CN103927875B - Based on the traffic overflow state identification method of video - Google Patents

Abstract

The invention discloses a kind of traffic overflow state identification method based on video, step is as follows: set up two video cameras at target cross junction, First video camera is monitored the crossing interior vehicle queuing behavior that direction occurs in overflow, and second video camera is monitored signal lamp crisscross with it; The video image camera supervised for First carries out pre-service, extracts video features: vehicle exists ratio and motion pixel ratio; Then the fuzzy recognizer of band feedback is utilized to carry out image recognition; Process for second camera supervised image, realize location and the identification of signal lamp; Output signal light recognition result; According to true and false overflow decision rule, draw traffic overflow result; Carry out the display of traffic overflow state and report to the police.The present invention is monitored crossing internal transportation information in real time by video camera, adopts the fuzzy recognition technology of band feedback to realize the automatic identification of traffic overflow state, for traffic congestion prevention, alleviate and provide support.

Description

Based on the traffic overflow state identification method of video

Technical field

The present invention relates to a kind of traffic overflow state identification method based on video.

Background technology

Traffic overflow is the bottleneck effect due to downstream intersection, makes section vehicle queue exceed road section length, thus causes Some vehicles to occupy the traffic behavior of crossing, upstream.The direct result that it brings causes traffic congestion, if do not controlled after occurring, jam will spread from single crossing to periphery, causes chain blocking up, and even produces lattice lock, cause the traffic paralysis of large area road network.

At present, the domestic and international theoretical research about traffic overflow and technology are applied and are mostly concentrated on its genesis mechanism and control method field, detection for traffic overflow mainly utilizes conventional coil detecting device to realize, document [GeroliminisN, SkabardonisA.Queuespilloversincitystreetnetworkswithsign al-controlledIntersections [C] .TransportationResearchBoard (TRB) 89thAnnualMeeting, Washington, D.C.2010:10-3498.] the overflow recognition methods based on coil checker occupation rate data is proposed, and give respective threshold, but coil checker has installation and maintenance complex procedures, high in cost of production shortcoming, and the laying distance of detecting device is depended on to a great extent by the traffic overflow testing result accuracy rate that coil checker carries out, therefore the method has very large limitation in actual applications.Document [Zhang Lidong, Jia Lei, red legend is emerging. based on the traffic overflow recognizer [J] of fuzzy theory. and computer utility, 2012,32 (8): 2378-2380.] propose the overflow recognition methods based on fuzzy theory, obtain the index such as vehicle queue's ratio and average speed and carry out fuzzy reasoning, thus the overflow situation identification under achieving simulated environment, but These parameters is difficult to obtain in actual applications, and therefore the method has limitation in actual applications equally.The core link of traffic overflow state recognition is the detection to crossing interior vehicle queueing condition, document [He Xiaofeng, Yang Yuzhen, Chen Yangzhou. the vehicle queue length based on Computer Vision detects [J]. traffic and computing machine, 2006, the method detecting and carry out there is detection again of first carrying out moving based on video 24 (5): 43-46.] is proposed, can Quick Test Vehicle vehicle queue length, but the dynamic formation process of vehicle queue cannot be provided, necessary condition can not be provided for overflow situation identification.The current domestic manual observation mode that adopts identifies traffic overflow more, and efficiency is lower.Therefore, the brand-new overflow detection technique that a kind of detection speed of research and probe is fast, accuracy rate is high, be convenient to safeguard has significant realistic meaning.

Find by carrying out retrieval to existing patent of invention and technology, being reported in of relevant " intersection traffic overflow situation video identification " is also a blank both at home and abroad, the present invention obtains traffic image by video capture, utilize the fuzzy recognition technology of band feedback, and combining image treatment technology, realize the identification to traffic overflow state; And by recognition result with pictograph in conjunction with type of alarm Real-time Feedback to traffic administration person.

Summary of the invention

Object of the present invention is exactly to solve the problem, a kind of traffic overflow state identification method based on video is provided, it has for the frequent traffic overflow phenomenon in urban transportation peak period, by high-definition camera, crossing internal transportation information is monitored in real time, realize to stationary vehicle queue up and moving state identification basis on, adopt the fuzzy recognition technology of band feedback, realize the judgement to vehicle queue's formation and dissipation state, finally realize the automatic identification of traffic overflow state, and meet real-time, the requirement such as by force of stability and adaptivity, for the prevention of urban traffic blocking, alleviate the advantage provided support.

To achieve these goals, the present invention adopts following technical scheme:

Based on a traffic overflow state identification method for video, comprise the steps:

Step (1): set up two high-definition cameras at target cross junction, wherein First high-definition camera is monitored the crossing interior vehicle queuing behavior that direction occurs in overflow, and particular location is the lateral mid-point near section end; The crisscross signal lamp of second high-definition camera pair and overflow direction is monitored, and particular location is the section end lateral mid-point with overflow direction vertical direction;

Step (2): comprise two steps arranged side by side: step (2-1) and step (2-2);

Step (2-1): the video image for First high-definition camera machine monitoring carries out pre-service, extracts video features: vehicle exists ratio and motion pixel ratio; Then the fuzzy recognizer of band feedback is utilized to carry out image recognition;

Step (2-2): the image for second high-definition camera machine monitoring processes, realizes location and the identification of signal lamp; Then output signal light recognition result;

Step (3): the result of combining step (2-1) and step (2-2), according to true and false overflow decision rule, draws traffic overflow result;

Step (4): carry out the display of traffic overflow state and report to the police.

The pretreated step of described step (2-1) is:

The acquisition of background updating and renewal, moving object detection, extract vehicle target in crossing;

The acquisition of described background updating adopts statistic histogram method, shown in formula specific as follows:

$P(x,y,k)=\left\{\begin{array}{ccc}P(x,y,k)+1& \mathrm{if}& I_i(x,y)=k\\ P(x,y,k)& \mathrm{if}& I_i(x,y)\≠k\end{array}\right.i=\mathrm{1,2,2},...,N---\left(1\right)$

B(x,y)＝k(2)

ifP(x,y,k)＝max(P)k＝0,1,2,…,255(3)

B (x, y) is the gray-scale value that the background image of foundation is corresponding at pixel (x, y) place, I _i(x, y) represents the gray-scale value that the i-th frame original image is corresponding at pixel (x, y) place, and N represents statistics totalframes.P (x, y, k) represents the number of times that pixel (x, y) place gray-scale value k occurs; Formula (3) represents at point (x, y) place, gets the gray-scale value of the maximum gray-scale value of occurrence number image as a setting.

The renewal of described background updating adopts selectivity background update method, and available formula (4) represents:

$B_i(x,y)=\left\{\begin{array}{cc}\mathrm{\α}{I}_i(x,y)+(1-\mathrm{\α}){\mathrm{\β}}_{i-1}(x,y)& G_i(x,y)=0\\ B_{i-1}(x,y)& G_i(x,y)\≠0\end{array}\right.---\left(4\right)$

B _i(x, y) and I _i(x, y) is respectively the i-th frame background image and the gray-scale value of original image at pixel (x, y) place, G _i(x, y) is that the i-th frame original image carries out the value of the differentiated binary image of background at pixel (x, y) place, and α is context update weight, 0≤α≤1.

The step that the vehicle of the extraction video features of described step (2-1) exists ratio is:

It is vehicle area coverage and the ratio of the surveyed area total area in crossing that vehicle exists than Q, adopts background difference and Moving Window to exist vehicle and extracts than Q characteristic quantity:

Background differential representation is as follows:

Make I _kfor kth frame gray level image, B _kfor background gray level image, by I _kwith B _kdiffer from, obtain background subtraction image M _k, to M _kcarry out medium filtering, and utilize Otsu algorithm to provide threshold value T ₁to M _kcarry out Iamge Segmentation, obtain binary image S _k; Calculating formula is as follows:

$S_k(x,y)=\left\{\begin{array}{cc}1& |I_k(x,y)-B_k(x,y)|>T_1\\ 0& \mathrm{otherwise}\end{array}\right.---\left(5\right)$

Moving Window is expressed as follows:

The horizontal detection window utilizing a height to be surveyed area vertical height 1/10th, rearwardly moves from queue head.Detection window is scanned from team's head to tail of the queue, if the horizontal zone meeting existence condition is maximum be numbered i, then L _k=i (0≤i≤10).With L _kthere is ratio in the vehicle characterizing the i-th two field picture.Median point due to data sequence has stronger robustness, and this algorithm adopts the time span of N continuous two field picture as the least unit time, and the intermediate value of N continuous frame testing result sequence compares result as the vehicle existence of unit interval.

Therefore, vehicle existence than Q characteristic quantity is:

Q＝med(L _k),(1≤k≤N)(6)

L _k＝max(I)(7)

$I=\left\{\begin{array}{cc}i|P_E\left(i\right)=\frac{E_i}{N_i}>T_E,& 0\≤i\≤10\end{array}\right\}---\left(8\right)$

Formula (8) represents that I meets the set that there are all zone numbers of detection threshold in kth two field picture; Formula (7) represents that the maximal value of getting in set I exists ratio to the vehicle characterizing kth two field picture, uses L _krepresent; Formula (6) represents in the video sequence of a N continuous frame, gets all L _kintermediate value Q characterize current N two field picture vehicle exist compare result.

The step of the motion pixel ratio of the extraction video features of described step (2-1) is:

Motion pixel is the ratio of motor image vegetarian refreshments and whole pixel in surveyed area than S, and definition is

S＝med(P _M(k)),(1≤k≤N)(9)

$P_M\left(k\right)=\frac{M_k}{N_k}---\left(10\right)$

Wherein, M _kfor motion number of pixels in kth two field picture surveyed area, N _kfor all number of pixels in kth two field picture surveyed area, formula (9) represents in the video sequence of a N continuous frame, gets all P _mk vehicle that the intermediate value S of () characterizes current N two field picture exists and compares result.

The fuzzy recognizer of band feedback is utilized to carry out image recognition in described step (2-1):

Exist than Q and motion pixel than the input of S as fuzzy recognizer using the vehicle gathered, after the obfuscation of fuzzy recognizer, fuzzy reasoning and ambiguity solution, fuzzy recognizer exports current overflow testing result; Wherein, fuzzy recognizer is logically made up of four parts, i.e. fuzzy diagnosis rule, fuzzification process, fuzzy reasoning process and deblurring process, through above four step process, this fuzzy recognizer can export fuzzy diagnosis result, the degree of overflow result namely in the present invention.

When vehicle is in queuing formation, dissipation transition state, " the current overflow testing result " that exported by fuzzy recognizer is as the feed back input amount of fuzzy recognizer, in the practical work process of fuzzy recognizer, feed back input amount is " last overflow testing result ", and this initialization of variable value is " without overflow ".

It is that { without N, low L, middle M, high H}, represent Current vehicle queuing ratio " without queuing up ", " low ", " medium " and " height " respectively that described vehicle exists than the fuzzy subset of Q;

The fuzzy subset of described motion pixel ratio S is that { blocking C, unobstructed F, transition T}, represent that the inner wagon flow traffic status in current crossing is " blocking ", " unobstructed " and " transition " respectively;

The fuzzy subset of described current overflow testing result O is that { without N, slight L, moderate M, severe H}, represent crossing flooded conditions "None", " slightly ", " moderate ", " severe " respectively.

Vehicle existence is respectively trapezoidal, trapezoidal and triangular function than the membership function of Q, motion pixel ratio S and current overflow testing result O ternary.Fuzzy inference rule: than Q be if Current vehicle exists motion pixel than S is last overflow testing result is then current overflow testing result is fuzzy implication relation adopts Mamdani " maximum-minimum " rationalistic method, obtains fuzzy control search table by gravity model appoach ambiguity solution.Described fuzzy control search table in the real-time testing process of traffic overflow for inquiry.This fuzzy recognizer introduces last overflow testing result as feedback quantity, thus realizes judging the flooded conditions of the wagon flow being in transient motion state.

The location of described step (2-2) signal lamp with identify it is adopt the signal lamp automatic identifying method under a kind of complex background, comprise Intensity segmentation, K mean cluster and prospect histogram analysis means, the color of signal lamp, directional information identified.The method concrete steps are as follows:

Image is converted into HSV space by A, and adopt Da-Jin algorithm to provide Intensity segmentation threshold value, its basic thought finds out a segmentation threshold, makes to be had maximum between-cluster variance or infima species internal variance by its pixel gray-scale value being divided into two classes.Iamge Segmentation based on brightness is regular such as formula shown in (11):

$f_3(x,y)=\left\{\begin{array}{cc}{f}_2(x,y)& V(x,y)\&GreaterEqual;T^{*}\\ 0& V(x,y)<T^{*}\end{array}\right.---\left(11\right)$

B extracts based on the signal lamp position of geometric properties: on the basis of Intensity segmentation, carry out gray processing process to image, then carry out contour detecting, and do boundary rectangle to the profile detected.Remember that single rectangular area is R, according to shape and the positional information of R, introduce new regulation and filtering is carried out to noise.New regulation comprises area information, positional information and aspect ratio information etc. as constraint condition.

Red, the green state of C definition signal lamp is signal lamp key message, and yellow and digital countdown state is non-critical information.The key message of signal lamp can be described by position coordinates, color, shape and direction.First adopt statistic of classification to obtain object detection area, rule is:

If any two candidate region R _ijcentral point L _ij(x _ij, y _ij) between Euclidean distance d meet

$d=\sqrt{{(x_{\mathrm{ij}1}-x_{\mathrm{ij}2})}^2+{(y_{\mathrm{ij}1}-y_{\mathrm{ij}2})}^2}<T6---\left(12\right)$

Then think 2 coincidences, otherwise think that do not overlap at 2.Wherein, R _ijbe the jth candidate region in the i-th two field picture, T6 is distance threshold.In target vides sequence, total N two field picture, comprises the signal period that several are complete; Wherein, in the i-th two field picture, the total number in candidate region is M _i, to L _ij(i=(1,2 ..., N), j=(1,2 ..., M _i)) classify according to above-mentioned rule, obtain not coincident central point L _k, k=(1,2 ..., S), proportion is respectively C in the video sequence _k, k=(1,2 ..., S), S is the number of not coincident central point.Because the appearance in the video sequence of signal lamp non-critical information and noise has the features such as randomness is strong, the duration is short, therefore basis

C _k＞T7(13)

All central points are screened, key signal lamp location sets D={R can be obtained _ij| i=(1,2 ..., N), j=(1,2 ..., K _i), wherein, T7 is proportion threshold value, K _iit is key signal lamp areal in the i-th two field picture.

D adopts K means clustering algorithm to obtain red, green color cluster center (r _c, g _c), then region R to be measured _ij(i=(1,2 ..., N2), j=(1,2 ..., K _i)) color decision rule be

$\mathrm{color}=\left\{\begin{array}{cc}R& |h_{\mathrm{ij}}-r_c|=\min \left\{\right|h_{\mathrm{ij}}-r_c|,|h_{\mathrm{ij}}-g_c\left|\right\}\\ G& |h_{\mathrm{ij}}-r_c|=\min \left\{\right|h_{\mathrm{ij}}-r_c|,|h_{\mathrm{ij}}-g_c\left|\right\}\end{array}\right.---\left(14\right)$

E is the flexible identification realized circular and arrow-shaped signal lamp, employing prospect histogram analysis method: summarize the regularity of distribution of signal with different type lamp prospect histogram under different scanning mode, and mated with generalise results by the histogram in region to be measured, and then draw signal lamp shape and directional information.

The step of described step (3) is: crossing interior vehicle queuing situation and signal lamp recognition result are comprehensively analyzed, and according to the true and false overflow decision rule in crossing, provides overflow recognition result; The frequency that described overflow recognition result provides sets according to user's actual needs;

Divide according to overflow character, overflow result comprises true overflow and false overflow two kinds of situations;

Divide according to degree of overflow, overflow result comprise without overflow, slightly, moderate and severe four kinds of situations.

The invention has the beneficial effects as follows:

1, the promptly and accurately identification to crossing overflow phenomena can be realized;

2, the shortcomings such as artificial cognition efficiency is low, cost is high, information transmission is slow are overcome;

3, high-definition camera has fast, accurately, is easy to the advantages such as maintenance, makes this invention have great practical value.

Accompanying drawing explanation

Fig. 1 is overall flow figure of the present invention;

Fig. 2 is video camera mounting means and system framework schematic diagram;

Fig. 3 is traffic overflow fuzzy recognizer;

Fig. 4 is overflow recognition result display interface;

Fig. 5 is that Moving Window detection vehicle exists schematic diagram.

Embodiment

Below in conjunction with accompanying drawing and embodiment, the invention will be further described.

As shown in Figure 1, a kind of traffic overflow state identification method based on video, comprises the steps:

Step (1): as shown in Figure 2, two high-definition cameras are set up at target cross junction, wherein First high-definition camera is monitored the crossing interior vehicle queuing behavior that direction occurs in overflow, and particular location is the lateral mid-point near section end; The crisscross signal lamp of second high-definition camera pair and overflow direction is monitored, and particular location is the section end lateral mid-point with overflow direction vertical direction;

Step (2): comprise two steps arranged side by side: step (2-1) and step (2-2);

Step (2-1): the video image for First high-definition camera machine monitoring carries out pre-service, extracts video features: vehicle exists ratio and motion pixel ratio; Then the fuzzy recognizer of band feedback is utilized to carry out image recognition;

Step (2-2): the image for second high-definition camera machine monitoring processes, realizes location and the identification of signal lamp; Then output signal light recognition result;

Step (3): the result of combining step (2-1) and step (2-2), according to true and false overflow decision rule, draws traffic overflow result;

Step (4): carry out the display of traffic overflow state and report to the police.

The pretreated step of described step (2-1) is:

The acquisition of background updating and renewal, moving object detection, extract vehicle target in crossing;

The acquisition of described background updating adopts statistic histogram method, shown in formula specific as follows:

$P(x,y,k)=\left\{\begin{array}{ccc}P(x,y,k)+1& \mathrm{if}& I_i(x,y)=k\\ P(x,y,k)& \mathrm{if}& I_i(x,y)\&NotEqual;k\end{array}\right.i=\mathrm{1,2,2},...,N---\left(1\right)$

B(x,y)＝k(2)

ifP(x,y,k)＝max(P)k＝0,1,2,…,255(3)

B (x, y) is the gray-scale value that the background image of foundation is corresponding at pixel (x, y) place, I _i(x, y) represents the gray-scale value that the i-th frame original image is corresponding at pixel (x, y) place, and N represents statistics totalframes.P (x, y, k) represents the number of times that pixel (x, y) place gray-scale value k occurs; Formula (3) represents at point (x, y) place, gets the gray-scale value of the maximum gray-scale value of occurrence number image as a setting.

The renewal of described background updating adopts selectivity background update method, and available formula (4) represents:

$B_i(x,y)=\left\{\begin{array}{cc}\mathrm{\&alpha;}{I}_i(x,y)+(1-\mathrm{\&alpha;}){\mathrm{\&beta;}}_{i-1}(x,y)& G_i(x,y)=0\\ B_{i-1}(x,y)& G_i(x,y)\&NotEqual;0\end{array}\right.---\left(4\right)$

B _i(x, y) and I _i(x, y) is respectively the i-th frame background image and the gray-scale value of original image at pixel (x, y) place, G _i(x, y) is that the i-th frame original image carries out the value of the differentiated binary image of background at pixel (x, y) place, and α is context update weight, 0≤α≤1.

The step that the vehicle of the extraction video features of described step (2-1) exists ratio is:

It is vehicle area coverage and the ratio of the surveyed area total area in crossing that vehicle exists than Q, adopts background difference and Moving Window to exist vehicle and extracts than Q characteristic quantity:

Background differential representation is as follows:

Make I _kfor kth frame gray level image, B _kfor background gray level image, by I _kwith B _kdiffer from, obtain background subtraction image M _k, to M _kcarry out medium filtering, and utilize Otsu algorithm to provide threshold value T ₁to M _kcarry out Iamge Segmentation, obtain binary image S _k; Calculating formula is as follows:

$S_k(x,y)=\left\{\begin{array}{cc}1& |I_k(x,y)-B_k(x,y)|>T_1\\ 0& \mathrm{otherwise}\end{array}\right.---\left(5\right)$

Moving Window is expressed as follows:

The horizontal detection window utilizing a height to be surveyed area vertical height 1/10th, rearwardly moves from queue head.Due to object be transformed into two dimensional surface space from three-dimensional space time, distance camera far away, less in two dimensional surface, therefore Moving Window detection vehicle there is the general geometric representation of method as shown in Figure 5.Detection window is scanned from team's head to tail of the queue, if the horizontal zone meeting existence condition is maximum be numbered i, then L _k=i (0≤i≤10).With L _kthere is ratio in the vehicle characterizing the i-th two field picture.Median point due to data sequence has stronger robustness, and this algorithm adopts the time span of N continuous two field picture as the least unit time, and the intermediate value of N continuous frame testing result sequence compares result as the vehicle existence of unit interval.

Therefore, vehicle existence than Q characteristic quantity is:

Q＝med(L _k),(1≤k≤N)(6)

L _k＝max(I)(7)

$I=\left\{\begin{array}{cc}i|P_E\left(i\right)=\frac{E_i}{N_i}>T_E,& 0\&le;i\&le;10\end{array}\right\}---\left(8\right)$

Formula (8) represents that I meets the set that there are all zone numbers of detection threshold in kth two field picture; Formula (7) represents that the maximal value of getting in set I exists ratio to the vehicle characterizing kth two field picture, uses L _krepresent; Formula (6) represents in the video sequence of a N continuous frame, gets all L _kintermediate value Q characterize current N two field picture vehicle exist compare result.

The step of the motion pixel ratio of the extraction video features of described step (2-1) is:

Motion pixel is the ratio of motor image vegetarian refreshments and whole pixel in surveyed area than S, and definition is

S＝med(P _M(k)),(1≤k≤N)(9)

$P_M\left(k\right)=\frac{M_k}{N_k}---\left(10\right)$

Wherein, M _kfor motion number of pixels in kth two field picture surveyed area, N _kfor all number of pixels in kth two field picture surveyed area, formula (9) represents in the video sequence of a N continuous frame, gets all P _mk vehicle that the intermediate value S of () characterizes current N two field picture exists and compares result.

The fuzzy recognizer of band feedback is utilized to carry out image recognition in described step (2-1):

Exist than Q and motion pixel than the input of S as fuzzy recognizer using the vehicle gathered, after the obfuscation of fuzzy recognizer, fuzzy reasoning and ambiguity solution, fuzzy recognizer exports current overflow testing result; When being in the transition states such as queuing formation, dissipation when vehicle, only cannot differentiate traffic overflow state with above-mentioned two indexes, therefore using the feed back input amount of the output " current overflow testing result " of fuzzy recognizer as fuzzy recognizer, in the practical work process of this fuzzy recognizer, the physical meaning of feed back input amount is " last overflow testing result ", and this initialization of variable value is " without overflow ".Therefore, fuzzy recognizer is real is triple input single output structure, as shown in Figure 3.

Wherein, fuzzy recognizer is logically made up of four parts, i.e. fuzzy diagnosis rule, fuzzification process, fuzzy reasoning process and deblurring process, through above four step process, this fuzzy recognizer can export fuzzy diagnosis result, the degree of overflow result namely in the present invention.

When vehicle is in queuing formation, dissipation transition state, " the current overflow testing result " that exported by fuzzy recognizer is as the feed back input amount of fuzzy recognizer, in the practical work process of fuzzy recognizer, feed back input amount is " last overflow testing result ", and this initialization of variable value is " without overflow ".

It is that { without N, low L, middle M, high H}, represent Current vehicle queuing ratio " without queuing up ", " low ", " medium " and " height " respectively that described vehicle exists than the fuzzy subset of Q;

The fuzzy subset of described motion pixel ratio S is that { blocking C, unobstructed F, transition T}, represent that the inner wagon flow traffic status in current crossing is " blocking ", " unobstructed " and " transition " respectively;

The fuzzy subset of described current overflow testing result O is that { without N, slight L, moderate M, severe H}, represent crossing flooded conditions "None", " slightly ", " moderate ", " severe " respectively.

Vehicle existence is respectively trapezoidal, trapezoidal and triangular function than the membership function of Q, motion pixel ratio S and current overflow testing result O ternary.Fuzzy inference rule: than Q be if Current vehicle exists motion pixel than S is last overflow testing result is then current overflow testing result is fuzzy implication relation adopts Mamdani " maximum-minimum " rationalistic method, obtains fuzzy control search table by gravity model appoach ambiguity solution.Described fuzzy control search table in the real-time testing process of traffic overflow for inquiry.This fuzzy recognizer introduces last overflow testing result as feedback quantity, thus realizes judging the flooded conditions of the wagon flow being in transient motion state.

The location of described step (2-2) signal lamp with identify it is adopt the signal lamp automatic identifying method under a kind of complex background, comprise Intensity segmentation, K mean cluster and prospect histogram analysis means, the color of signal lamp, directional information identified.The method concrete steps are as follows:

Image is converted into HSV space by A, and adopt Da-Jin algorithm to provide Intensity segmentation threshold value, its basic thought finds out a segmentation threshold, makes to be had maximum between-cluster variance or infima species internal variance by its pixel gray-scale value being divided into two classes.Iamge Segmentation based on brightness is regular such as formula shown in (11):

$f_3(x,y)=\left\{\begin{array}{cc}{f}_2(x,y)& V(x,y)\&GreaterEqual;T^{*}\\ 0& V(x,y)<T^{*}\end{array}\right.---\left(11\right)$

B extracts based on the signal lamp position of geometric properties: on the basis of Intensity segmentation, carry out gray processing process to image, then carry out contour detecting, and do boundary rectangle to the profile detected.Remember that single rectangular area is R, according to shape and the positional information of R, introduce new regulation and filtering is carried out to noise.New regulation comprises area information, positional information and aspect ratio information etc. as constraint condition.

Red, the green state of C definition signal lamp is signal lamp key message, and yellow and digital countdown state is non-critical information.The key message of signal lamp can be described by position coordinates, color, shape and direction.First adopt statistic of classification to obtain object detection area, rule is:

If any two candidate region R _ijcentral point L _ij(x _ij, y _ij) between Euclidean distance d meet

$d=\sqrt{{(x_{\mathrm{ij}1}-x_{\mathrm{ij}2})}^2+{(y_{\mathrm{ij}1}-y_{\mathrm{ij}2})}^2}<T6---\left(12\right)$

Then think 2 coincidences, otherwise think that do not overlap at 2.Wherein, R _ijbe the jth candidate region in the i-th two field picture, T6 is distance threshold.In target vides sequence, total N two field picture, comprises the signal period that several are complete; Wherein, in the i-th two field picture, the total number in candidate region is M _i, to L _ij(i=(1,2 ..., N), j=(1,2 ..., M _i)) classify according to above-mentioned rule, obtain not coincident central point L _k, k=(1,2 ..., S), proportion is respectively C in the video sequence _k, k=(1,2 ..., S), S is the number of not coincident central point.Because the appearance in the video sequence of signal lamp non-critical information and noise has the features such as randomness is strong, the duration is short, therefore basis

C _k＞T7(13)

All central points are screened, key signal lamp location sets D={R can be obtained _ij| i=(1,2 ..., N), j=(1,2 ..., K _i), wherein, T7 is proportion threshold value, K _iit is key signal lamp areal in the i-th two field picture.

D adopts K means clustering algorithm to obtain red, green color cluster center (r _c, g _c), then region R to be measured _ij(i=(1,2 ..., N2), j=(1,2 ..., K _i)) color decision rule be

$\mathrm{color}=\left\{\begin{array}{cc}R& |h_{\mathrm{ij}}-r_c|=\min \left\{\right|h_{\mathrm{ij}}-r_c|,|h_{\mathrm{ij}}-g_c\left|\right\}\\ G& |h_{\mathrm{ij}}-r_c|=\min \left\{\right|h_{\mathrm{ij}}-r_c|,|h_{\mathrm{ij}}-g_c\left|\right\}\end{array}\right.---\left(14\right)$

E is the flexible identification realized circular and arrow-shaped signal lamp, employing prospect histogram analysis method: summarize the regularity of distribution of signal with different type lamp prospect histogram under different scanning mode, and mated with generalise results by the histogram in region to be measured, and then draw signal lamp shape and directional information.

The step of described step (3) is: crossing interior vehicle queuing situation and signal lamp recognition result are comprehensively analyzed, and according to the true and false overflow decision rule in crossing, provides overflow recognition result; The frequency that described overflow recognition result provides sets according to user's actual needs;

Divide according to overflow character, overflow result comprises true overflow and false overflow two kinds of situations;

Divide according to degree of overflow, overflow result comprise without overflow, slightly, moderate and severe four kinds of situations.

The mode be combined with text prompt by real-time change curve is shown by overflow result, and result provided frequency for 5 seconds, as shown in Figure 4.

By reference to the accompanying drawings the specific embodiment of the present invention is described although above-mentioned; but not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various amendment or distortion that creative work can make still within protection scope of the present invention.

Claims (9)

1., based on a traffic overflow state identification method for video, it is characterized in that, comprise the steps:

Step (1): set up two high-definition cameras at target cross junction, wherein First high-definition camera is monitored the crossing interior vehicle queuing behavior that direction occurs in overflow, and particular location is the lateral mid-point near section end; The crisscross signal lamp of second high-definition camera pair and overflow direction is monitored, and particular location is the section end lateral mid-point with overflow direction vertical direction;

Step (2): comprise two steps arranged side by side: step (2-1) and step (2-2);

Step (2-1): the video image for First high-definition camera machine monitoring carries out pre-service, extracts video features: vehicle exists ratio and motion pixel ratio; Then the fuzzy recognizer of band feedback is utilized to carry out image recognition;

Step (2-2): the image for second high-definition camera machine monitoring processes, realizes location and the identification of signal lamp; Then output signal light recognition result;

Step (3): the result of combining step (2-1) and step (2-2), according to true and false overflow decision rule, draws traffic overflow result;

Step (4): carry out the display of traffic overflow state and report to the police;

The step of the motion pixel ratio of the extraction video features of described step (2-1) is:

Motion pixel is the ratio of motor image vegetarian refreshments and whole pixel in surveyed area than S, and definition is

S＝med(P _M(k)),(1≤k≤N)(9)

$P_M\left(k\right)=\frac{M_k}{N_k}---\left(10\right)$

Wherein, M _kfor motion number of pixels in kth two field picture surveyed area, N _kfor all number of pixels in kth two field picture surveyed area, formula (9) represents in the video sequence of a N continuous frame, gets all P _mk vehicle that the intermediate value S of () characterizes current N two field picture exists and compares result.

2. a kind of traffic overflow state identification method based on video as claimed in claim 1, it is characterized in that, the pretreated step of described step (2-1) is: the acquisition of background updating and renewal, moving object detection, extracts vehicle target in crossing.

3. a kind of traffic overflow state identification method based on video as claimed in claim 1, is characterized in that, the step that the vehicle of the extraction video features of described step (2-1) exists ratio is:

It is vehicle area coverage and the ratio of the surveyed area total area in crossing that vehicle exists than Q, adopts background difference and Moving Window to exist vehicle and extracts than Q characteristic quantity:

Therefore, vehicle existence than Q characteristic quantity is:

Q＝med(L _k),(1≤k≤N)(6)

L _k＝max(I)(7)

$I=\left\{\begin{array}{cc}i|P_E\left(i\right)=\frac{E_i}{N_i}>T_E,& 0\&le;i\&le;10\end{array}\right\}---\left(8\right)$

Formula (8) represents that I meets the set that there are all zone numbers of detection threshold in kth two field picture; Formula (7) represents that the maximal value of getting in set I exists ratio to the vehicle characterizing kth two field picture, uses L _krepresent; Formula (6) represents in the video sequence of a N continuous frame, gets all L _kintermediate value Q characterize current N two field picture vehicle exist compare result;

I meets the zone number that there is detection threshold in kth two field picture; P _eratio is there is in (i) for meeting the vehicle that there is the i-th surveyed area of detection threshold in kth two field picture; T _efor current detection region memory is at the detection threshold of vehicle; N _ifor number of pixels all in surveyed area; E _ifor detection window moves to the number of pixels in the i-th surveyed area.

4. a kind of traffic overflow state identification method based on video as claimed in claim 1, is characterized in that, utilizes the fuzzy recognizer of band feedback to carry out image recognition in described step (2-1):

Exist than Q and motion pixel than the input of S as fuzzy recognizer using the vehicle gathered, after the obfuscation of fuzzy recognizer, fuzzy reasoning and ambiguity solution, fuzzy recognizer exports current overflow testing result; Wherein, fuzzy recognizer is logically made up of four parts, i.e. fuzzy diagnosis rule, fuzzification process, fuzzy reasoning process and deblurring process, and through above four step process, this fuzzy recognizer exports fuzzy diagnosis result;

When vehicle is in queuing formation, dissipation transition state, " the current overflow testing result " that exported by fuzzy recognizer is as the feed back input amount of fuzzy recognizer, in the practical work process of fuzzy recognizer, feed back input amount is " last overflow testing result ", and feed back input amount initialization value is " without overflow ";

It is that { without N, low L, middle M, high H}, represent Current vehicle queuing ratio " without queuing up ", " low ", " medium " and " height " respectively that described vehicle exists than the fuzzy subset of Q;

The fuzzy subset of described motion pixel ratio S is that { blocking C, unobstructed F, transition T}, represent that the inner wagon flow traffic status in current crossing is " blocking ", " unobstructed " and " transition " respectively;

The fuzzy subset of described current overflow testing result O is that { without N, slight L, moderate M, severe H}, represent crossing flooded conditions "None", " slightly ", " moderate ", " severe " respectively;

Vehicle existence is respectively trapezoidal, trapezoidal and triangular function than the membership function of Q, motion pixel ratio S and current overflow testing result O ternary; Fuzzy inference rule: than Q be if Current vehicle exists motion pixel than S is last overflow testing result is then current overflow testing result is fuzzy implication relation adopts Mamdani " maximum-minimum " rationalistic method, obtains fuzzy control search table by gravity model appoach ambiguity solution; Described fuzzy control search table in the real-time testing process of traffic overflow for inquiry; Described fuzzy recognizer introduces last overflow testing result as feedback quantity, thus realizes judging the flooded conditions of the wagon flow being in transient motion state.

5. a kind of traffic overflow state identification method based on video as claimed in claim 1, it is characterized in that, location and the identification of described step (2-2) signal lamp adopt the signal lamp automatic identifying method under a kind of complex background, comprise Intensity segmentation, K mean cluster and prospect histogram analysis means, the color of signal lamp, directional information are identified.

6. a kind of traffic overflow state identification method based on video as claimed in claim 1, it is characterized in that, the step of described step (3) is: crossing interior vehicle queuing situation and signal lamp recognition result are comprehensively analyzed, according to the true and false overflow decision rule in crossing, provide overflow recognition result; The frequency that described overflow recognition result provides sets according to user's actual needs;

Divide according to overflow character, overflow result comprises true overflow and false overflow two kinds of situations;

Divide according to degree of overflow, overflow result comprise without overflow, slightly, moderate and severe four kinds of situations.

7. a kind of traffic overflow state identification method based on video as claimed in claim 2, is characterized in that, the acquisition of described background updating adopts statistic histogram method, shown in formula specific as follows:

$P(x,y,k)=\left\{\begin{array}{ccc}P(x,y,k)+1& if& I_i(x,y)=k\\ P(x,y,k)& if& I_i(x,y)\&NotEqual;k\end{array}\right.i=1,2,3,...,N---\left(1\right)$

B(x,y)＝k(2)

ifP(x,y,k)＝max(P)k＝0,1,2,…,255(3)

B (x, y) is the gray-scale value that the background image of foundation is corresponding at pixel (x, y) place, I _i(x, y) represents the gray-scale value that the i-th frame original image is corresponding at pixel (x, y) place, and N represents statistics totalframes; P (x, y, k) represents the number of times that pixel (x, y) place gray-scale value k occurs; Formula (3) represents at point (x, y) place, gets the gray-scale value of the maximum gray-scale value of occurrence number image as a setting.

8. a kind of traffic overflow state identification method based on video as claimed in claim 2, is characterized in that, the renewal of described background updating adopts selectivity background update method, represents by formula (4):

$B_i(x,y)=\left\{\begin{array}{cc}{\mathrm{\&alpha;I}}_i(x,y)+(1-\mathrm{\&alpha;})B_{i-1}(x,y)& G_i(x,y)=0\\ B_{i-1}(x,y)& G_i(x,y)\&NotEqual;0\end{array}\right\}---\left(4\right)$

B _i(x, y) and I _i(x, y) is respectively the i-th frame background image and the gray-scale value of original image at pixel (x, y) place, G _i(x, y) is that the i-th frame original image carries out the value of the differentiated binary image of background at pixel (x, y) place, and α is context update weight, 0≤α≤1.

9. a kind of traffic overflow state identification method based on video as claimed in claim 3, it is characterized in that, background differential representation is as follows:

Make I _kfor kth frame gray level image, B _kfor background gray level image, by I _kwith B _kdiffer from, obtain background subtraction image M _k, to M _kcarry out medium filtering, and utilize Otsu algorithm to provide threshold value T ₁to M _kcarry out Iamge Segmentation, obtain binary image S _k; Calculating formula is as follows:

$S_k(x,y)=\left\{\begin{array}{cc}1& |I_k(x,y)-B_k(x,y)|>T_1\\ 0& otherwise\end{array}\right.---\left(5\right)$

Wherein, (x, y) represents pixel, x and y represents horizontal ordinate and the ordinate of each pixel of image respectively;

Moving Window is expressed as follows:

The horizontal detection window utilizing a height to be surveyed area vertical height 1/10th, rearwardly moves from queue head; Detection window is scanned from team's head to tail of the queue, if the horizontal zone meeting existence condition is maximum be numbered i, then L _k=i (0≤i≤10); With L _kthere is ratio in the vehicle characterizing kth two field picture; Median point due to data sequence has stronger robustness, adopts the time span of N continuous two field picture as the least unit time, and the intermediate value of N continuous frame testing result sequence compares result as the vehicle existence of unit interval.