian distribution model.The corresponding average of first distribution is used for the initialization background model, that is:

?

6. if this coupling Gaussian distribution of this pixel is consistent with last time, then counter adds 1, otherwise zero clearing.

Figure 2011103145761100002DEST_PATH_IMAGE009

10 Gaussian Background model modifications finish, if be written into next frame, then circulation gets into step 3 again.This moment, target can not be updated in the background model because rest time is long, and because upgrading link does not stop, can better adaptability be arranged to luminosity variation and complex background.

Claims

1. a method of cutting apart based on the static object of Gaussian Background model is characterized in that, comprises following concrete steps:

First two field picture according to prestore background picture or video makes up the Gaussian Background model, and each Color Channel of each pixel makes up K Gauss model and distributes, and is written into the corresponding initial parameter of model, comprises learning rate win_size, matching threshold Std_threshold, weights threshold value bg_stdshold, variance initial value var_init, target Minimum Area threshold value Area_threshold, repeat counter counter threshold value counter_threshold, initialization each point Gaussian distribution model, storage allocation space;

When a new frame is written into; K Gaussian distribution to each passage of each pixel and background model mated successively; And return corresponding matching result, and carrying out parameter update according to matching result, the parameter update formula selects basis whether to arrive the frame number decision of learning rate; The same distribution of each coupling then this multiplicity adds one, and then zero clearing does not match;

To the Gaussian distribution weights of each point and descending ordering of ratio of variance, according to coupling distribute corresponding weights and before weights with whether whether mate background greater than this point of threshold value decision, not matching then is the foreground point;

Connected domain scanning is carried out in confirmed foreground point in the entire image, filter, it is labeled as background parts much smaller than the foreground point of target object size;

Foreground point and background dot to identifying scan, and whether the multiplicity that detects each point surpasses threshold value, and surpass threshold value and then carry out the parameter reduction, and the multiplicity zero clearing;

Gauss model upgrade to finish, and returns above-mentioned 2) step finishes until video.

2.

。

3. the method for cutting apart based on the static object of Gaussian Background model according to claim 2; It is characterized in that said 1) in the step, for preventing the possibility of the static existence of target object in first frame; When being written into background image in advance, with this image setting for repeating Th in advance _{Pre_num}Inferior, Th _{Pre_num}Value is greater than the half value of learning rate win_size.

4.

Figure 2011103145761100001DEST_PATH_IMAGE002

。

5.

Figure 2011103145761100001DEST_PATH_IMAGE004

; If this coupling Gaussian distribution of this pixel is consistent with last time; Then multiplicity adds 1, otherwise zero clearing.

6.

Figure 2011103145761100001DEST_PATH_IMAGE006

7.

。

8. the method for cutting apart based on the static object of Gaussian Background model according to claim 7; It is characterized in that; Said 5) in the step, picture element scan is pursued in the foreground point that collects, if its corresponding multiplicity surpasses setting threshold counter_threshold; Then with the multiplicity zero clearing, and carry out the parameter reduction according to the following equation;

Other distributed areas are belonged to Gaussian distribution:

。