CN106709906B - A kind of method of background difference and prospect detection - Google Patents

Abstract

The invention discloses a kind of methods that background difference and prospect detect, first building minot vector, for saving the reference data and statistical information of each pixel；The dynamic of each position of background image is defined, representing dynamic level of pixel series is set according to representing dynamic level of pixel division rule and dynamic is attributed to a certain representing dynamic level of pixel；Optimum bandwidth is obtained using dynamic KDE algorithm；An adaptive threshold value is designed to each pixel, the value of threshold value and the dynamic of pixel are related；Elimination operation is carried out to the reference data bottom in minot vector using Tetris update scheme, to eliminate noise jamming, improves the quality of reference data.The experimental results showed that validity of the new method in prospect detection application is compared with fresh approach, have in three indexs one better than all methods, other two slightly below fresh approach.

Description

A kind of method of background difference and prospect detection

Technical field

The invention belongs to technical field of computer vision more particularly to a kind of background difference and prospect inspections in precise and high efficiency The method for surveying modeling.

Background technique

In computer vision application, accurate background difference model is a basic and important step of foreground extraction Suddenly.Under normal conditions, before observing and assessing an image to be sorted, by gathering and studying a series of continuous images Produce the background model of initialization.But this method for choosing a fixed reference window may play static background Effect, but ignores the research to the dynamic background with dynamic variation range.For the size of fixed reference window, setting one A upper limit value can to avoid the omission to dynamic background, but will increase computation burden especially and have complicated calculations ingredient (such as Gauss).It being found by carrying out analysis to the background for having dynamic, dynamic part tends to take up a part of entire background, or It says, entire background can be divided into different ranks using dynamic.The difference of sound rank according to belonging to region and set difference Threshold value, dynamic area, such as the leaf waved, the ripples etc. to ripple, in face of a new pixel to determine its whether be When background, having one, the range that can be received is bigger than wider tolerance, so as to be set as biggish threshold value； And whether the variation range of static region (also referred to as stability region) is relatively narrow, receiving to seem when a new pixel is background Compare with caution, usually only very close to its scope of subject when can just be considered, be otherwise classified as prospect.Therefore, at this time Threshold value fully taking into account actual situation, i.e., sometimes beyond a small range it is also possible, it should be set as lesser Value, to meet the fluctuation of stable background area once in a while.

In past ten years, a stable background model how is kept, it is accurate to establish characterization either statically or dynamically scene The appearance (being often referred to background difference) of middle object has attracted the concern of numerous researchers.Initially, the principle of background difference model is The difference (document 1) between two picture frames is simply calculated, one of frame is simple background image, another is then worked as Previous frame.But when encountering background variation mistake can occur for this simple calculate.Therefore, in face of the uncertainty of data set, Researchers propose based on statistical technology come analysis background model.Wren et al. has used a single Gauss to come to back Scape models (document 2), it is assumed that scene is relatively stationary Pfinder.But Stauffer and Grimson et al. have found Pfinder does not work (document 3) for outdoor scene, they model pixel value using mixed Gaussian, and this method is to rear The many improved methods come are with milestone formula meaning (document 4-8).These improved methods obtain in many scenes Many progress.However, researcher often can not be by the method for parametrization come the feature modeling to natural scene.Elgammal Et al. select a normal function as kernel function (document 9), then by assuming that bandwidth be diagonal matrix reduce calculate it is general Rate density P_rComplexity, and assess by mean absolute deviation the Continuous Gray Scale value of sampled pixel.For one in t The pixel at quarter, they give P_r(x_t) it is provided with a global threshold th.If P_r(x_t) < th, the pixel be considered as prospect otherwise For background.Elgammal et al. combines two kinds of update mechanisms (selection updates and blind update) for more new model, and calculates The intersection of testing result come eliminate mistake prospect.Mittall and Paragios (document 10) et al. define one by spherical shape Estimate the hybrid density estimator of operator and sample point estimation operator (document 11) composition.In assorting process, they are used A kind of statistical approximation method (such as method of sampling) obtains the adaptive threshold of an increment to realize lower mistake warning With high verification and measurement ratio.Yaser Sheikh and Mubarak Shah et al. thinks (document 12), spatially the correlation of neighborhood pixels Property be it is critically important, they explicitly model background and prospect, and this method achieves dynamic element such as camera vibration, ripples etc. The testing result of high quality.In addition to these statistical methods, researcher also proposed a lot of other methods, such as optical flow method (document 13), super-pixel method (document 14), dynamic texture send out (document 15,16), binary pattern (document 17,18), lasso trick method (document 19) etc..

Document 1:R.Jain and H.Nagel, " On the Analysis of Accumulative Difference Pictures from Image Sequences of Real World Scenes,”IEEE Trans.Pattern Analysis and Machine Intelligence,1979.N.Oliver,B.Rosario,and A.Pentland,“A Bayesian Computer Vision System for Modeling Human Interactions,”IEEE Trans.Pattern Analysis and Machine Intelligence,2000.

Document 2:Wren C., Azarbayejani A., Darrelll t., Pentland A.Pfinder:Real- Time Tracking of the Human Body,IEEE Transactions on Pattern Analysis and Machine Intelligence,Volume 19,No.7,pages 780-785,July 1997.

Document 3:Stauffer C., Grimson W.Adaptive background mixture models for real-time rracking,CVPR 1999,pages 246-252,1999.

Document 4:Z.Zivkovic and F.van der Heijden " Efficient adaptive density estimation per image pixel for the task of background subtraction",Pattern Recognition Letters,vol.27,no.7,pages 773-780,2006.

Document 5:M.Heikkila and M.Pietikainen.A texture-based method for modeling the background and detectin moving obj ects.IEEE Transactions on Pattern Analysis and Machine Intelligence,28(4):657-662,2006.

Document 6:H.Lin, J.Chuang, T.Liu.Regularized background adaptation:a novel learning rate control scheme for Gaussian mixture modeling.IEEE Trans.Image Process.20(2011)822-836.

Document 7:X.Liu and C.Qi.Future-data driven modeling of complex backgrounds using mixture of gausians.Neurocomputing,119:439-453,2013.

Document 8:Shimada, A., Nagahara, H., Taniguchi, R.:Background modeling based on bidirectional analysis.In:CVPR,pp.1979–1986(2013).

Document 9:Elgammal A., Harwood D., Davis L.Non-parametric Model for Background Subtraction,ECCV 2000,pages 751-767,Dublin,Ireland,June 2000.

Document 10:A.Mittal and N.Paragios.Motion-based background subtraction using adaptive kernel density estimation.In IEEE International Conference on Computer Vision and Pattern Recognition,2004.

Document 11:M.P.Wand and M.C.Jones.Kernel Smoothing.Chapman and Hall, 1995.

Document 12:Sheikh Yaser, and Mubarak Shah. " Bayesian modeling of dynamic scenes for obj ect detection."IEEE transactions on pattern analysis and machine intelligence27.11(2005):1778-1792.

Document 13:Spatiotemporal Background Subtraction Using Minimum Spanning Tree and Optical Flow.Eccv.2014.

Document 14:Wand, Matt P., and M.Chris Jones.Kernel smoothing.Crc Press, 1994.,Feldman-Haber S,Keller Y.A probabilistic graph-based framework for plug-and-play multi-cue visual tracking.IEEE Transactions on Image Processing.2014May；23(5):2291-301.

Document 15:Mumtaz A, Zhang W, Chan AB.Joint motion segmentation and background estimation in dynamic scenes.InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2014(pp.368-375).

Document 16:A.B.Chan and N.Vasconcelos, " Modeling, clustering, and segmenting video with mixtures of dynamic textures,”IEEE TPAMI,2008.

Document 17:P.-L.St-Charles, G.-A.Bilodeau, and R.Bergevin.SuBSENSE:A universal change detection method with local adaptive sensitivity.IEEE Transactions on Image Processing,24(1):359–373,2015.

Document 18:Guo, Lili, Dan Xu, and ZhenpingQiang. " Background Subtraction Using Local SVD Binary Pattern."In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops,pp.86-94.2016.

Document 19:background subtraction via generalized fused lasso foreground modeling.Cvpr.2015.

Summary of the invention

For the existing problem not high in accuracy based on the model of statistical method, matter of the present invention from reference data Amount, the calculating of optimum bandwidth, 4 points of the selection of adaptive threshold and more new strategy set out, propose a kind of new background difference with The method of prospect detection.

The technical scheme adopted by the invention is that: a kind of method of background difference and prospect detection, it is characterised in that: building Minot vector, for saving the reference data and statistical information of each pixel；Define the dynamic of each position of background image Property, representing dynamic level of pixel series is set according to representing dynamic level of pixel division rule, and dynamic is attributed to a certain representing dynamic level of pixel；Using dynamic KDE algorithm obtains optimum bandwidth；An adaptive threshold value, the value of threshold value and the dynamic of pixel are designed to each pixel It is related；Using Tetris update scheme, elimination operation is carried out to the reference data bottom in minot vector, to eliminate Noise jamming.

Compared with prior art, the present invention has innovation below and advantage:

The invention proposes a new background difference modeling methods based on Density Estimator.Devise an originality Data structure -- minot vector (MV), for saving the reference data and statistical information of each pixel, which is promoted Computational efficiency.It proposes the definition of dynamic and assigns this attribute to each pixel according to the variation degree of gray value.? When probability density calculates, dynamic KDE algorithm is used and realizes for the first time, this makes accuracy be improved.Different from other One global threshold algorithm of setting of document devises an adaptive threshold value, the value and picture of threshold value to each pixel The dynamic of element is related.The more new strategy of a completely new background model is proposed, to the most bottom of the reference data in minot vector Layer carries out elimination operation, to eliminate noise jamming.Compared with current method, the method effectively improves background differential mode The efficiency and accuracy rate of type.

Detailed description of the invention

Fig. 1 is the flow diagram of the embodiment of the present invention.

The square matrix schematic diagram that the minot vector proposed in Fig. 2 embodiment of the present invention is constituted.

The schematic diagram of dynamic background and static background is separated in Fig. 3 embodiment of the present invention.

Threshold calculations schematic diagram in Fig. 4 embodiment of the present invention.

Background model more new algorithm schematic diagram in Fig. 5 embodiment of the present invention.

The representative sub-scene video schematic diagram of one extracted in six kinds of classifications in Fig. 6 embodiment of the present invention.

Specific embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, with reference to the accompanying drawings and embodiments to this hair It is bright to be described in further detail, it should be understood that implementation example described herein is merely to illustrate and explain the present invention, not For limiting the present invention.

Referring to Fig.1, the method for a kind of background difference provided by the invention and prospect detection, constructs minot vector, for protecting Deposit the reference data and statistical information of each pixel；The dynamic of each position of background image is defined, representing dynamic level of pixel is set Dynamic is attributed to a certain representing dynamic level of pixel according to representing dynamic level of pixel division rule by series；It is obtained using dynamic KDE algorithm optimal Bandwidth；An adaptive threshold value is designed to each pixel, the value of threshold value and the dynamic of pixel are related；Utilize Russia Square update scheme carries out elimination operation to the reference data bottom in minot vector, to eliminate noise jamming.

In this implementation construct minot vector, specific implementation the following steps are included:

Step A1: the one-dimensional vector that building length is 263 is as minot vector；

Step A2: for each position of background image, minot vector is designed as depositing from the unit of subscript 0 to 255 Storing up gray value is corresponding lower target quantity；

Step A3: minot vector is designed as to the statistics of the storage each position of background image from the unit of subscript 256 to 262 Information, wherein the sum (total) of the storage gray value of Unit 256, Unit 257 store the variation amplitude (span of gray value Position), Unit 258 and 259 store minimum and maximum gray value (min and max) respectively, and the storage of Unit 260 gray scale is formed Cluster number (cluster), the storage of Unit 261 gray scale is formed by the number (kurtosis) of kurtosis, Unit 262 Store the sum (base) (as shown in Figure 2) of bottom gray value.

The dynamic that each position of background image is defined in the present embodiment is the gray scale of pixel in a period of time The variation range of value.Then setting representing dynamic level of pixel series is such as evenly dividing according to division rule, dynamic is attributed to a certain Representing dynamic level of pixel, the effect after classification are as shown in Figure 3.

Calculate optimum bandwidth in this implementation, specific implementation the following steps are included:

Step B1: it chooses a normal distribution and estimates as pilotAnd use observation data I=I₁,I₂,...,I_NMeter Calculate variances sigma², wherein N is the quantity for observing data；

Step B2: local bandwidth h is calculated using following formula_loc:

Wherein, iqr is quartile section；

Step B3: dynamic factor η is defined_i:

Wherein, geom isGeometric mean, α is sensitivity parameter, α ∈ [- 1,0]；

Step B4: optimum bandwidth is obtained are as follows:

H_opt=η_ih_loc(3)

See Fig. 4, an adaptive threshold value is designed to each pixel in the present embodiment, specific implementation includes following Step:

Step C1: according to formulaCalculate each observation data (I₁,I₂,...,I_m) probability density in other data(wherein, m is not repeat ash in observation data The number of angle value, K () indicate kernel function)；

Step C2: rightIt carries out ascending sort and removes duplicate keys, i.e.,m'< M × (m-1) obtains the threshold value array Thresholds [m'] that length is m'；Wherein I=I₁,I₂,...,I_m；

Step C3: the dynamic rank dr obtained when being divided according to representing dynamic level of pixel sets its initial threshold as Thresholds [dr]。

Tetris update scheme in the present embodiment, specific implementation the following steps are included:

Step D1: setting update condition is (as being equal to maximum dynamic in background image when the number of estimative picture frame Value)；

Step D2: when meeting update condition, by minot vector, the value of non-zero unit subtracts 1 from subscript 0 to 255；

Step D3: updating statistical mark position information, (Fig. 5 illustrates three kinds of situations: being (a) static region, is (b) dynamic area Domain (c) is adaptability to environmental change).

Compared below by way of the validity of qualitative and quantitative experimental verification the method for the present invention, and with newest method Compared with.

The present invention is in addition to comparing (KDE (Elgammal A., Harwood with the classical way based on statistics frame D.,Davis L.Non-parametric Model for Background Subtraction,ECCV2000,pages 751-767,Dublin,Ireland,June 2000.)、GMM(Zivkovic Z.Improved adaptive Gaussian mixture model for background subtraction[C]//Pattern Recognition,2004.ICPR 2004.Proceedings of the 17th International Conference on.IEEE,2004,2:28-31.) With PBAS (M.Hofmann, P.Tiefenbacher, and G.Rigoll.Background segmentation with feedback:The pixel-based adaptive segmenter.In IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops,pages38–43, Providence, RI, United states, 2012)) outside, while and newest at present and two methods that performance is most outstanding into Row comparative experiments, respectively SuBSENSE (P.-L.St-Charles, G.-A.Bilodeau, and R.Bergevin.SuBSENSE:A universal change detection method with local adaptive Sensitivity.IEEE Transactions on Image Processing, 24 (1): 359-373,2015.) and IUTIS-5(S.Bianco,G.Ciocca,R.Schettini"How far can you get by combining change Detection algorithms? " .Submitted to IEEE Transactions on Image Processing, 2015.)。

The hardware environment of this method is PC machine, and processor is Intel Xeon E3-1230V2 3.3GHz, and memory is 24G, Software runtime environment is Matlab2016.Data set be famous ChangeDetection.net2012 (Goyette N, Jodoin P M,Porikli F,et al.Changedetection.net:A new change detection benchmark dataset[C]//2012IEEE Computer Society Conference on Computer Vision And Pattern Recognition Workshops.IEEE, 2012:1-8.), there are six scene types altogether, are respectively as follows: Baseline, dynamicBackground, cameraJitter, intermittentObjectMotion, shadow and Thermal contains 4 to 6 sub- scene videos in each scene type again.In qualitative analysis, Fig. 6 is shown respectively from this The representative sub-scene video of one extracted in six classifications, i.e. a of the first row arrange f column be respectively office, Boats, sidewalk, sofa, peopleInShade and library.Second row to the end a line be respectively KDE, GMM, PBAS, SuBSENSE, IUTIS-5, context of methods and GroundTruth.It can be seen from the figure that context of methods is surveyed in prospect spy Accuracy on it is well many than being equally based on statistical KDE, GMM and PBAS, while than newest method SuBSENSE and IUTIS-5 is also slightly better, and very close with GroundTruth.

In quantitative analysis, table 1 gives this method and 5 kinds of methods for comparing whole scene videos in data set Three kinds of indexs comparison, respectively Recall, Precision and F-Measure.As can be seen from the table, this method exists It is better than all methods in Recall index, is that second and third are good respectively in F-Measure and Precision index.

Comparison of the 1 six kinds of methods of table on whole scene videos about Recall, Precision and F-Measure index As a result；The runic of number indicates optimal result, the excellent result of the expression second with asterisk

Method	Recall	Precision	F-Measure
				KDE	0.7442	0.6843	0.6719
GMM	0.6964	0.7079	0.6596
				PBAS	0.7840	0.8160	0.7532
SuBSENSE	0.8281	0:8576*	0.8260
				IUTIS-5	0.8471*	0.8913	0.8542
Ours	0.8522	0.8421	0:8343*

A kind of new method detected for background difference and prospect proposed by the present invention can effectively improve the standard of model True property.

It should be understood that the part that this specification does not elaborate belongs to the prior art.

It should be understood that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this The limitation of invention patent protection range, those skilled in the art under the inspiration of the present invention, are not departing from power of the present invention Benefit requires to make replacement or deformation under protected ambit, fall within the scope of protection of the present invention, this hair It is bright range is claimed to be determined by the appended claims.

Claims (5)

1. a kind of method of background difference and prospect detection, it is characterised in that: building minot vector, for saving each pixel Reference data and statistical information；The dynamic of each position of background image is defined, representing dynamic level of pixel series is set, according to dynamic Grade classification rule, is attributed to a certain representing dynamic level of pixel for dynamic；Optimum bandwidth is obtained using dynamic KDE algorithm；To each Pixel designs an adaptive threshold value, and the value of threshold value and the dynamic of pixel are related；Utilize Tetris update scheme Elimination operation is carried out to the reference data bottom in minot vector, to eliminate noise jamming；

The building minot vector, specific implementation the following steps are included:

Step A1: the one-dimensional vector that building length is 263 is as minot vector；

Step A2: for each position of background image, minot vector is designed as storage ash from the unit of subscript 0 to 255 Angle value is corresponding lower target quantity；

Step A3: minot vector is believed from the statistics that the unit of subscript 256 to 262 is designed as the storage each position of background image Breath, wherein the sum of the storage gray value of Unit 256, Unit 257 store the variation amplitude of gray value, Unit 258 and 259 are deposited respectively Minimum and maximum gray value is stored up, the storage of Unit 260 gray scale is formed by the number of cluster, and the storage of Unit 261 gray scale is formed by The number of kurtosis, Unit 262 store the sum of bottom gray value.

2. the method for background difference according to claim 1 and prospect detection, which is characterized in that the definition background image Each position dynamic, be the variation range of the gray value of pixel in a period of time.

3. the method for background difference according to claim 1 and prospect detection, which is characterized in that described to obtain optimal band Width, specific implementation the following steps are included:

Step B1: it chooses a normal distribution and estimates as pilotAnd use observation data I=I₁,I₂,...,I_NCalculating side Poor σ², wherein N is the quantity for observing data；

Step B2: local bandwidth h is calculated using following formula_loc:

Wherein, iqr is quartile section；

Step B3: dynamic factor η is defined_i:

Wherein, geom isGeometric mean, α is sensitivity parameter, α ∈ [- 1,0]；

Step B4: optimum bandwidth is obtained are as follows:

H_opt=η_ih_loc(3)。

4. the method for background difference according to claim 1 and prospect detection, which is characterized in that described to give each pixel Design an adaptive threshold value, specific implementation the following steps are included:

Step C1: according to formulaEach observation data I is calculated at other Probability density in dataWherein, I value I₁,I₂,...,I_m, m is not repeat gray scale in observation data The number of value, K () indicate kernel function；

Step C2: rightIt carries out ascending sort and removes duplicate keys, i.e., Obtain the threshold value array Thresholds [m'] that length is m'；Wherein I=I₁,I₂,...,I_m；

Step C3: the dynamic rank dr obtained when being divided according to representing dynamic level of pixel sets its initial threshold as Thresholds [dr].

5. the method for background difference according to claim 1 and prospect detection, which is characterized in that the Tetris is more New departure, specific implementation the following steps are included:

Step D1: setting update condition；

Step D2: when meeting update condition, by minot vector, the value of non-zero unit subtracts 1 from subscript 0 to 255；

Step D3: statistical mark position information is updated.