CN202563526U - Transportation vehicle detection and recognition system based on video - Google Patents
Transportation vehicle detection and recognition system based on video Download PDFInfo
- Publication number
- CN202563526U CN202563526U CN2012201121029U CN201220112102U CN202563526U CN 202563526 U CN202563526 U CN 202563526U CN 2012201121029 U CN2012201121029 U CN 2012201121029U CN 201220112102 U CN201220112102 U CN 201220112102U CN 202563526 U CN202563526 U CN 202563526U
- Authority
- CN
- China
- Prior art keywords
- vehicular traffic
- vehicle
- sample image
- image
- characteristic
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Images
Abstract
The utility model discloses a transportation vehicle detection and recognition system based on video. The transportation vehicle detection and recognition system based on the video comprises a target characteristic data bank acquiring device, a motion target detection device and a transportation vehicle recognition device. The transportation vehicle detection and recognition system establishes and stores a transportation vehicle sample target characteristic data bank through the target characteristic data bank acquiring device. After the motion target detection device detects a motion target area, the transportation vehicle recognition device establishes a vehicle search window and recognizes vehicles. The transportation vehicle detection and recognition system based on the video solves the technical problems that a detection system based on video in existing technology is poor in distinguishing transportation vehicle ability and low in distinguishing accuracy under the conditions of obvious background motions, queueing of vehicles, shielding of the vehicles and the like, and has the advantages of being good in vehicle distinguishing ability and high in accuracy.
Description
Technical field
The utility model relates to Flame Image Process and mode identification technology, specifically is that a kind of vehicular traffic based on video detects recognition system.
Background technology
In recent years; Fast development along with electronic industry especially novel sensor and mass storage devices; The video monitoring technology is with its dirigibility and the comprehensive favor that more and more receives people, and it is modern high technologies such as core, integrated use computing machine, control technology with the real-time dynamic information; Fast video image is handled in real time flexibly, reached real-time monitoring.Vehicle detection is the important step in the traffic system, and for traffic monitoring, traffic control provide information, vehicle checking method commonly used comprises coil detection, detections of radar, laser detection etc.Installation and maintenance engineering amount that coil detects are big, destroy the road surface, influence the life-span of road, and laser detection and detections of radar cost height not only, and easily human body is worked the mischief.
Vehicle detection recognition technology based on video can extract information of vehicles from video image, not only flexible, does not destroy the road surface, and can a large amount of detection information be provided for traffic monitoring, for traffic administration provides visual information.As emerging vehicle checking method, receive people's attention based on the vehicle detection recognition technology of video day by day.
Chinese patent document CN101719217A has proposed a kind of model recognition system and method based on the relaxation algorithm; This system comprises video camera, image pick-up card, Flame Image Process and the display device that links to each other with data line; This method is to use earlier the camera acquisition automobile video frequency; The image of gathering is set up based on Gauss's average background model, with this as detecting foreground area; Through frame difference method of operating, foreground area segmented extraction from background at moving target place is come out as foreground image again, with the relaxation algorithm foreground image is handled and obtained vehicle commander and overall height, judge the vehicle result according to national vehicle criterion of identification then.Chinese patent document CN102332167A discloses the object detection method of vehicle and pedestrian in a kind of intelligent transportation monitoring; Sequence of frames of video is carried out the initialization of background model, independently set up the mixed Gaussian background component model of saturation degree component and luminance component and get the component average; Present frame in the sequence of frames of video with background frames phase difference, is carried out removing shade and noise again and carrying out morphologic filtering after the binary conversion treatment to the prospect frame; Upgrade weights, average and the variance of the component of the mixed Gaussian background model that obtains with upgrading the factor; The Jeffrey value of each distribution in the mixed Gaussian background model after moving target pixel point value that will mate and the renewal compares, and utilizes the Jeffrey value to judge whether the moving target pixel belongs to the foreground point.Chinese patent document CN102222346A discloses a kind of vehicle detection and tracking, at first each two field picture in the video is set up the Gaussian Background model; Utilize frame difference method that adjacent two frames are done difference processing, obtain rough moving region and stagnant zone; Stagnant zone to obtaining carries out context update, and the moving region is not upgraded; Background image to after the renewal of current frame image and acquisition is done difference, obtains the accurate movement zone; Utilize each pixel matching process that the adjacent two frame moving region images that obtain are found out the overlapping region, and compare overlapping region and given threshold size; If greater than given threshold value, then judging whether to take place target, the overlapping region overlaps; If then calculate the length breadth ratio of the first frame moving region in adjacent two frames, through this this moving vehicle of ratio detection and tracking; If not, then be judged as same vehicle; If, then obtaining the minimum boundary rectangle of a plurality of target frames less than given threshold value, the overlapping region comes correctly to vehicle detection and tracking.
Above-mentioned three pieces of patent documentations are in the vehicle detection process, basically all based on the single pixel modeling of image; The identification vehicle, utilization be image single-point information, judge that current pixel point is foreground pixel (vehicle pixel) or background pixel (non-vehicle pixel); Do not consider the relativity problem between the pixel, i.e. the surface information of image is so can exist following problem: for example in the vehicle detection process; Branch slightly rock or the situation of video camera slight jitter under, above-mentioned patent documentation can satisfy vehicle detection identification, but rocks apparent in view at branch; During the apparent in view situation of DE Camera Shake; Vehicle detection identification based on single pixel modeling can produce a lot of wrong reports, and especially abundant more in image edge information, the wrong report that is produced is disturbed just many more; When the situation that vehicle adhesion and vehicle block; Can not effectively distinguish each independent vehicle; Can become several vehicle identification a vehicle and become several vehicles or the like to a vehicle identification, thereby influence the tracking of each independent vehicle, and then behavioural analysis (illegal the driving in the wrong direction of influence back vehicle; Make a dash across the red light illegal lane change etc.); In addition, for pedestrian's target and the vehicle target in the traffic image, can not distinguish effectively.
In video identification technology; Need carry out feature extraction to image information; The characteristic of extracting comprises Harr characteristic, integration characteristic and angle character, and the Haar characteristic is divided three classes: edge feature, linear feature, central feature and diagonal line characteristic are combined into feature templates.Adularescent and two kinds of rectangles of black in the feature templates, and the eigenwert that defines this template be the white rectangle pixel with deduct the black rectangle pixel and.The integration characteristic refer in the image arbitrarily the numerical value of any equal from the upper left corner of image to the rectangular area that this point is constituted in institute a pixel sum is arranged.The gradient feature description information that changes of regional areas such as edge of image, angle point, have stronger robustness for the variation of illumination, be widely used in target signature description, images match and the target detection.
The utility model content
For this reason; The utility model is to be solved be existing vehicle detection based on video background motion significantly, vehicle adhesion and vehicle block etc. and to distinguish vehicular traffic ability, technical matters that resolving accuracy is not high under the situation, provide a kind of and have preferably that vehicle separating capacity, high-precision vehicular traffic based on video detect recognition system.
For solving the problems of the technologies described above, the technical scheme that the utility model adopts is following:
A kind of vehicular traffic based on video detects recognition system, comprising:
(1) the target signature database obtains equipment; Store the vehicular traffic sample image carried out pattern-recognition training after; According to the vehicular traffic sample object property data base that normalization type, angular type and type of vehicle obtain, said target signature database obtains equipment and comprises
The sample image pretreater; From traffic video, cut out non-vehicular traffic sample image and said vehicular traffic sample image; Said vehicular traffic sample image and said non-vehicular traffic sample image are carried out the image pre-service, and said image pre-service comprises normalization processing and gray processing processing;
Vehicular traffic sample image sorter; Comprise the normalization sorter that is connected with said sample image pretreater; Said normalization sorter is connected with the vehicle class sorter with the angle sorter, and the said vehicular traffic sample image after normalization is handled carries out angle classification and vehicle class classification;
The sample image feature deriving means; Be connected with said vehicular traffic sample image sorter; Said non-vehicular traffic sample image and sorted said vehicular traffic sample image are extracted the characteristics of image after its gray processing is handled, and said characteristics of image comprises integration characteristic, Haar characteristic and gradient characteristic;
Sample image pattern-recognition training aids; Be connected with said sample image feature deriving means; The said integration characteristic of said vehicular traffic sample image and said non-vehicular traffic sample image, said Haar characteristic and said gradient characteristic are carried out the pattern-recognition training, and Adaboost algorithm or SVM algorithm are adopted in said pattern-recognition training;
The target signature archival memory is connected with said sample image pattern-recognition training aids, stores said sample image pattern-recognition training aids and obtains vehicular traffic sample object property data base according to normalization type, angular type and type of vehicle;
(2) moving object detection equipment is connected with traffic video equipment, from the traffic video image sequence that obtains, detects motion target area;
(3) vehicular traffic identification equipment; Carry out data transmission with said moving object detection equipment; Load said vehicular traffic sample object property data base, the said motion target area that said moving object detection recognition of devices is gone out carries out vehicle identification, and said vehicular traffic identification equipment comprises
The vehicle search window is provided with a vehicle search window in said motion target area;
The search window feature deriving means is connected with said vehicle search window, obtains integration characteristic, Haar characteristic and the gradient characteristic of the gray level image in the said vehicle search window;
Comparative device; Be connected with said target signature archival memory with said search window feature deriving means; Said integration characteristic, said Haar characteristic and the said gradient characteristic of said search window feature deriving means acquisition and the said vehicular traffic sample object property data base in the said target signature archival memory are compared, the vehicle in the said vehicle search window is discerned.
Described vehicular traffic detects recognition system, and said normalization sorter is handled the sample image that obtains 24 * 24,32 * 32,48 * 48 3 kinds of sizes with said vehicular traffic sample image and said non-vehicular traffic sample image normalization.
Described vehicular traffic detects recognition system; Said angle sorter is classified according to body sway angle in the said vehicular traffic sample image, is divided into :-30 ° to+30 ° ,-25 ° to-55 ° and+25 ° to+55 ° ,-50 ° to-80 ° and+50 ° to+80 ° three types.
Described vehicular traffic detects recognition system, and said vehicle class sorter is divided into three types of compact car, the large car that does not comprise public transit vehicle and public transit vehicles with said vehicular traffic sample image.
The technique scheme of the utility model is compared prior art and is had the following advantages:
(1) the described vehicular traffic based on video of the utility model detects recognition system; Comprise that the target signature database obtains equipment; Moving object detection equipment and vehicular traffic identification equipment; Obtain in the equipment at said target signature database; Come vehicular traffic sample image and non-vehicular traffic sample image are carried out normalization and gray processing processing through the sample image pretreater, pretreated vehicle sample image is carried out angle classification and vehicle class classification, come non-vehicular traffic sample image and sorted vehicle sample are carried out feature extraction through the sample image feature deriving means through vehicular traffic sample image sorter; Through sample image pattern-recognition training aids the said characteristic of extracting is trained and pattern-recognition then, set up vehicular traffic sample object property data base and store.After the moving object detection Equipment Inspection goes out motion target area, by the vehicular traffic identification equipment set up the vehicle search window go forward side by side the driving identification.In setting up vehicular traffic sample object property data base process, the vehicular traffic sample image is classified according to normalization kind, angle kind and vehicle class, carry out feature extraction then, improved accuracy of identification and operation time; In the training process of vehicle sample; The utility model utilizes Adaboost algorithm for pattern recognition or SVM algorithm identified vehicular traffic; This algorithm in the vehicular traffic image between each pixel correlativity take into account; What the Adaboost algorithm for pattern recognition was discerned utilization to vehicle is the surface information of image, and dot information and surface information are organically combined, the interference that can overcome that branch rocks, situation such as DE Camera Shake and vehicle adhesion is brought; Non-vehicle target (like the pedestrian) in the traffic image can be distinguished, and can better, more effectively carry out vehicle detection identification.In the process of carrying out vehicle identification; Come said motion target area is searched for through definition vehicle search window; Can come the setting search step-length according to the size of motion target area, guarantee the speed and the precision of search, improve search efficiency and search quality.The described vehicular traffic based on video of the utility model detects recognition system; Through setting up vehicular traffic target signature mathematical model; Research vehicle target Feature Extraction Technology is set up vehicular traffic sample object property data base, the automatic renewal technology of research background; Realize the pure Video Detection of traffic events and traffic behavior; It calculates accurately, the few characteristics of erroneous judgement provide accurately, practicality, technological means intuitively, can adapt to the road conditions of urban transportation more complicated well, has high practicality.
(2) the described vehicular traffic of the utility model detects recognition system; The sample image that obtains 24 * 24,32 * 32,48 * 48 3 kinds of sizes is handled in said vehicular traffic sample image and said non-vehicular traffic sample image normalization; These three kinds of image patterns are evenly distributed in the size of the video image that obtains, and help later vehicle sample training and video image is carried out vehicle identification.
(3) the described vehicular traffic of the utility model detects recognition system; Classify according to body sway angle in the said vehicular traffic sample image, be divided into :-30 ° to+30 ° ,-25 ° to-55 ° and+25 ° to+55 ° ,-50 ° to-80 ° and+50 ° to+80 ° three types.Adopt this mode classification; Both comprehensive vehicle obliquities is reacted the characteristics of image of all angles comprehensively, reduces the kind of angle again as far as possible; Thereby reduce the data type in the vehicular traffic sample object property data base, improve the matching speed in the identifying.
(4) the described vehicular traffic of the utility model detects recognition system; Said vehicle class sorter is divided into three types of compact car, the large car that does not comprise public transit vehicle and public transit vehicles with said vehicular traffic sample image; Because these three types of vehicles have not only been summarized the type of most vehicles; And the data message of these vehicles is that traffic administration and road conditions detect more information is provided, and therefore divides three types for this reason with vehicle class, the classification when being convenient to vehicle identification.
(5) the described vehicular traffic of the utility model detects recognition system; The training process of vehicle sample training need be collected a large amount of vehicular traffic sample images and non-vehicular traffic sample image; Non-vehicular traffic sample image major part derives from the non-vehicular traffic part in the traffic video image, and the vehicular traffic sample image is as positive sample, and non-vehicular traffic sample image is as negative sample; Positive sample and negative sample are included in the training aids of vehicle detection together; Obtain to have the vehicle template of characteristic by force through training, can effectively must distinguish vehicle image and non-vehicle image, improve the precision of vehicle identification.
Description of drawings
For the content that makes the utility model is more clearly understood,, the utility model is done further detailed explanation, wherein below according to the specific embodiment of the utility model and combine accompanying drawing
Fig. 1 is that the vehicular traffic based on video of an embodiment of the utility model detects the recognition system structured flowchart;
Fig. 2 detect for the vehicular traffic based on video shown in Figure 1 recognition system recognition methods set up vehicular traffic sample object property data base process flow diagram;
Fig. 3 is the moving object detection process flow diagram that detects the recognition methods of recognition system based on the vehicular traffic of video shown in Figure 1;
Fig. 4 detect for the vehicular traffic based on video shown in Figure 1 recognition system recognition methods motion target area is carried out the vehicle identification process flow diagram;
Fig. 5 and Fig. 6 carry out the simulated effect figure that the moving region is detected for using system embodiment illustrated in fig. 1 to a two field picture;
Fig. 7 and Fig. 8 simulated effect figure for using system embodiment illustrated in fig. 1 that image is discerned.
Reference numeral is expressed as among the figure: 1-moving object detection equipment; 2-target signature database obtains equipment; 21-sample image pretreater; 22-vehicular traffic sample image sorter; 23-sample image feature deriving means; 24-sample image pattern-recognition training aids; 25-target signature archival memory; 3-vehicular traffic identification equipment; 31-vehicle search window; 32-search window feature deriving means; The 33-comparative device.
Embodiment
Embodiment 1:
Provide the concrete embodiment that the described a kind of vehicular traffic based on video of the utility model detects recognition system below, referring to shown in Figure 1, the vehicular traffic based on video in the present embodiment detects recognition system, comprising:
(1) the target signature database obtains equipment 2; Store the vehicular traffic sample image carried out pattern-recognition training after; The vehicular traffic sample object property data base that obtains according to normalization type, angular type and type of vehicle; Comprise sample image pretreater 21, vehicular traffic sample image sorter 22, sample image feature deriving means 23, sample image pattern-recognition training aids 24 and target signature archival memory 25, respectively said units is described in detail below.
Sample image pretreater 21; From traffic video, cut out non-vehicular traffic sample image and said vehicular traffic sample image; Said vehicular traffic sample image and said non-vehicular traffic sample image are carried out the image pre-service, and said image pre-service comprises normalization processing and gray processing processing.In the present embodiment; Take from traffic video, to cut out 8000 vehicular traffic sample images and 12000 non-vehicular traffic sample images; 8000 said vehicular traffic sample images and 12000 said non-vehicular traffic sample images are carried out normalization processing and gray processing processing, and the sample image that obtains 24 * 24,32 * 32,48 * 48 3 kinds of sizes is handled in said normalization.
Vehicular traffic sample image sorter 22 is connected with said sample image pretreater 21, and the said vehicular traffic sample image after normalization is handled carries out angle classification and vehicle class classification.The classification of said angle is meant according to body sway angle in the said vehicular traffic sample image classifies (is 0 ° of direction with the scene visual direction), is divided in the present embodiment :-30 ° to+30 ° ,-25 ° to-55 ° and+25 ° to+55 ° ,-50 ° to-80 ° and+50 ° to+80 ° three types.The classification of said vehicle class is divided into three types of compact car, the large car that does not comprise public transit vehicle and public transit vehicles with said vehicular traffic sample image in the present embodiment.
Sample image feature deriving means 23; Be connected with said vehicular traffic sample image sorter 22, said non-vehicular traffic sample image and sorted said vehicular traffic sample image are extracted integration characteristic, Haar characteristic and gradient characteristic after its gray processing is handled.
The integration characteristic be meant in the image arbitrarily the numerical value of any equal from the upper left corner of image to the rectangular area that this point is constituted in institute have a pixel and.The Haar characteristic is divided three classes: edge feature, linear feature, central feature and diagonal line characteristic are combined into feature templates.Adularescent and two kinds of rectangles of black in the feature templates, and the eigenwert that defines this template be the white rectangle pixel with deduct the black rectangle pixel and.Confirming that the quantity of Harr-like characteristic after the characteristic formp just depends on the size of training sample image matrix; Feature templates is placed arbitrarily in subwindow; A kind of form is called a kind of characteristic, and the characteristic of finding out all subwindows is the basis of carrying out the weak typing training.The gradient feature description information that changes of regional areas such as edge of image, angle point, have stronger robustness for the variation of illumination, be widely used in target signature description, images match and the target detection.The integration characteristic of image, Haar characteristic, gradient characteristic have detailed explanation and concrete preparation method in recognition of face, the identification of head shoulder of OpenCV the inside, adopt method of the prior art can obtain integration characteristic, Haar characteristic and the gradient characteristic of vehicular traffic sample image.The full name of OpenCV is Open Source Computer Vision Library, and OpenCV was set up by Intel in 1999, was provided support by Willow Garage now.OpenCV is a cross-platform computer vision storehouse based on BSD licence mandate (increasing income) distribution, may operate on Linux, Windows and the Mac OS operating system.Its lightweight and efficient---constitute by a series of C functions and a small amount of C++ class, the interface of language such as Python, Ruby, MATLAB is provided simultaneously, realized a lot of general-purpose algorithms of Flame Image Process and computer vision aspect.Integration characteristic, Haar characteristic, the gradient characteristic of the method that provides among the OpenCV in can the Direct Recognition image.
Sample image pattern-recognition training aids 24 is connected with said sample image feature deriving means 23, and the said integration characteristic of said vehicular traffic sample image, said Haar characteristic and said gradient characteristic are carried out Adaboost algorithm pattern recognition training.Adaboost is the general-purpose algorithm that a kind of Weak Classifier that will be better than conjecture at random arbitrarily is combined into strong classifier.Viola combines the Adaboost theory at first with the Haar rectangular characteristic; Being applied to people's face detects; Realized the real-time face detection, this method is that a plurality of Weak Classifiers based on single characteristic are cascaded into a strong classifier, then a plurality of strong classifiers is cascaded into the object detector of a completion.The training process of AdaBoost vehicle sample is consistent basically with people's face training process of OpenCV the inside in the present embodiment.
Target signature archival memory 25; Be connected with said sample image pattern-recognition training aids 24, store said sample image pattern-recognition training aids 24 and obtain 27 kinds of vehicular traffic sample object property data bases according to normalization type, angular type and type of vehicle.
(2) moving object detection equipment 1; Be connected with traffic video equipment; The traffic video equipment here can be for being arranged on video capture device such as camera, the video camera etc. on crossing or the highway section; Also can be the video equipment of gathering of noting, from the traffic video image sequence that obtains, detect motion target area.
(3) the vehicular traffic identification equipment 3; Carry out data transmission with said moving object detection equipment 1; Load 27 kinds of said vehicular traffic sample object property data bases of said target signature archival memory 25 storages, the said motion target area that said moving object detection equipment 1 is identified carries out vehicle identification.
Said vehicular traffic identification equipment 3 comprises: vehicle search window 31 is provided with a vehicle search window 31 in said motion target area.
Search window feature deriving means 32 is connected with said vehicle search window 31, obtains integration characteristic, Haar characteristic and the gradient characteristic of the gray level image in the said vehicle search window 31.
Comparative device 33; Be connected with said target signature archival memory 25 with said search window feature deriving means 32; Said integration characteristic, said Haar characteristic and the said gradient characteristic of said search window feature deriving means 32 acquisitions and the said vehicular traffic sample object property data base in the said target signature archival memory 25 are compared, the vehicles in the said vehicle search window 31 are discerned.
A kind of vehicular traffic that uses the described vehicular traffic based on video of present embodiment to detect recognition system detects recognition methods, comprises the steps:
1. set up vehicular traffic sample object property data base, as shown in Figure 2:
S01: from traffic video, cut out 8000 vehicular traffic sample images and 12000 non-vehicular traffic sample images; Carry out normalization processing and gray processing processing through 21 pairs of said vehicular traffic sample images of sample image pretreater and said non-vehicular traffic sample image, the sample image that obtains 24 * 24,32 * 32,48 * 48 3 kinds of sizes is handled in said normalization.
S02: the said vehicular traffic sample image after the normalization processing is carried out angle classification and vehicle class classification; Calling vehicular traffic sample image sorter 22 classifies; Said angle classification is meant according to body sway angle in the said vehicular traffic sample image classifies; Be divided into :-30 ° to+30 ° ,-25 ° to-55 ° and+25 ° to+55 ° ,-50 ° to-80 ° and+50 ° to+80 ° three types, the classification of said vehicle class refers to said vehicular traffic sample image is divided into three types of compact car, the large car that does not comprise public transit vehicle and public transit vehicles.After the classification, extract integration characteristic, Haar characteristic and gradient characteristic after its gray processing is handled through 23 pairs of sorted said vehicular traffic sample images of sample image feature deriving means.
S03: integration characteristic, Haar characteristic and gradient characteristic to said vehicular traffic sample image are carried out Adaboost algorithm pattern recognition training; Obtain 27 kinds of vehicular traffic sample object property data bases according to normalization type, angular type and type of vehicle, here 27 kinds is to make up 27 kinds that (3 * 3 * 3) obtain according to 3 kinds of normalization types, 3 kinds of angular type and 3 kinds of type of vehicle.
2. read the traffic video image sequence, from said traffic video image sequence, detect motion target area.Can adopt in the prior art moving target detecting method to realize here, for example based on the movement technique of code book.
3. load said vehicular traffic sample object property data base, said motion target area is carried out vehicle identification, as shown in Figure 4:
D01: vehicle search window of definition in each said motion target area;
D02: the integration characteristic, Haar characteristic and the gradient characteristic that obtain the gray level image in the said vehicle search window;
D03: said integration characteristic, Haar characteristic and gradient characteristic in the said vehicle search window and said vehicular traffic sample object property data base are compared; Vehicle in the said vehicle search window is discerned; Obtain the angle and the type of vehicle of the highest auto model of matching degree, as vehicle angles and the type of vehicle in the said vehicle search window.
In the present embodiment; Wherein the training of vehicular traffic sample is according to reducing vehicular traffic sample image normalization size 24 * 24,32 * 32,48 * 48 types; Angular type-30 ° to+30 ° one type ,-25 ° to-55 ° and+25 ° to+55 ° one type ,-50 ° to-80 ° and+50 ° to+80 ° one type; Type of vehicle compact car, large car (not comprising public transit vehicle), public transit vehicle carry out classification based training jointly; Be divided into 27 types altogether according to normalization size, angle, type of vehicle, train respectively then, obtain 27 kinds of vehicular traffic target signature database models altogether.Embodiment that the most can conversion, the normalization type can be set to other types according to image size commonly used, as several kinds in the types such as 24 * 24,28 * 28,32 * 32,36 * 36,48 * 48,56 * 56.Angular type also can be set to more multiclass as required and segment; As-20 ° to+20 ° one type ,-20 ° to-50 ° and+20 ° to+50 ° one type ,-50 ° to-80 ° and+50 ° to+80 ° one type; Be divided into the 3-6 class can, be provided with as required.Some types that type of vehicle also can be divided into other are types such as shipping is called a taxi, minibus as increasing, and the type of vehicle of telling as required is provided with.
Embodiment 2:
On the basis of the foregoing description, the step that said vehicular traffic detects recognition methods can also adopt specific method to detect motion target area in 2..The size that said step can be approached motion target area through the mode that enlarges search window successively in 3..Concrete step is following:
2. read the traffic video image sequence, the process that from said traffic video image sequence, detects motion target area is following, referring to shown in Figure 3:
T01: read the traffic video image sequence; Carry out the embossment computing; The principle of embossment computing is to be adjacent upper left pixel to a pixel of image to carry out the difference computing and add a constant (in order to keep the gray scale of image; The constant here can be set to 128), make dark area increase some brightness.Carry out adopting mixed Gauss model then after the embossment computing, said traffic video image sequence is carried out background modeling, obtain bianry image said traffic video image sequence.
Mixed Gaussian background modeling basic thought is: at first define the color of K each pixel of STA representation, wherein generally between 3～5, the K value is big more for the value of K, and it is strong more to handle the fluctuation ability, and speed is slow more.Each state among the K is represented with a gaussian kernel function, the pixel value of these state part expression backgrounds, and remainder is then represented the pixel value of sport foreground.Use variable X like each pixel color value
tExpression, its probability density function can be represented by following K three-dimensional Gaussian function:
η (x, μ
I, t, ∑
I, t) be t i Gaussian distribution constantly, its average is μ
I, t, covariance matrix is a ∑
I, t=(σ
I, t)
2, ω
I, tBe the weight of i Gaussian distribution, and have
Wherein
N representes X in the formula
tDimension.
Along with the continuous variation of scene, the mixed Gauss model of each pixel all needs study constantly to upgrade, and method is with the K in the mixed Gauss model gauss component ω
I, t/ σ
I, tCurrent pixel value X is used in descending ordering then
tCompare one by one with K gauss component in the mixed Gauss model, if current pixel value X
tWith the average μ in i the gauss component
I, tBetween difference less than this gauss component standard deviation sigma
I, t2.5 to 3.5 times, then mixed Gauss model is upgraded, renewal equation is following:
ω
i，t+1＝(1-α)ω
i，t-αM
i，t (3)
μ
i，t+1＝(1-ρ)μ
i，t-ρx
i (4)
(σ
i.t+1)
2＝(1-ρ)(σ
i，t)
2-ρ(x
i-σ
i，t)
T(x
i-μ
i，t) (5)
ρ＝α/ω
i，t (6)
Wherein α is the learning rate of mixed Gauss model, as i gauss component and X
tDuring coupling, M
I, tBe 1, otherwise be 0; If current pixel value X
tWhen all not matching with K gauss component in the mixed Gauss model; Then coming last gauss component in this pixel mixed Gauss model is replaced by new gauss component; The average of new gauss component replaces with current pixel value; Primary standard difference and weight use as default, and after upgrading completion, the weights of each gauss component are again by normalization.
In process to traffic video image Gauss modeling; To mixed Gauss model, carried out the discrete digital processing, optimized the arithmetic speed of mixed Gaussian so greatly; According to the emulation of actual scene, set up the mixed Gaussian background model of K=3 in the present embodiment.
T02: said bianry image is carried out morphologic filtering, use the connected domain analysis to carry out the target area then and demarcate, obtain some motion target areas;
T03: for said motion target area, when the single pixel of traffic video image was in said motion target area, then this said pixel did not carry out the context update process, otherwise carries out the context update process.
Described vehicular traffic detects recognition methods; In the said process that detects motion target area, adopt mixed Gauss model that the traffic video image sequence is carried out background modeling; The bianry image that obtains is carried out morphologic filtering and connected domain analysis; For said motion target area, when the single pixel of traffic video image was in said motion target area, then this said pixel did not carry out the context update process; Otherwise carry out the context update process, improved the accuracy of detection of motion target area.Through the bianry image after the mixed Gaussian modeling is carried out morphologic filtering, can reduce picture noise, help to obtain motion target area.Adopt the zone of this detection moving target; The method that can background modeling and Adaboost pattern-recognition be combined; The Adaboost pattern-recognition need not searched for entire image; Only need behind the mixed Gaussian background modeling, to extract to search for the identification vehicle in the area-of-interest, improve arithmetic speed greatly; Under the smaller situation of the area-of-interest that the mixed Gaussian modeling is extracted, the step-length that can reduce to search for during Adaboost identification improves precision; Under the bigger situation of area-of-interest that the mixed Gaussian modeling is extracted; Adaboost can enlarge step-size in search during identification a little, realizes the adjustment of News Search step-length, and no longer adopts single fixing scaling; Under the situation that guarantees arithmetic speed; Improve precision, the while under remote situation, still can be detected identification in the vehicular traffic target accurately; Can overcome unfavorable factors such as illumination, shade, leaf are rocked, DE Camera Shake, vehicle adhesion and vehicle are blocked just and can be solved preferably.
3. load said vehicular traffic sample object property data base, said motion target area carried out vehicle identification:
D01: vehicle search window of definition in each said motion target area;
D02: the integration characteristic, Haar characteristic and the gradient characteristic that obtain the gray level image in the said vehicle search window;
D03: when said vehicle search window and said vehicular traffic sample object property data base are compared; The said vehicular traffic sample image after handling according to normalization and the sample image size of said non-vehicular traffic sample image are carried out proportional zoom to said vehicle search window, and integration characteristic, Haar characteristic and gradient characteristic and the said vehicular traffic sample object property data base of the gray level image in the said vehicle search window mated.That is to say vehicle search window of definition in each motion target area; Said vehicle search window increases according to certain step-length automatically; According to the size of said vehicle search window, automatically with 27 vehicular traffic sample object property data bases in sample type be that 24 * 24,32 * 32,48 * 48 target signature data compare identification.Successfully spend less than matching threshold if mate in this process, think that then said vehicle search window is non-vehicle region, otherwise, think that then said vehicle search window is a vehicle region, obtain the angular type and the type of vehicle of coupling.
Described vehicular traffic detects recognition methods; When said vehicle search window and said vehicular traffic sample object property data base were compared, the said vehicular traffic sample image after handling according to normalization and the sample image size of said non-vehicular traffic sample image were carried out proportional zoom to said vehicle search window.At first; Choose less vehicle search window and carry out feature extraction and vehicle identification; Progressively enlarge the vehicle search window then and carry out feature extraction and vehicle identification again; The vehicle search window is until expanding the motion target area size to, and vehicle identification finishes, and finally obtains the angular type of vehicle identification and the kind type of vehicle.In this vehicle identification process; When the distant situation of video scene, can choose less step-size in search, during the closer situation of video scene; Can choose the bigger step-length of searching, the accuracy that in the recognition efficiency that improves vehicle, has improved vehicle identification like this.
The described vehicular traffic of the utility model detects recognition system; Be to adopt the method that background modeling and pattern-recognition are combined, can the vehicle that detection recognizes be classified according to angle and type of vehicle that while vehicular traffic target is under remote situation; Still can detect accurately and recognize; Improved accuracy of identification, can overcome simultaneously unfavorable factors such as illumination, shade, leaf are rocked, DE Camera Shake, vehicle adhesion and vehicle are blocked just and can be solved preferably.
Fig. 5 and Fig. 6 carry out the simulated effect figure that the moving region is detected for using the described vehicular traffic of the foregoing description to detect recognition system to a two field picture; Wherein expression is the motion detection regional extent in the wire frame among Fig. 6; Fig. 5 carries out the two-value classification results that modeling obtains to traffic video image sequence utilization mixed Gauss model; Fig. 6 carries out target to Fig. 5 to demarcate two rectangle frames that processing obtains; Be respectively vehicle and pedestrian, two motion target areas also just arranged here---vehicle region and pedestrian zone.
Fig. 7 and Fig. 8 detect the simulated effect figure that recognition system is discerned image for using the described vehicular traffic of the foregoing description; Fig. 7 carries out vehicle identification to the motion target area that Fig. 5 obtains; Here face has filtered out the pedestrian, only is left our interested vehicular traffic; Fig. 8 is that the another one time point recognizes vehicular traffic, has identified two vehicular traffic targets, and one is compact car, and another one is a large car.
As other can conversion embodiment, said Adaboost algorithm can use the SVM algorithm to replace.SVM (support vector machine SVMs) is a kind of trainable machine learning method; The main thought of SVM may be summarized to be 2 points: it is to analyze to the linear separability situation for (1); For the inseparable situation of linearity; Be converted into high-dimensional feature space and make its linear separability through using the Nonlinear Mapping algorithm will hang down the linear inseparable sample of the dimension input space, become possibility thereby make high-dimensional feature space adopt linear algorithm that the nonlinear characteristic of sample is carried out linear analysis; (2) it based on the structural risk minimization theory in feature space construction optimum segmentation lineoid, make learner obtain global optimization, and satisfy certain upper bound with certain probability in the expected risk of whole sample space.Because SVM algorithm training speed is fast, but vehicle identification speed is slow, need a special hardware accelerator support this moment; Other is with above-mentioned embodiment.Can realize the purpose of the utility model equally, belong to the protection domain of the utility model.
Obviously, the foregoing description only be for explanation clearly done for example, and be not qualification to embodiment.For the those of ordinary skill in affiliated field, on the basis of above-mentioned explanation, can also make other multi-form variation or change.Here need not also can't give exhaustive to all embodiments.And conspicuous variation of being extended out thus or change still are among the protection domain of the invention.
Claims (1)
1. the vehicular traffic based on video detects recognition system, it is characterized in that, comprising:
(1) the target signature database obtains equipment; Store the vehicular traffic sample image carried out pattern-recognition training after; According to the vehicular traffic sample object property data base that normalization type, angular type and type of vehicle obtain, said target signature database obtains equipment and comprises
The sample image pretreater; From traffic video, cut out non-vehicular traffic sample image and said vehicular traffic sample image; Said vehicular traffic sample image and said non-vehicular traffic sample image are carried out the image pre-service, and said image pre-service comprises normalization processing and gray processing processing;
Vehicular traffic sample image sorter; Comprise the normalization sorter that is connected with said sample image pretreater; Said normalization sorter is connected with the vehicle class sorter with the angle sorter, and the said vehicular traffic sample image after normalization is handled carries out angle classification and vehicle class classification;
The sample image feature deriving means; Be connected with said vehicular traffic sample image sorter; Said non-vehicular traffic sample image and sorted said vehicular traffic sample image are extracted the characteristics of image after its gray processing is handled, and said characteristics of image comprises integration characteristic, Haar characteristic and gradient characteristic;
Sample image pattern-recognition training aids; Be connected with said sample image feature deriving means; The said integration characteristic of said vehicular traffic sample image and said non-vehicular traffic sample image, said Haar characteristic and said gradient characteristic are carried out the pattern-recognition training, and Adaboost algorithm or SVM algorithm are adopted in said pattern-recognition training;
The target signature archival memory is connected with said sample image pattern-recognition training aids, stores said sample image pattern-recognition training aids and obtains vehicular traffic sample object property data base according to normalization type, angular type and type of vehicle;
(2) moving object detection equipment is connected with traffic video equipment, from the traffic video image sequence that obtains, detects motion target area;
(3) vehicular traffic identification equipment; Carry out data transmission with said moving object detection equipment; Load said vehicular traffic sample object property data base, the said motion target area that said moving object detection recognition of devices is gone out carries out vehicle identification, and said vehicular traffic identification equipment comprises
The vehicle search window is provided with a vehicle search window in said motion target area;
The search window feature deriving means is connected with said vehicle search window, obtains integration characteristic, Haar characteristic and the gradient characteristic of the gray level image in the said vehicle search window;
Comparative device; Be connected with said target signature archival memory with said search window feature deriving means; Said integration characteristic, said Haar characteristic and the said gradient characteristic of said search window feature deriving means acquisition and the said vehicular traffic sample object property data base in the said target signature archival memory are compared, the vehicle in the said vehicle search window is discerned.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2012201121029U CN202563526U (en) | 2012-03-22 | 2012-03-22 | Transportation vehicle detection and recognition system based on video |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2012201121029U CN202563526U (en) | 2012-03-22 | 2012-03-22 | Transportation vehicle detection and recognition system based on video |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN202563526U true CN202563526U (en) | 2012-11-28 |
Family
ID=47213156
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2012201121029U Expired - Fee Related CN202563526U (en) | 2012-03-22 | 2012-03-22 | Transportation vehicle detection and recognition system based on video |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN202563526U (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102637257A (en) * | 2012-03-22 | 2012-08-15 | 北京尚易德科技有限公司 | Video-based detection and recognition system and method of vehicles |
| CN103136935A (en) * | 2013-01-11 | 2013-06-05 | 东南大学 | Method for tracking sheltered vehicles |
| CN103400117A (en) * | 2013-07-29 | 2013-11-20 | 电子科技大学 | Method for positioning and tracking personnel in well on basis of compute vision |
| CN104680787A (en) * | 2015-02-04 | 2015-06-03 | 上海依图网络科技有限公司 | Method for detecting congestion condition of roads |
| CN104978584A (en) * | 2014-04-14 | 2015-10-14 | 航天信息股份有限公司 | Fumigation medicine detection method and device based on video monitoring |
| CN105138525A (en) * | 2014-05-30 | 2015-12-09 | 株式会社日立制作所 | Traffic video processing device and method, and retrieval device and method |
| CN105160333A (en) * | 2015-09-30 | 2015-12-16 | 深圳市华尊科技股份有限公司 | Vehicle model identifying method and vehicle model identifying device |
| CN107644528A (en) * | 2017-08-02 | 2018-01-30 | 浙江工业大学 | A kind of vehicle queue length detection method based on vehicle tracking |
| CN107730972A (en) * | 2017-05-10 | 2018-02-23 | 西安艾润物联网技术服务有限责任公司 | The method and apparatus that video identification controls banister |
| CN109815856A (en) * | 2019-01-08 | 2019-05-28 | 深圳中兴网信科技有限公司 | Status indication method, system and the computer readable storage medium of target vehicle |
| CN110751085A (en) * | 2019-10-17 | 2020-02-04 | 南京航空航天大学 | Mouse behavior recognition system |
| CN112465691A (en) * | 2020-11-25 | 2021-03-09 | 北京旷视科技有限公司 | Image processing method, image processing device, electronic equipment and computer readable medium |
| CN113538891A (en) * | 2020-04-17 | 2021-10-22 | 无锡锦铖人工智能科技有限公司 | Intelligent vehicle counting system |
-
2012
- 2012-03-22 CN CN2012201121029U patent/CN202563526U/en not_active Expired - Fee Related
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102637257B (en) * | 2012-03-22 | 2014-07-02 | 北京尚易德科技有限公司 | Video-based detection and recognition system and method of vehicles |
| CN102637257A (en) * | 2012-03-22 | 2012-08-15 | 北京尚易德科技有限公司 | Video-based detection and recognition system and method of vehicles |
| CN103136935A (en) * | 2013-01-11 | 2013-06-05 | 东南大学 | Method for tracking sheltered vehicles |
| CN103136935B (en) * | 2013-01-11 | 2015-04-15 | 东南大学 | Method for tracking sheltered vehicles |
| CN103400117A (en) * | 2013-07-29 | 2013-11-20 | 电子科技大学 | Method for positioning and tracking personnel in well on basis of compute vision |
| CN103400117B (en) * | 2013-07-29 | 2016-02-10 | 电子科技大学 | A kind of personnel in the pit's positioning and tracing method based on computer vision |
| CN104978584B (en) * | 2014-04-14 | 2019-01-29 | 航天信息股份有限公司 | Stifling drug detection method and device based on video monitoring |
| CN104978584A (en) * | 2014-04-14 | 2015-10-14 | 航天信息股份有限公司 | Fumigation medicine detection method and device based on video monitoring |
| CN105138525A (en) * | 2014-05-30 | 2015-12-09 | 株式会社日立制作所 | Traffic video processing device and method, and retrieval device and method |
| CN105138525B (en) * | 2014-05-30 | 2019-10-01 | 株式会社日立制作所 | Traffic video processing unit and method and retrieval device and method |
| CN104680787A (en) * | 2015-02-04 | 2015-06-03 | 上海依图网络科技有限公司 | Method for detecting congestion condition of roads |
| CN105160333A (en) * | 2015-09-30 | 2015-12-16 | 深圳市华尊科技股份有限公司 | Vehicle model identifying method and vehicle model identifying device |
| CN105160333B (en) * | 2015-09-30 | 2018-08-17 | 深圳市华尊科技股份有限公司 | A kind of model recognizing method and identification device |
| CN107730972A (en) * | 2017-05-10 | 2018-02-23 | 西安艾润物联网技术服务有限责任公司 | The method and apparatus that video identification controls banister |
| CN107644528A (en) * | 2017-08-02 | 2018-01-30 | 浙江工业大学 | A kind of vehicle queue length detection method based on vehicle tracking |
| CN109815856A (en) * | 2019-01-08 | 2019-05-28 | 深圳中兴网信科技有限公司 | Status indication method, system and the computer readable storage medium of target vehicle |
| CN110751085A (en) * | 2019-10-17 | 2020-02-04 | 南京航空航天大学 | Mouse behavior recognition system |
| CN113538891A (en) * | 2020-04-17 | 2021-10-22 | 无锡锦铖人工智能科技有限公司 | Intelligent vehicle counting system |
| CN112465691A (en) * | 2020-11-25 | 2021-03-09 | 北京旷视科技有限公司 | Image processing method, image processing device, electronic equipment and computer readable medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102637257B (en) | Video-based detection and recognition system and method of vehicles | |
| CN202563526U (en) | Transportation vehicle detection and recognition system based on video | |
| Wei et al. | Multi-vehicle detection algorithm through combining Harr and HOG features | |
| Jensen et al. | Vision for looking at traffic lights: Issues, survey, and perspectives | |
| Menon et al. | Detection and recognition of multiple license plate from still images | |
| CN105373794B (en) | A kind of licence plate recognition method | |
| CN104951784B (en) | A kind of vehicle is unlicensed and license plate shading real-time detection method | |
| CN102289686B (en) | Method for identifying classes of moving targets based on transfer learning | |
| CN102509091B (en) | Airplane tail number recognition method | |
| Alefs et al. | Road sign detection from edge orientation histograms | |
| Lin et al. | A license plate recognition system for severe tilt angles using mask R-CNN | |
| CN102867417B (en) | Taxi anti-forgery system and taxi anti-forgery method | |
| Romdhane et al. | An improved traffic signs recognition and tracking method for driver assistance system | |
| CN103870803A (en) | Vehicle license plate recognition method and system based on coarse positioning and fine positioning fusion | |
| CN102364496A (en) | Method and system for identifying automobile license plates automatically based on image analysis | |
| Momin et al. | Vehicle detection and attribute based search of vehicles in video surveillance system | |
| CN102902983B (en) | A kind of taxi identification method based on support vector machine | |
| CN104615986A (en) | Method for utilizing multiple detectors to conduct pedestrian detection on video images of scene change | |
| CN104978567A (en) | Vehicle detection method based on scenario classification | |
| CN102915433A (en) | Character combination-based license plate positioning and identifying method | |
| CN103310006A (en) | ROI extraction method in auxiliary vehicle driving system | |
| CN104463232A (en) | Density crowd counting method based on HOG characteristic and color histogram characteristic | |
| Chen et al. | Robust and real-time traffic light recognition based on hierarchical vision architecture | |
| Mammeri et al. | North-American speed limit sign detection and recognition for smart cars | |
| Ketcham et al. | Recognizing the Illegal Parking Patterns of Cars on the Road in Front of the Bus Stop Using the Support Vector Machine |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20121128 Termination date: 20200322 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |