CN106156776A - A kind of illumination recognition methods in traffic video monitoring - Google Patents

Abstract

The invention discloses the illumination recognition methods in a kind of traffic video monitoring, to the traffic video monitoring image obtained in real time, retain filtering by edge and reduce noise jamming, and the edge of objects in images can be retained, do not lose the detailed information in original image, then the image obtained after utilizing original image and filtering calculates the Y-PSNR (PSNR) foundation as illumination identification, by contrast PSNR value with classification thresholds so that it is determined that illumination is dark or bright.Additionally, the present invention is by statistical history PSNR value, the most adaptive adjustment classification thresholds.This method have the advantage that accuracy of identification is high, higher degree of accuracy is ensured in the case of large area shade, scene complexity, noise are many, owing to have employed dynamic threshold, it is possible to prevent the wrong identification in the case of change in weather, seasonal variations, for Changes in weather, there is the strongest stability.

Description

A kind of illumination recognition methods in traffic video monitoring

Technical field

The present invention relates to Computer Vision Recognition technology, particularly relate to the illumination recognition methods in traffic video monitoring.

Background technology

Along with the fast development of Urbanization in China, urban population and vehicle synchronous increase severely, traffic problems have become city The huge difficult problem that development faces.For cracking traffic quagmire, in all parts of the country except extensive development track traffic, all by sight Invest traffic video monitoring.Traffic video monitoring technology is based on image procossing, the novel video monitoring skill of pattern recognition Art.It is used for finding the object of motion in image, and it is tracked, analyzes, find "abnormal" behavior in time, touch Transmit messages alert and take other measures to intervene.

In current traffic video monitoring system, in the case of various complex illumination cannot being met due to single analysis system Scene demand, because all taking two or more system to carry out integrated treatment in the industry.Divide according to different light conditions Do not use different analysis systems can greatly strengthen the effect of video monitoring.Thus, the judgement to light conditions just becomes The core of system call studies a question.

At present, the technology of weather identification mainly has two classes: quiet hour threshold method and statistics of histogram method.

Wherein, quiet hour threshold method considers that the light conditions on daytime is substantially better than the light conditions at night, uses pre- If time threshold, use different monitoring systems in different time period.The major defect of this method has two: the One, adaptive ability is poor, and threshold value is completely dependent on manually adjusting；Second, it is impossible to effectively detect what vile weather caused Visibility reduces situation.

The statistics of histogram method pixel value by Statistical monitor image, utilizes statistical nature to distinguish all kinds of light conditions. The method can detect vile weather to a certain extent, but algorithm stability is the strongest.For complex scene, in scene Have that noise in large area shade, scene is many, in scene in the case of the various complexity of object, it is impossible to enough ensure good knowledge Other effect.And the method needs to gather image in advance during realizing and carries out early stage training to determine threshold value, it is achieved side Formula is more complicated.

Therefore, it is necessary to a kind of self adaptation of research is good, the illumination recognition methods that stability is strong, this is conducive to improving video Monitoring effect.Thus reduce traffic department and process complicated, the workload of mistake transport information, use manpower and material resources sparingly.Simultaneously Significant for the vehicle accident detection in video monitoring and vehicle peccancy detection etc..

Summary of the invention

The technical problem to be solved is to provide a kind of illumination identification in traffic intersection video monitoring system Method, its accuracy of identification is high, stability is strong, and self adaptation can identify accuracy with guarantee under complicated environmental condition.

The technical solution adopted in the present invention is: a kind of illumination recognition methods, comprises the following steps:

Obtain n-th frame crossing image I currently pending in the monitor video of crossing the most in real time_N, wherein, the initial value of n is 0, 0≤n≤N-1, N represent the totalframes of the crossing picture comprised in the crossing monitor video of acquisition in real time；

If 2. n ≠ 0, then perform step 4.；If n=0, initialize the statistics of PSNR (Y-PSNR) value Amount S (PSNR value acquisition mode is in step 6. middle explanation)；Statistic S=(s₀, s₁..., s_M-1) it is a M dimension group, s₀, s₁..., s_M-1Representing M integer variable, M is the maximum of PSNR value；Initialization mode is: by s₀, s₁..., s_M-1 All it is entered as 0；Then proceed to perform step 3.；

3. classification thresholds T is entered as empirical value, then proceedes to perform step 5.；

4., as n%100=0, update classification thresholds T, do not update；Then step is performed 5.；

5. to current n-th frame image I_NIt is filtered processing, obtains filtered image O_N；

6. according to the image I not having filtering_NOutput image O with filtering_NCalculating PSNR value, calculation isWherein MAX represents maximum gradation value；| ω | represents the number of all pixels in image Amount；I_iAnd O_iThe original image I represented respectively_NWith filtering output image O_NPixel value at ith pixel, i represents pixel Sequence number, 0≤i≤| ω |-1；

7. by step 6. in calculated PSNR value add statistic S, method particularly includes: if PSNR≤1, then Assignment s₀=s₀+1；If 1 ＜ PSNR≤2, then assignment s₁=s₁+1；If ... M-1 ＜ PSNR, Then assignment s_M-1=s_M-1+1；

8. comparison step 6. in calculated PSNR value and present threshold value T；If PSNR >=T, it is identified result For " illumination becomes clear "；If PSNR is ＜ T, it is identified result for " illumination is dark ".

The detailed process of described step 4. middle renewal classification thresholds T is:

4. the discrete probability distribution P=(p of PSNR value-1, is calculated₀, p₁..., p_M-1), calculation is Wherein s_iIt is statistic S=(s₀, s₁..., s_M-1Integer value in), 0≤i≤M-1；

4.-2, Split Index Y=(y is calculated₀, y₁..., y_M-1), calculation is Wherein p_kIt is discrete probability distribution P=(p₀, p₁..., p_M-1Probit in), 0≤i≤M-1；

4.-3, Split Index Y=(y is selected₀, y₁..., y_M-1The element y that in), numerical value is maximum_i, integer value i of its correspondence is just It it is new classification thresholds.

The detailed process of described step 5. middle Filtering Processing is:

5.-1, image I to be filtered is represented respectively with w and h_NWidth and height；Firstly, for pixel each in image Neighborhood calculates corresponding average and the difference of two squares；For ith pixel, its neighborhood territory pixel integrates as ω_i, ω_iRepresent institute in image Having the collection of pixels less than r of the distance with ith pixel, r is preset parameter；ω_iCorresponding average is Wherein i, j represent the sequence number of pixel, dist (i, j) represent ith pixel and jth pixel away from From, I_jRepresent the pixel value of jth pixel, | ω_i| represent set ω_iIn the quantity of all pixels；ω_iThe corresponding difference of two squares is ${\mathrm{\&sigma;}}_i^2=\frac{1}{\left|{\mathrm{\&omega;}}_i\right|}{\mathrm{\&Sigma;}}_{\mathrm{dist}(i,j)<r}{(I_j-{\mathrm{\&mu;}}_i)}^2;$

5.-2, filtering parameter is calculated, for ith pixel correspondence parameter alpha_i, computational methods are WhereinBeing the 5. calculated difference of two squares in-1, ε is preset parameter；

5.-3, filtering parameter is calculated, for ith pixel correspondence parameter beta_i, computational methods are Wherein μ_jIt is 5. calculated average in-1,Being the 5. calculated difference of two squares in-1, ε is fixing ginseng Number；

-4 5., utilize 5. calculated filtering parameter in-2 and 5.-3, calculate filter result；For ith pixel, Its corresponding filtering is output as O_i=α_iI_i+β_i；Wherein I_iRepresent the pixel value of jth pixel.

Compared with prior art, it is an advantage of the current invention that:

1) the inventive method combines the judgement of PSNR value and the dynamic threshold of real time modifying, it is possible to adapt to various complex environment. Comparing and judge according to time threshold, this method can adapt to the change of day length automatically.Compare statistics of histogram side Method, this method can, noise complicated in large area shade, scene many in the case of ensure higher degree of accuracy.

2) in the weather decision process of the inventive method, 5. step uses filtering, can retain figure while reducing noise The marginal information of object in Xiang, does not lose image information.When acting on weather identification, it is possible to be effectively kept fine day object Sharp-edged feature, and reduce the interference that PSNR value is calculated by noise to greatest extent.

3) the inventive method uses dynamic threshold, the most just according to real-time statistical data amendment present threshold value. Dynamic threshold is prevented from the wrong identification in the case of change in weather, seasonal variations.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of the inventive method；

Detailed description of the invention

Below in conjunction with accompanying drawing embodiment, the present invention is described in further detail.

The present invention proposes the illumination recognition methods in a kind of traffic video monitoring, and overall procedure block diagram is as it is shown in figure 1, concrete Comprise the following steps:

The present invention solves the technical scheme that above-mentioned technical problem used: a kind of illumination recognition methods, comprises the following steps:

Obtain n-th frame crossing image I currently pending in the monitor video of crossing the most in real time_N, wherein, the initial value of n is 0, 0≤n≤N-1, N represent the totalframes of the crossing picture comprised in the crossing monitor video of acquisition in real time；In this enforcement In example, N determines according to actual crossing monitor video, generally just can meet the accuracy of recognition result more than 100.

If 2. n ≠ 0, then perform step 4.；If n=0, initialize the statistics of PSNR (Y-PSNR) value Amount S (PSNR value acquisition mode is in step 6. middle explanation)；Statistic S=(s₀, s₁..., s_M-1) it is a M dimension group, s₀, s₁..., s_M-1Representing M integer variable, M is the maximum of PSNR value, and M value is 100 in instances.Initially Change mode is: by s₀, s₁..., s_M-1All it is entered as 0；Then proceed to perform step 3..

3. classification thresholds T being entered as empirical value, in this example, T is entered as 50.Then proceed to perform step 5..

4., as n%100=0, it is meant that n is the integral multiple of 100, update classification thresholds T, do not update；Then Perform step 5..

In this embodiment, step detailed process 4. is as follows:

4. the discrete probability distribution P=(p of PSNR value-1, is calculated₀, p₁..., p_M-1), calculation is Wherein s_iIt is statistic S=(s₀, s₁..., s_M-1Integer value in), 0≤i≤M-1；

4.-2, Split Index Y=(y is calculated₀, y₁..., y_M-1), calculation is Wherein p_kIt is discrete probability distribution P=(p₀, p₁..., p_M-1Probit in), 0≤i≤M-1；

4.-3, Split Index Y=(y is selected₀, y₁..., y_M-1The element y that in), numerical value is maximum_i, integer value i of its correspondence is just It it is new classification thresholds.

5. to current n-th frame image I_NIt is filtered processing, obtains filtered image O_N；

In this embodiment, step detailed process 5. is as follows:

5.-1, image I to be filtered is represented respectively with w and h_NWidth and height, in this example, w=600, h=500； Neighborhood firstly, for pixel each in image calculates corresponding average and the difference of two squares；For ith pixel, its neighborhood Set of pixels is ω_i, ω_iRepresenting that in image, all distances with ith pixel are less than the collection of pixels of r, r is preset parameter, In this example, r=20.0；ω_iCorresponding average isWherein i, j represent the sequence number of pixel, (i j) represents ith pixel and the distance of jth pixel, I to dist_jRepresent the pixel value of jth pixel, | ω_i| represent set ω_i In the quantity of all pixels, in this example | ω_i| span is within [400,1600]；ω_iThe corresponding difference of two squares is ${\mathrm{\&sigma;}}_i^2=\frac{1}{\left|{\mathrm{\&omega;}}_i\right|}{\mathrm{\&Sigma;}}_{\mathrm{dist}(i,j)<r}{(I_j-{\mathrm{\&mu;}}_i)}^2;$

5.-2, filtering parameter is calculated, for ith pixel correspondence parameter alpha_i, computational methods are WhereinBeing the 5. calculated difference of two squares in-1, ε is preset parameter, ε=0.01 in this example；

5.-3, filtering parameter is calculated, for ith pixel correspondence parameter beta_i, computational methods are Wherein μ_jIt is 5. calculated average in-1,Being the 5. calculated difference of two squares in-1, ε is fixing ginseng Number, ε=0.01 in this example；

-4 5., utilize 5. calculated filtering parameter in-2 and 5.-3, calculate filter result；For ith pixel, Its corresponding filtering is output as O_i=α_iI_i+β_i；Wherein I_iRepresent the pixel value of jth pixel.

6. according to the image I not having filtering_NOutput image O with filtering_NCalculating PSNR value, calculation isWherein MAX represents maximum gradation value；| ω | represents the number of all pixels in image Amount, | ω |=600 × 500 in this example；I_iAnd O_iThe original image I represented respectively_NWith filtering output image O_NAt i-th picture Pixel value at Su, i represents the sequence number of pixel, 0≤i≤| ω |-1；

7. by step 6. in calculated PSNR value add statistic S, method particularly includes: if PSNR≤1, then Assignment s₀=s₀+1；If 1 ＜ PSNR≤2, then assignment s₁=s₁+1；If ... M-1 ＜ PSNR, Then assignment s_M-1=s_M-1+1；

8. comparison step 6. in calculated PSNR value and present threshold value T；If PSNR >=T, it is identified result For " illumination becomes clear "；If PSNR is ＜ T, it is identified result for " illumination is dark ".

For more strongly suggesting feasibility and the effectiveness of the inventive method, the accuracy performance of the inventive method is entered by we Row test.

We Various Seasonal, different crossing monitor video in intercepted 10000 two field pictures at random, and to each frame figure The state of weather of picture is manually demarcated.And test quiet hour threshold value, statistics of histogram and the inventive method.Its In, daytime fine day, the greasy weather on daytime belong to the situation that light conditions is good, be classified as the sunny classification of the colour of sky, remaining situation all belongs to In colour of sky darkness classification.

Statistical result is as listed in table 1.

Table 1 weather judges accuracy statistical result

	Quiet hour threshold value	Statistics of histogram	The inventive method
				Daytime fine day	86.4%	91.1%	99.7%
Greasy weather on daytime	95.0%	80.3%	98.1%
				Cloudy day on daytime	13.4%	74.2%	89.4%
Rainy day on daytime	27.1%	55.2%	95.3%

Night fine day	75.8%	93.6%	99.1%
				Rainy day at night	77.7%	79.3%	99.7%

From table 1 it follows that for various situations, the inventive method can reach the highest accuracy.

As shown in table 1, quiet hour threshold method is the lowest for the differentiation accuracy at cloudy day on daytime and rainy day on daytime, and this Class erroneous judgement can affect follow-up vehicle identification tracking effect.In the inventive method, employ dynamic threshold calculation, right Overcast and rainy can adjust threshold value timely, it is ensured that judge accuracy.Additionally, dynamic threshold also is able to adapt in Various Seasonal Daytime the different situation of duration, the most by day fine day, greasy weather and fine day at night, in the case of the rainy day, this method is just Really rate is also apparently higher than quiet hour threshold method.

Contrast statistics of histogram method as a result, it is possible to find, the method greasy weather by day, rainy day on daytime and night In the case of rainy day, it determines accuracy ratio is relatively low.Because noise in image is many in the case of these are several, can be to directly system Meter produces the biggest interference.And 1. this method have employed guiding filtering due to the step of weather decision process, significantly drop Low signal-to-noise ratio so that judge that in the case of these are several accuracy is higher.

Various situations in contrast table 1, it can be seen that the algorithm stability of the inventive method is substantially better than other two method.

Claims (3)

1. a traffic lights recognition methods, it is characterised in that comprise the following steps:

Obtain n-th frame crossing image I currently pending in the monitor video of crossing the most in real time_N, wherein, the initial value of n is 0, 0≤n≤N-1, N represent the totalframes of the crossing picture comprised in the crossing monitor video of acquisition in real time；

If 2. n ≠ 0, then perform step 4.；If n=0, initialize the statistics of PSNR (Y-PSNR) value Amount S (PSNR value acquisition mode is in step 6. middle explanation)；Statistic S=(s₀, s₁..., s_M-1) it is a M dimension group, s₀, s₁..., s_M-1Representing M integer variable, M represents PSNR value maximum；Initialization mode is: by s₀, s₁..., s_M-1 All it is entered as 0；Then proceed to perform step 3.；

3. classification thresholds T is entered as empirical value, then proceedes to perform step 5.；

4., as n%100=0, update classification thresholds T, do not update；Then step is performed 5.；

5. to current n-th frame image I_NIt is filtered processing, obtains filtered image O_N；

6. according to the image I not having filtering_NOutput image O with filtering_NCalculating PSNR value, calculation isWherein MAX represents maximum gradation value；| ω | represents the number of all pixels in image Amount；I_iAnd O_iThe original image I represented respectively_NWith filtering output image O_NPixel value at ith pixel, i represents pixel Sequence number, 0≤i≤| ω |-1；

7. by step 6. in calculated PSNR value add statistic S, method particularly includes: if PSNR≤1, then Assignment s₀=s₀+1；If 1 ＜ PSNR≤2, then assignment s₁=s₁+1；If ... M-1 ＜ PSNR, Then military value s_M-1=s_M-1+1；

8. comparison step 6. in calculated PSNR value and present threshold value T；If PSNR >=T, it is identified result For " illumination becomes clear "；If PSNR is ＜ T, it is identified result for " illumination is dark ".

A kind of illumination recognition methods the most according to claim 1, it is characterised in that described step 4. middle self adaptation Ground determines that the detailed process of current light classification thresholds is:

4. the discrete probability distribution P=(p of PSNR value-1, is calculated₀, p₁..., p_M-1), calculation is Wherein s_iIt is to add up S=(s₀, s₁..., s_M-1Integer value in), 0≤i≤M-1；

4.-2, Split Index Y=(y is calculated₀, y₁..., y_M-1), calculation is Wherein p_kIt is discrete probability distribution P=(p₀, p₁..., p_M-1Probit in), 0≤i≤M-1；

4.-3, Split Index Y=(y is selected₀, y₁..., y_M-1The element y that in), numerical value is maximum_i, integer value i of its correspondence is just It it is new classification thresholds.

A kind of illumination recognition methods the most according to claim 1, it is characterised in that described step concrete mistake 5. Cheng Wei:

5.-1, image I to be filtered is represented respectively with w and h_NWidth and height；Firstly, for pixel each in image Neighborhood calculates corresponding average and the difference of two squares；For ith pixel, its neighborhood territory pixel integrates as ω_i, ω_iRepresent institute in image Having the collection of pixels less than r of the distance with ith pixel, r is preset parameter；ω_iCorresponding average is Wherein i, j represent the sequence number of pixel, dist (i, j) represents ith pixel and the distance of jth pixel, I_jRepresent the pixel value of jth pixel, | ω_i| represent set ω_iIn the quantity of all pixels；ω_iThe corresponding difference of two squares is ${\mathrm{\&sigma;}}_i^2=\frac{1}{\left|{\mathrm{\&omega;}}_i\right|}{\mathrm{\&Sigma;}}_{\mathrm{dist}(i,j)<r}{(I_j-{\mathrm{\&mu;}}_i)}^2;$

5.-2, filtering parameter is calculated, for ith pixel correspondence parameter alpha_i, computational methods are WhereinBeing the 5. calculated difference of two squares in-1, ε is preset parameter；

5.-3, filtering parameter is calculated, for ith pixel correspondence parameter beta_i, computational methods are Wherein μ_jIt is 5. calculated average in-1,Being the 5. calculated difference of two squares in-1, ε is fixing ginseng Number；

-4 5., utilize 5. calculated filtering parameter in-2 and 5.-3, calculate filter result；For ith pixel, Its corresponding filtering is output as O_i=α_iI_i+β_i；Wherein I_iRepresent the pixel value of jth pixel.