CN105913454A - Pixel coordinate locus prediction method of motion object in video image - Google Patents

Abstract

The invention discloses a pixel coordinate locus prediction method of a motion object in a video image, comprising steps of obtaining motion object history pixel locus information from the video image, putting forward a fractional model fitting locus formula based on a camera image-forming principle, establishing a overdetermined equation set by combining with a history locus in order to solve coefficients of the fractional model in order to determine a relation between a pixel coordinate and time of a motion object which performs uniform linear motion in a real space in a video image, and accurately predicting a pixel coordinate of the object in a future moment. The pixel coordinate locus prediction method does not need to provide camera parameters or any information about a road surface in advance and can predict the image position of the future moment through the history pixel motion locus and corresponding time interval information of the object performing uniform linear motion in the real space in the video image, and is simple and practical in operation and high in accuracy.

Description

A kind of pixel coordinate trajectory predictions method of moving object in video sequences

Technical field:

The invention belongs to video signal processing technology field, specifically one and utilize moving target history pixel Trace information determines the Forecasting Methodology of its pixel coordinate and corresponding time relationship.

Background technology:

Along with economic development and the raising of living standards of the people, the motorization level of China improves rapidly, the thing followed Traffic congestion and accident also get more and more, and therefore supervise for road vehicle and bring huge pressure to traffic police, and intelligence Can traffic increasingly be subject to people's attention as a kind of technology that can effectively solve appeal problem.

Along with the resolution of Current traffic CCTV camera is more and more higher, monitoring range is the most increasing, is passing through traffic CCTV camera obtains in the real time video image on road, by image processing techniques detecting and tracking road vehicle target, If it is known that all driving traces of a certain uniform motion target in current time image, need to predict the figure of its future time Image position, it is impossible to use simple uniform motion pattern, because camera lens image-forming principle determines road in traffic video image Actual range be not the same with corresponding pixel distance.

The most accurate Forecasting Methodology is to need camera parameters, and antenna height, attitude and road surface distance etc. count According to the relation set up between pixel distance and actual range, then derive target by the uniform motion pattern in real world The relation of pixel motion distance and time, but for needs such for traffic cameras are the most commonly used, in reality In operation, the error of some data cannot obtain very greatly the most at all.

The present invention then provides a kind of convenient practicality, and can ensure the Forecasting Methodology of precision, to solve appeal problem.

Summary of the invention

Goal of the invention: according to camera lens image-forming principle, uniform motion object pixel coordinate in video image and time Between the uniform motion pattern that is no longer complies with in real world of relation, if to determine according to existing object pixel Grid Track Coordinate and the relation of time, use simple linear fit or fitting of a polynomial track, it was predicted that the result gone out often precision is not Enough, error is bigger；And if not according to existing track, directly setting up mathematical modulo by camera parameters and real road data Type, realizes extremely complex even cannot realizing in practical operation.For the problems referred to above, the present invention proposes a kind of convenient practicality, And the method that precision can be ensured, rely on uniform motion target history pixel motion track in video image and corresponding time Information, the picture position of its future time the most measurable.

Technical scheme: the present invention provides a kind of pixel coordinate trajectory predictions method of moving object in video sequences, the party Method, based on the uniform motion object pixel Grid Track in fractional model matching video image, comprises the steps:

Step 1: in camera supervised scope, by image processing techniques, obtains moving target based on pixel coordinate system The historical track information of some frame of video (each frame of video represents a sampling instant, the most some sampling instants).

Step 2: choose n frame (the i.e. n that the pixel speed trend of target is correct from the frame of video track acquired in step 1 The individual moment), its object pixel coordinate is: (p_x1,p_y1),(p_x2,p_y2)…(p_xn,p_yn), the most corresponding sampling instant t₁,t₂…t_n, Wherein sampling instant initial value t₁=0, n are the integer not less than 3, because x or y direction model has three unknown numbers, need Could set up solving equations more than or equal to 3, in general, the historical track information according to having obtained chooses correct tracing point Number, the correct tracing point certainly chosen is the most, determined by forecast model more accurate.

Step 3: x coordinate of each point selected in step 2 and y-coordinate are set up respectively a pixel coordinate and time Between Fraction Functions relation:Wherein a_x,b_x,c_x,a_y,b_y,c_yFor unknown system to be determined Number, obtains about x with about the over-determined systems of y:

$\left\{\begin{array}{c}{t}_1{p}_{x1}{a}_x+p_{x1}{b}_x-c_x=t_1\\ .\\ .\\ .\\ t_n{p}_{xn}{a}_x+p_{xn}{b}_x-c_x=t_n\end{array}\right.---\left(1\right)$

$\left\{\begin{array}{c}{t}_1{p}_{y1}{a}_y+p_{y1}{b}_y-c_y=t_1\\ .\\ .\\ .\\ t_n{p}_{yn}{a}_y+p_{yn}{b}_y-c_y=t_n\end{array}\right.---\left(2\right)$

Owing to unknown coefficient to be determined has contained and camera suspension height, angle, focal length, target speed and just The information that the factors such as beginning position are relevant, this Forecasting Methodology need not any information of any information on actual road surface and camera, Equation can be set up ask by uniform motion target history pixel motion track in video image and corresponding temporal information Solve these unknowm coefficients, determine its pixel coordinate track and the movement relation formula between the time.

Step 4: two over-determined systems obtained in solution procedure 3, obtains one group of solution: a_x',b_x',c_x',a_y',b_y', c_y', so that it is determined that x, y-coordinate track and the relation between the time:

$p_x=\frac{t+{c_x}^{\′}}{{a_x}^{\′}t+{b_x}^{\′}}---\left(5\right)$

$p_y=\frac{t+{c_y}^{\′}}{{a_y}^{\′}t+{b_y}^{\′}}---\left(6\right)$

Step 5: the moment to be predicted substitutes into the relational expression of pixel coordinate and time, to predict that this moving target will be when future The pixel coordinate carved, the pixel coordinate such as t isWhereinRepresent to Under round.

Step 1 obtains the concrete grammar of the historical track information of moving target some frame of video based on pixel coordinate system For:

After obtaining real-time video flow image data by video flowing frame decoding, first carry out Image semantic classification and remove noise, Then realize the detection of moving target with background subtraction (wherein, to process can obtain through image binaryzation and morphological image To the elementary contour of detected moving target, use the motion mesh that boundary rectangle method labelling based on objective contour is obtained Mark), follow the tracks of the same moving target between upper frame of video finally by the method that boundary rectangle overlapping area is maximum, thus extract The moving parameter information such as the history driving trace going out each moving target.

History pixel track selected in step 2 is counted n=6.

Step 2 is chosen from the existing m frame track information of target the concrete grammar of n satisfactory object pixel coordinate For:

1) uniformly choose wherein n frame, be spaced between the most each selected frameFrame, it is judged that between each adjacent two frames of selected frame The mean pixel velocity variations trend of target whether meet the requirements, between each adjacent two frames of i.e. selected frame, object pixel y sits Mark certain trend that whether meets of average speed: when target makees the uniform motion away from camera lens, each adjacent the two of selected frame Between frame, object pixel y-coordinate average speed should be more and more less；Otherwise, object pixel y between each adjacent two frames of selected frame Coordinate average speed should be increasing, calculates object pixel y-coordinate average speed formula between each adjacent two frames of selected frame ForWherein i, j represent the most different two frames；

2) if the mean pixel velocity variations trend of the target between certain adjacent two frame selected is incorrect, find the most successively The most adjacent frame with correct trend is to replace this frame, and all frames after this frame of gravity treatment, until selected frame is each The mean pixel velocity variations trend of the target between adjacent two frames all meets the requirements.

Step 4 solves over-determined systems by the method solving unique least square solution, method particularly includes:

Over-determined systems about x is written as form:

$\left(\begin{array}{ccc}{t}_1{p}_{x1}& p_{x1}& -1\\ .& .& .\\ .& .& .\\ .& .& .\\ t_n{p}_{xn}& p_{xn}& -1\end{array}\right)\left(\begin{array}{c}{a}_x\\ b_x\\ c_x\end{array}\right)=\left(\begin{array}{c}{t}_1\\ .\\ .\\ .\\ t_n\end{array}\right)---\left(3\right)$

IfThen formula 3 is expressed as:

A=(R^H·R)^-1·R^H·T (4)

Solution formula 4, obtains the relational expression of object pixel coordinate system x direction coordinate and time:

$p_x=\frac{t+{c_x}^{\′}}{{a_x}^{\′}t+{b_x}^{\′}}---\left(5\right)$

In like manner, the relational expression of object pixel coordinate system y direction coordinate and time is obtained:

$p_y=\frac{t+{c_y}^{\′}}{{a_y}^{\′}t+{b_y}^{\′}}---\left(6\right)$

Beneficial effect: the invention provides the picture of uniform motion target in a kind of video image based on target histories track Element coordinate and the Forecasting Methodology of time relationship, get rid of detection effect of noise by choosing correct trend point, more practical；Relatively In general linear fit and fitting of a polynomial, fraction approximating method precision of prediction improves a lot；Relative to directly by Know that intrinsic parameters of the camera, external position morphological parameters and real road data founding mathematical models solve pixel coordinate and time The method of relation, operation is greatly simplified, it is not necessary to field test data, pervasive and can ensure precision.

Accompanying drawing explanation

Fig. 1 is the image processing flow figure of the embodiment of the present invention；

Fig. 2 is the prediction matched curve error schematic diagram of the embodiment of the present invention；

Fig. 3 is the object pixel Grid Track prediction algorithm flow chart of the embodiment of the present invention.

Detailed description of the invention

Below in conjunction with specific embodiment, it is further elucidated with the present invention, it should be understood that these embodiments are merely to illustrate the present invention Rather than restriction the scope of the present invention, after having read the present invention, the those skilled in the art's various equivalences to the present invention The amendment of form all falls within the application claims limited range.

A kind of pixel coordinate trajectory predictions method of moving target being applied in traffic video monitoring, including walking as follows Rapid:

Step 1: in camera supervised scope, by image processing techniques, obtains moving target based on pixel coordinate system Historical track information, as it is shown in figure 1, its method particularly includes:

After obtaining real-time video flow image data by video flowing frame decoding, first carry out Image semantic classification and remove noise, Then realize the detection of moving target with background subtraction (wherein to process can obtain through image binaryzation and morphological image The elementary contour of detected moving target, uses the moving target that boundary rectangle method labelling based on objective contour is obtained), Follow the tracks of the same moving target between upper frame of video finally by the method that boundary rectangle overlapping area is maximum, thus extract every The moving parameter information such as the history driving trace of individual moving target.

Step 2: the object pixel coordinate that the speed trend in n moment in track acquired in selecting step 1 is correct: (p_x1, p_y1),(p_x2,p_y2)…(p_xn,p_yn), the most corresponding time t₁,t₂…t_n, wherein n=6, time initial value t₁=0；

Step 3: the x coordinate in each moment selected in step 2 and y-coordinate are set up respectively a pixel coordinate with The Fraction Functions relation of time:Wherein a_x,b_x,c_x,a_y,b_y,c_yTo be determined for the unknown Coefficient, available over-determined systems:

$\left\{\begin{array}{c}{t}_1{p}_{x1}{a}_x+p_{x1}{b}_x-c_x=t_1\\ .\\ .\\ .\\ t_n{p}_{xn}{a}_x+p_{xn}{b}_x-c_x=t_n\end{array}\right.,$

$\left\{\begin{array}{c}{t}_1{p}_{y1}{a}_y+p_{y1}{b}_y-c_y=t_1\\ .\\ .\\ .\\ t_n{p}_{yn}{a}_y+p_{yn}{b}_y-c_y=t_n\end{array}\right.;$

Step 4: two over-determined systems obtained in solution procedure 3, can solve one group of solution: obtain in solution procedure 3 Two over-determined systems, one group of solution: a can be obtained_x',b_x',c_x',a_y',b_y',c_y', so that it is determined that x, y-coordinate track with Relation between time:

$p_x=\frac{t+{c_x}^{\′}}{{a_x}^{\′}t+{b_x}^{\′}},$

$p_y=\frac{t+{c_y}^{\′}}{{a_y}^{\′}t+{b_y}^{\′}};$

Step 5: if now the pixel coordinate of moving target and the relation of time are just it was determined that need to predict this motion Target is at the pixel coordinate in certain moment following, it is only necessary to is substituted in the moment, such as, predicts the pixel coordinate of t, be thenWhereinRepresent and round downwards.

Specific in the present embodiment, the pixel coordinate trajectory predictions method of moving object in video sequences includes walking as follows Rapid:

Step 1: by image processing flow figure as shown in Figure 1, in acquisition monitor video, a certain moving target is based on picture Element coordinate system historical track information, 85 frame track information altogether, this target by camera lens far-end towards camera lens proximal movement, institute Should be as the time with speed trend more and more less, the 29th frame information is pixel coordinate during this target of initial discovery；

Step 2: accurately choose 6 satisfactory object pixel coordinates from the existing trace information of target, between first waiting Every choosing 6 frames, it is spaced 14 frames between the most each pixel coordinate, thus obtains the pixel coordinate dot information in wherein 6 frame moment such as Shown in table 1, it is judged that its speed trend is the most correct, finds that last poimt-to-point speed trend is incorrect, gravity treatment the 58th frame data.

Table 1: reconnaissance trace information at equal intervals

Frame sequence	114	100	86	72	58	44
							px	260	298	329	366	392	416
py	124	89	68	52	44	30

Owing to the most front four frame data have been correct, from the 5th start to reselect near meet speed trend Point, regains 6 dot informations as shown in table 2；

Table 2: correct locus of points information

Frame sequence	114	100	86	72	57	52
							px	260	298	329	366	391	401
py	124	89	68	52	41	39

Step 3: substitute into fractional model, set up following over-determined systems:

$\left\{\begin{array}{c}401\×52\×a_x+401\×b_x-c_x=52\\ 391\×57\×a_x+391\×b_x-c_x=57\\ 366\×72\×a_x+366\×b_x-c_x=72\\ 329\×86\×a_x+329\×b_x-c_x=86\\ 298\×100\×a_x+298\×b_x-c_x=100\\ 260\×114\×a_x+260\×b_x-c_x=114\end{array}\right.,$

$\left\{\begin{array}{c}39\×52\×a_y+39\×b_y-c_y=52\\ 41\×57\×a_y+41\×b_y-c_y=57\\ 52\×72\×a_y+52\×b_y-c_y=72\\ 68\×86\×a_y+68\×b_y-c_y=86\\ 89\×100\×a_y+89\×b_y-c_y=100\\ 124\×114\×a_y+124\×b_y-c_y=114\end{array}\right.$

Step 4: solved by least square solution: a_x=0.0012, b_x=-0.3662, c_x=-174.7588, a_y=- 0.0390, b_y=6.2578, c_y=109.8697；

Step 5: the pixel coordinate of moving target and the relation of time are just it was determined that fractional model prediction curve such as Fig. 2 Shown in, if now needing the pixel coordinate predicting this moving target at certain frame following, it is only necessary to frame sequence is substituted into, such as Predict the pixel coordinate of t frame, be thenWhereinTable Show and round downwards.

By this method, the trace information of the postorder frame of this moving target obtains the most, actual path information and the present invention The curve of error of forecast model is as in figure 2 it is shown, matched curve is the most identical with former track, respond well.

Claims (5)

1. the pixel coordinate trajectory predictions method of a moving object in video sequences, it is characterised in that: the method is based on fraction Uniform motion object pixel Grid Track in models fitting video image, comprises the steps:

Step 1: in camera supervised scope, by image processing techniques, if it is based on pixel coordinate system to obtain moving target The historical track information of dry frame of video；

Step 2: choose the n frame that the pixel speed trend of target is correct, its mesh from the some frame of video tracks acquired in step 1 Mark pixel coordinate is: (p_x1,p_y1),(p_x2,p_y2)…(p_xn,p_yn), the most corresponding sampling instant t₁,t₂…t_n, wherein sampling instant Initial value t₁=0, n are the integer not less than 3；

Step 3: the x and y coordinates of this target in each frame selected in step 2 is set up a pixel coordinate and time respectively Fraction Functions relation:Wherein a_x,b_x,c_x,a_y,b_y,c_yFor unknown coefficient to be determined, obtain To about x with about the over-determined systems of y:

Step 4: two over-determined systems obtained in solution procedure 3, obtains one group of solution: a_x',b_x',c_x',a_y',b_y',c_y', from And determine x, y-coordinate track and the relation between the time:

Step 5: is substituted into the relational expression of pixel coordinate and time, to predict that this moving target is at future time instance the moment to be predicted Pixel coordinate, the pixel coordinate such as t isWhereinRepresent and take downwards Whole.

The pixel coordinate trajectory predictions method of moving object in video sequences the most according to claim 1, it is characterised in that: Step 1 obtains the historical track information of moving target some frame of video based on pixel coordinate system method particularly includes:

After obtaining real-time video flow image data by video flowing frame decoding, first carry out Image semantic classification and remove noise, then Realize the detection of moving target with background subtraction, wherein process through image binaryzation and morphological image and obtain being detected fortune The elementary contour of moving-target, uses the moving target that boundary rectangle method labelling based on objective contour is obtained, finally by Same moving target between the upper frame of video of method tracking that boundary rectangle overlapping area is maximum, thus extract each motion mesh The moving parameter information such as target history driving trace.

The pixel coordinate trajectory predictions method of moving object in video sequences the most according to claim 1, it is characterised in that: History pixel track selected in step 2 is counted n=6.

The pixel coordinate trajectory predictions method of moving object in video sequences the most according to claim 1, it is characterised in that: Step 2 chooses n satisfactory object pixel coordinate from the existing m frame track information of target method particularly includes:

1) uniformly choose wherein n frame, be spaced between the most each selected frameFrame, it is judged that the mesh between each adjacent two frames of selected frame Whether target mean pixel velocity variations trend meets the requirements, and between each adjacent two frames of i.e. selected frame, object pixel y-coordinate is put down All speed whether meet certain trend: when target makees the uniform motion away from camera lens, each adjacent two frames of selected frame it Between object pixel y-coordinate average speed should be more and more less；Otherwise, object pixel y-coordinate between each adjacent two frames of selected frame Average speed should be increasing, calculates object pixel y-coordinate average speed formula between each adjacent two frames of selected frame and isWherein i, j represent the most different two frames；

2) if the mean pixel velocity variations trend of the target between certain adjacent two frame selected is incorrect, find the most successively around The most adjacent frame with correct trend is to replace this frame, and all frames after this frame of gravity treatment, until selected frame is each adjacent The mean pixel velocity variations trend of the target between two frames all meets the requirements.

The pixel coordinate trajectory predictions method of moving object in video sequences the most according to claim 1, it is characterised in that: Step 4 solves over-determined systems by the method solving unique least square solution, method particularly includes:

Over-determined systems about x is written as form:

IfThen formula 1 is expressed as:

A=(R^H·R)^-1·R^H·T (4)

Solution formula 4, obtains the relational expression of object pixel coordinate system x direction coordinate and time:

In like manner, the relational expression of object pixel coordinate system y direction coordinate and time is obtained:

。