CN103914853A - Method for processing target adhesion and splitting conditions in multi-vehicle tracking process - Google Patents

Abstract

The invention discloses a method for processing target adhesion and splitting in the multi-vehicle tracking process. The method includes the first step of carrying out processing in a vehicle tracking algorithm based on a maximum overlapping area and in a vehicle tracking algorithm based on Mean shift, and the second step of processing the target adhesion and splitting conditions in the vehicle tracking process, wherein the first step includes the steps of carrying out normal vehicle tracking according to the vehicle tracking algorithm based on the maximum overlapping area and tracking adhering and split vehicles according to the vehicle tracking algorithm based on the Mean shift, and the second step includes the steps of analyzing matching conditions of a moving object in previous frames and the moving object in latter frames to judge the conditions of the moving object and carrying out tracking through different algorithms according to the feedback conditions of the moving object. According to the method, on the basis that the detection sequence of the vehicles is given, the non-hundred-percent reliability of the detection sequence is considered, the previous frames where the object appears are analyzed, the adhesing vehicles and the normal vehicles are distinguished through marking, and consequently the vehicles are tracked in two ways. The method is simple, easy to realize and less in time consumption.

Description

The disposal route of target adhesion and division situation when many vehicle trackings

Technical field

The present invention relates to the technical field of vehicle tracking, particularly the disposal route of target adhesion and division situation when a kind of many vehicle trackings of reality.

Background technology

Along with more and more robotization of traffic control system, intellectuality, Intelligent traffic management systems also becomes hot research object.Vehicle tracking is vital module in Intelligent traffic management systems, and it has merged the knowledge of numerous ambits such as computer vision, Computer Image Processing, machine learning.

Desirable, traffic is simple, and in the perfect situation of testing result, current track algorithm, as Mean shift, Camshift, template tracking, particle filter and feature point tracking etc. all can solve tracking problem, but reality is really not so, under complicated traffic conditions, vehicle there will be adhesion, and in the faulty situation of testing result, vehicle there will be division.

Though front people have proposed a lot of solutions to these problems, major part has all only been considered the solution to adhesion problems, does not consider due to the incorrect fragmentation problem causing due to testing result in tracing process; And do not consider the situation that many cars stick together, and disposal route while only just considering the situation that two cars stick together; Finally, do not emphasize the importance that is judged as correct target or false target when vehicle appears at guarded region, because if just there is adhesion when vehicle has just appeared at guarded region, and be all generally to give tacit consent on the first appearance for normal vehicle according to front people's method, so, if no longer adhesion between vehicle in follow-up tracing process, front people's method will think that vehicle divides, thereby process, so just occurred wrong tracking.In like manner, if testing result division when vehicle has just appeared at guarded region, the tracking that also can lead to errors according to front people's method.

Summary of the invention

The object of the invention is to overcome the shortcoming and deficiency of prior art, the disposal route of target adhesion and division situation when a kind of many vehicle trackings of reality are provided.

Object of the present invention is achieved through the following technical solutions:

When many vehicle trackings, a disposal route for target adhesion and division situation, comprises the steps:

S1, vehicle tracking algorithm based on Maximum overlap area and Mean shift;

S11, the vehicle tracking algorithm of utilization based on Maximum overlap area carry out normal vehicle tracking;

S12, the tracking that utilizes the vehicle tracking algorithm based on Mean shift to carry out adhesion and divide vehicle;

Processing to target adhesion and division situation in S2, vehicle tracking process;

The match condition of S21, moving target to front and back frame is analyzed, and judges the state of moving target;

S22, utilize different algorithms to follow the tracks of according to the situation of the feedback of status of moving target.

Preferably, in step S11, the vehicle tracking algorithm of Maximum overlap area is:

Upper left corner coordinate (the x of vehicle in known k frame ₁, y ₁) and lower right corner coordinate (x ₂, y ₂), the upper left corner coordinate (x of vehicle in k+1 frame ₃, y ₃) and lower right corner coordinate (x ₄, y ₄), upper left corner coordinate and the left side, the lower right corner of establishing overlapping region are respectively: (x ₅, y ₅), (x ₆, y ₆):

$\left\{\begin{array}{c}{x}_5=\max (x_1,x_3)\\ x_6=\min (x_2,x_4)\\ y_5=\max (y_1,y_3)\\ y_6=\min (y_2,y_4)\end{array}\right.---\left(1\right)$

Can be tried to achieve respectively the upper left corner and the lower right corner of overlapping region by formula (1);

$\left\{\begin{array}{c}{x}_5<x_6\\ y_5<y_6\end{array}\right.---\left(2\right)$

If (2) formula is set up, illustrate that there is lap two rectangular areas, otherwise two rectangular areas are not overlapping, if (2) formula is set up, two rectangular area overlapping area S are

S=(x ₆-x ₅)*(y ₆-y ₅) （3）

Adjacent two frames detect in sequences, and the overlapping the maximum of area is the correct person of coupling, give target that this coupling is correct to number accordingly.

Preferably, step S12 is specially:

S121, manual search target window, set up object module

$q_u\left(\hat{y}\right)=\hat{C}{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{\hat{y}-{\hat{x}}_i}{\hat{h}}\left|\right|}^2\right)\mathrm{\&delta;}[b\left({\hat{x}}_i\right)-u]---\left(4\right)$

In formula, $\hat{C}=\frac{1}{{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{\hat{y}-{\hat{x}}_i}{\hat{h}}\left|\right|}^2\right)}$

$\mathrm{\&delta;}[b\left({\hat{x}}_i\right)-u]=\left\{\begin{array}{cc}1& b\left({\hat{x}}_i\right)=u\\ 0& b\left({\hat{x}}_i\right)\&NotEqual;u\end{array}\right.$

for the centre coordinate of object module, be the coordinate of i pixel, for the pixel value of point, the radius that h is To Template;

refer to the distance of i pixel to object module center;

for kernel function, when the larger k of x value (x) value less, otherwise when the less k of x value (x) is worth greatlyr, the weight of giving the closer to central point is larger;

for discrete impulse function, when time, equal 1; When time, equal 0, in the time that the pixel value traversing equals given pixel value, just numeration is once;

S122, set up candidate family at next frame

$C=\frac{1}{{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)}$

$\mathrm{\&delta;}[b\left(x_i\right)-u]=\left\{\begin{array}{cc}1& b\left(x_i\right)=u\\ 0& b\left(x_i\right)\&NotEqual;u\end{array}\right.$

In above formula, y ₀it is the centre coordinate of the To Template of previous frame;

S123, using Bhattacharyya coefficient as similarity function, obtain the iterative Bhattacharyya Coefficient Definition of Mean shift and be:

In the time that ρ (y) is maximum, illustrate that p (y) and q have maximum similarity;

Its first order Taylor is:

In formula,

(4), in formula, Section 1 and y are irrelevant, only need Section 2 to obtain maximal value so will try to achieve the value of ρ (y); Can obtain maximal value to Section 2 differentiate in (4)

Make f ' (y)=0, have:

So just, obtain Mean shift iterative:

Wherein, g (X)=-k ' is (X)

Because kernel function is given large weights for the pixel value of target's center, and give little weights for the pixel at wide center, thereby adopt kernel function histogram model modeling to guarantee that Mean shift algorithm itself has good robustness to blocking with the variation of background.

Preferably, in step S21, the state of moving target is normal condition, splitting status and adhesion state, and step S21 is specially:

S211, front and back two frame vehicles are to illustrate that one to one the state of this moving target is normal condition;

S212, before determining, in the frame target situation that is normal target, in lower frame, there is area rapid drawdown, and non-enter monitoring cable place judge that dividing appears in target while for no reason there is fresh target;

S213, before determining, frame target be normal target in the situation that, occurs that in lower frame area increases suddenly, and non-go out monitoring cable place judge that adhesion appears in target while having target to suddenly disappear.

Preferably, step S22 is specially:

S221, one of drafting enter the line of guarded region, think that when target enters this line this is fresh target, judge whether this target meets vehicle feature;

S222, the target detecting in sequence in adjacent two frames is mated with Maximum overlap area algorithm, in the time there is following three kinds of situations, think that the match is successful: for without counting and tagger is mated one by one; As long as just think that for what have tagger to have to match the match is successful; As long as just think it the match is successful for what have numeration person to have to match;

S223, draw a line that rolls guarded region away from, roll this line away from when target and think that target rolls away from, delete the storage organization of this car.

Preferably, step S221 further comprises the steps:

In the time that this target meets vehicle feature, give this fresh target with numbering; If be less than the scope of minimum vehicle, judge that it is not a complete car, now do not award numbering, but count 0, wait for several frames after more clear testing result accurately numbering; If this target is greater than the scope of minimum vehicle, so think that this target is that several cars have occurred that adhesion phenomenon has caused wrong testing result, now gives this target with numbering, and give this target with mark, to represent that this target is as undesired target.

Preferably, step S222 further comprises the steps:

If in coupling time, is found not numbered by matcher, but there is numeration number, it is described vehicle in as fresh target is imperfect to stay, now, if having, it is matched, for its numeration adds 1 operation;

In the time of coupling, find there is mark by matcher, whether the total area that contrasts target in its two frame varies widely, if had, mark is removed, the target matching is with numbering, toply give original numbering, remaining give successively new numbering, if area does not have that great changes will take place, award corresponding numbering and mark;

In the time of coupling, target is one to one, directly give its corresponding numbering, in the time that coupling is unsuccessful, if there is the multiple object matchings that detect sequence in k frame to the same target that detects sequence in k+1 frame, now show in k frame it is correct result, but during to k+1 frame, there is the situation that several cars merge, now with Mean shift algorithm, those the several targets in k frame have been departed from the tracking of k+1 frame detection sequence;

If there is an object matching of sequence in k frame to the multiple targets that detect sequence in k+1 frame, correct result while now showing in k frame, but to k+1 frame there is the situation of a car division, now with Mean shift algorithm, that target in k frame is departed from the tracking of k+1 frame detection sequence.

The present invention has following advantage and effect with respect to prior art:

1, be very simply easy to realize based on Maximum overlap area algorithm, greatly improved the speed of algorithm,

Can reach the requirement of real-time follow-up.

2, considered that vehicle feature that whether feature that vehicle has just entered guarded region meets vehicle judges the correctness of testing result when target occurs first, this is just for the tracking to adhesion vehicle afterwards lays the foundation.

3, be mainly to follow not reach threshold value that continuous n frame all occurs not to the principle of numbering when target for the processing of division.

4, for adhesion, and the Mean shift algorithm that vehicle fragmentation problem utilization in tracing process departs from testing result is followed the tracks of.

Accompanying drawing explanation

Fig. 1 is target following and to division adhesion situation Processing Algorithm process flow diagram;

Fig. 2 is that Maximum overlap area is asked method schematic diagram.

Embodiment

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited to this.

Embodiment

The disposal route of target adhesion and division situation while the invention discloses a kind of many vehicle trackings, as shown in Figure 1, step is:

(1), the vehicle tracking algorithm based on Maximum overlap area and Mean shift;

1., utilize the vehicle tracking algorithm based on Maximum overlap area to carry out normal vehicle tracking;

2., utilize the vehicle tracking algorithm based on Mean shift to carry out the tracking of adhesion and division vehicle;

(2), the processing to target adhesion and division situation in vehicle tracking process;

The match condition of the moving target 3., to front and back frame is analyzed, and judges the state of moving target;

4., utilize different algorithms to follow the tracks of according to the situation of the feedback of status of moving target;

In concrete enforcement, video camera is fixed the monitor video under scene, obtained the testing result sequence of each frame by moving object detection module, according to testing result sequence, utilize the combinational algorithm of Maximum overlap area algorithm and Mean Shift algorithm to follow the tracks of, and the state that may occur in tracing process moving target is analyzed, thereby correct solution division and adhesion situation reach and follow the tracks of correct object.The running environment of the disposal route of target adhesion and division situation when many vehicle trackings of the present invention: Intel (R) Core (TM) i7-3770K CPU@3.5GHz8GB RAM, developing instrument is VS2010, development language is C++ & OpenCV, video resolution is: 448 × 336, processing speed to video has reached 38fps, met the requirement of real-time processing, accuracy has reached 97.30%.

To describe concrete methods of realizing in detail according to step below:

1, the vehicle tracking algorithm based on Maximum overlap area and Mean shift;

Maximum overlap area algorithm is a kind of New Algorithm that the present invention proposes, this algorithm is mainly the speed in order to improve tracking, to reach the requirement of real-time follow-up, Maximum overlap area track algorithm is mainly to carry out the matching relationship between corresponding target in consecutive frame according to the sequence of vehicle detection result, this algorithm is simply easy to realize, consuming time little, but it too relies on testing result, if mistake appears in testing result, can affect tracking results, so Maximum overlap area algorithm and Mean shift algorithm are carried out to combination, because Mean shift algorithm is self-existent a kind of track algorithm, it does not rely on detection sequence, thereby thereby can assist Maximum overlap area track algorithm to improve correctness and the robustness of following the tracks of.

In the time two frames detection sequences being mated with Maximum overlap area algorithm, according to the rule of coupling and the threshold value of setting, when occurring that a car in present frame matches many cars in former frame, now illustrate that merging and blocking has appearred in vehicle; When many cars in present frame all match same car in former frame, the situation that has occurred division in testing result is now described.In the time running into above two kinds of situations, return to previous frame, the vehicle Mean shift algorithm that occurs this situation is followed the tracks of.

Above algorithm is described as follows:

1.1 based on Maximum overlap area algorithm, as shown in Figure 2:

1) ask the area of overlapping region

illustrate:

This step is mainly to provide method how to ask overlapping area.

algorithm:

Upper left corner coordinate (the x of vehicle in known k frame ₁, y ₁) and lower right corner coordinate (x ₂, y ₂), the upper left corner coordinate (x of vehicle in k+1 frame ₃, y ₃) and lower right corner coordinate (x ₄, y ₄).If the top-left coordinates of overlapping region and bottom right coordinate are respectively (x ₅, y ₅), (x ₆, y ₆):

$\left\{\begin{array}{c}{x}_5=\max (x_1,x_3)\\ x_6=\min (x_2,x_4)\\ y_5=\max (y_1,y_3)\\ y_6=\min (y_2,y_4)\end{array}\right.---\left(1\right)$

Can be tried to achieve respectively the upper left corner and the lower right corner of overlapping region by formula (1).

$\left\{\begin{array}{c}{x}_5<x_6\\ y_5<y_6\end{array}\right.---\left(2\right)$

If (2) formula is set up, illustrate that there is lap two rectangular areas, otherwise two rectangular areas are not overlapping.If (2) formula is set up, two rectangular area overlapping area S are:

S=(x ₆-x ₅)*(y ₆-y ₅) （3）

2) matching principle

illustrate:

A given target, finds out the target that overlaps area maximum in next frame by the method for exhaustion.

algorithm:

Adjacent two frames detect in sequences, and the overlapping the maximum of area is the correct person of coupling, give target that this coupling is correct to number accordingly.

M=max{s ₁,s ₂,s ₃,…,s _n}

1.2 based on Mean shift algorithm

1) manual search target window, sets up object module

illustrate:

Given target window, utilizes COLOR COMPOSITION THROUGH DISTRIBUTION probability to set up object module.

algorithm:

$q_u\left(\hat{y}\right)=\hat{C}{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{\hat{y}-{\hat{x}}_i}{\hat{h}}\left|\right|}^2\right)\mathrm{\&delta;}[b\left({\hat{x}}_i\right)-u]---\left(4\right)$

In formula, $\hat{C}=\frac{1}{{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{\hat{y}-{\hat{x}}_i}{\hat{h}}\left|\right|}^2\right)}$

$\mathrm{\&delta;}[b\left({\hat{x}}_i\right)-u]=\left\{\begin{array}{cc}1& b\left({\hat{x}}_i\right)=u\\ 0& b\left({\hat{x}}_i\right)\&NotEqual;u\end{array}\right.$

for the centre coordinate of object module, be the coordinate of i pixel, for the pixel value of point, the radius that h is To Template.

refer to the distance of i pixel to object module center.

for kernel function, when the larger k of x value (x) value less, otherwise when the less k of x value (x) is worth greatlyr, the weight of giving the closer to central point is larger.

for discrete impulse function, when time, equal 1; When time, equal 0.In the time that the pixel value traversing equals given pixel value, just numeration is once.

2) set up candidate family at next frame

illustrate:

Same 1), set up the color probability Distribution Model of next frame.

algorithm:

$C=\frac{1}{{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)}$

$\mathrm{\&delta;}[b\left(x_i\right)-u]=\left\{\begin{array}{cc}1& b\left(x_i\right)=u\\ 0& b\left(x_i\right)\&NotEqual;u\end{array}\right.$

In above formula, y ₀it is the centre coordinate of the To Template of previous frame.

3), using Bhattacharyya coefficient as similarity function, obtain Mean shift iterative

illustrate:

Using Bhattacharyya coefficient as similarity function, make candidate family approach step by step object module, obtain Mean shift iterative.

algorithm:

Bhattacharyya Coefficient Definition is:

In the time that ρ (y) is maximum, illustrate that p (y) and q have maximum similarity;

Its first order Taylor is:

In formula,

(4), in formula, Section 1 and y are irrelevant, only need Section 2 to obtain maximal value so will try to achieve the value of ρ (y); Can obtain maximal value to Section 2 differentiate in (4)

Make f ' (y)=0, have:

So just, obtain Mean shift iterative:

Wherein, g (X)=-k ' is (X)

Because kernel function is given large weights for the pixel value of target's center, and give little weights for the pixel at wide center, thereby adopt kernel function histogram model modeling to guarantee that Mean shift algorithm itself has good robustness to blocking with the variation of background.

2, the processing to target adhesion and division situation in vehicle tracking process;

Due to illumination, the impact of the factors such as weather, present detection technique does not reach perfect effect, and there is division in the vehicle occasional that detects, and what sometimes many cars leaned on too closely may detect them to become a car.So, but in tracing process, there will be some unreasonable situations that really can exist, must process and just can make tracking results rationally with correct these situations.

2.1 fresh targets occur

illustrate:

Judge which kind of situation is that fresh target occurs.

algorithm:

Manually draw a line that enters guarded region, think that when target enters this line this is fresh target, judge whether this target meets vehicle feature (can provide in advance enter the length of line part different automobile types and the scope of width).In the time that this target meets vehicle feature, give this fresh target with numbering; If be less than the scope of minimum vehicle, judge that it is not a complete car, now do not award numbering, but count 0, wait for several frames after more clear testing result accurately numbering; If this target is greater than the scope of minimum vehicle, so think that this target is that several cars have occurred that adhesion phenomenon has caused wrong testing result, now gives this target with numbering, and give this target with mark, to represent that this target is as undesired target.

In 2.2 monitoring ranges, follow the tracks of situation analysis

illustrate:

According to the situation of coupling, as one to one, one-to-many, many-one, provides disposal route.

algorithm:

The target detecting in sequence in adjacent two frames is mated with Maximum overlap area algorithm, in the time there is following three kinds of situations, think that the match is successful: for without counting and tagger is mated one by one; As long as just think that for what have tagger to have to match the match is successful; As long as just think it the match is successful for what have numeration person to have to match.

If in coupling time, is found not numbered by matcher, but there is numeration number, it is described vehicle in as fresh target is imperfect to stay, now, if having, it is matched, for its numeration adds 1 operation.In the time that reaching 5, its numeration gives its numbering.

Find there is mark by matcher when when coupling, whether the total area that contrasts target in its two frame varies widely, if had, mark is removed, and the target matching is with numbering, toply gives original numbering, remaining gives successively new numbering.If area does not have that great changes will take place, award corresponding numbering and mark.

In the time of coupling, target is one to one, directly gives its corresponding numbering.

In the time that coupling is unsuccessful, if there is the multiple object matchings that detect sequence in k frame to the same target that detects sequence in k+1 frame, now show in k frame it is correct result, but during to k+1 frame, there is the situation that several cars merge, now with Mean shift algorithm, those the several targets in k frame have been departed from the tracking of k+1 frame detection sequence.

If there is an object matching of sequence in k frame to the multiple targets that detect sequence in k+1 frame, correct result while now showing in k frame, but to k+1 frame there is the situation of a car division, now with Mean shift algorithm, that target in k frame is departed from the tracking of k+1 frame detection sequence.

2.3 targets are rolled away from

illustrate:

Judge which kind of situation is that target is rolled away from.

algorithm:

Manually draw a line that rolls guarded region away from, roll this line away from when target and think that target rolls away from, delete the storage organization of this car.

Above-described embodiment is preferably embodiment of the present invention; but embodiments of the present invention are not restricted to the described embodiments; other any do not deviate from change, the modification done under Spirit Essence of the present invention and principle, substitutes, combination, simplify; all should be equivalent substitute mode, within being included in protection scope of the present invention.

Claims (7)

1. when vehicle tracking more than, a disposal route for target adhesion and division situation, is characterized in that, comprises the steps:

S1, vehicle tracking algorithm based on Maximum overlap area and Mean shift;

S11, the vehicle tracking algorithm of utilization based on Maximum overlap area carry out normal vehicle tracking;

S12, the tracking that utilizes the vehicle tracking algorithm based on Mean shift to carry out adhesion and divide vehicle;

Processing to target adhesion and division situation in S2, vehicle tracking process;

The match condition of S21, moving target to front and back frame is analyzed, and judges the state of moving target;

S22, utilize different algorithms to follow the tracks of according to the situation of the feedback of status of moving target.

2. when many vehicle trackings according to claim 1, the disposal route of target adhesion and division situation, is characterized in that, in step S11, the vehicle tracking algorithm of Maximum overlap area is:

Upper left corner coordinate (the x of vehicle in known k frame ₁, y ₁) and lower right corner coordinate (x ₂, y ₂), the upper left corner coordinate (x of vehicle in k+1 frame ₃, y ₃) and lower right corner coordinate (x ₄, y ₄), upper left corner coordinate and the left side, the lower right corner of establishing overlapping region are respectively: (x ₅, y ₅), (x ₆, y ₆):

$\left\{\begin{array}{c}{x}_5=\max (x_1,x_3)\\ x_6=\min (x_2,x_4)\\ y_5=\max (y_1,y_3)\\ y_6=\min (y_2,y_4)\end{array}\right.---\left(1\right)$

Can be tried to achieve respectively the upper left corner and the lower right corner of overlapping region by formula (1);

$\left\{\begin{array}{c}{x}_5<x_6\\ y_5<y_6\end{array}\right.---\left(2\right)$

If (2) formula is set up, illustrate that there is lap two rectangular areas, otherwise two rectangular areas are not overlapping, if (2) formula is set up, two rectangular area overlapping area S are

S=(x ₆-x ₅)*(y ₆-y ₅) （3）

Adjacent two frames detect in sequences, and the overlapping the maximum of area is the correct person of coupling, give target that this coupling is correct to number accordingly.

3. when many vehicle trackings according to claim 1, the disposal route of target adhesion and division situation, is characterized in that, step S12 is specially:

S121, manual search target window, set up object module

$q_u\left(\hat{y}\right)=\hat{C}{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{\hat{y}-{\hat{x}}_i}{\hat{h}}\left|\right|}^2\right)\mathrm{\&delta;}[b\left({\hat{x}}_i\right)-u]---\left(4\right)$

In formula, $\hat{C}=\frac{1}{{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{\hat{y}-{\hat{x}}_i}{\hat{h}}\left|\right|}^2\right)}$

$\mathrm{\&delta;}[b\left({\hat{x}}_i\right)-u]=\left\{\begin{array}{cc}1& b\left({\hat{x}}_i\right)=u\\ 0& b\left({\hat{x}}_i\right)\&NotEqual;u\end{array}\right.$

for the centre coordinate of object module, be the coordinate of i pixel, for the pixel value of point, the radius that h is To Template;

refer to the distance of i pixel to object module center;

for kernel function, when the larger k of x value (x) value less, otherwise when the less k of x value (x) is worth greatlyr, the weight of giving the closer to central point is larger;

for discrete impulse function, when time, equal 1; When time, equal 0, in the time that the pixel value traversing equals given pixel value, just numeration is once;

S122, set up candidate family at next frame

$C=\frac{1}{{\mathrm{\&Sigma;}}_{i=1}^{n}k\left({\left|\right|\frac{y-x_i}{h}\left|\right|}^2\right)}$

$\mathrm{\&delta;}[b\left(x_i\right)-u]=\left\{\begin{array}{cc}1& b\left(x_i\right)=u\\ 0& b\left(x_i\right)\&NotEqual;u\end{array}\right.$

In above formula, y ₀it is the centre coordinate of the To Template of previous frame;

S123, using Bhattacharyya coefficient as similarity function, obtain the iterative Bhattacharyya Coefficient Definition of Mean shift and be:

In the time that ρ (y) is maximum, illustrate that p (y) and q have maximum similarity;

Its first order Taylor is:

In formula,

(4), in formula, Section 1 and y are irrelevant, only need Section 2 to obtain maximal value so will try to achieve the value of ρ (y); Can obtain maximal value to Section 2 differentiate in (4)

Make f ' (y)=0, have:

So just, obtain Mean shift iterative:

Wherein, g (X)=-k ' is (X)

Because kernel function is given large weights for the pixel value of target's center, and give little weights for the pixel at wide center, thereby adopt kernel function histogram model modeling to guarantee that Mean shift algorithm itself has good robustness to blocking with the variation of background.

4. when many vehicle trackings according to claim 1, the disposal route of target adhesion and division situation, is characterized in that, in step S21, the state of moving target is normal condition, splitting status and adhesion state, it is characterized in that, step S21 is specially:

S211, front and back two frame vehicles are to illustrate that one to one the state of this moving target is normal condition;

S212, before determining, in the frame target situation that is normal target, in lower frame, there is area rapid drawdown, and non-enter monitoring cable place judge that dividing appears in target while for no reason there is fresh target;

S213, before determining, frame target be normal target in the situation that, occurs that in lower frame area increases suddenly, and non-go out monitoring cable place judge that adhesion appears in target while having target to suddenly disappear.

5. when many vehicle trackings according to claim 1, the disposal route of target adhesion and division situation, is characterized in that, step S22 is specially:

S221, one of drafting enter the line of guarded region, think that when target enters this line this is fresh target, judge whether this target meets vehicle feature;

S222, the target detecting in sequence in adjacent two frames is mated with Maximum overlap area algorithm, in the time there is following three kinds of situations, think that the match is successful: for without counting and tagger is mated one by one; As long as just think that for what have tagger to have to match the match is successful; As long as just think it the match is successful for what have numeration person to have to match;

S223, draw a line that rolls guarded region away from, roll this line away from when target and think that target rolls away from, delete the storage organization of this car.

6. when many vehicle trackings according to claim 5, the disposal route of target adhesion and division situation, is characterized in that, step S221 further comprises the steps:

In the time that this target meets vehicle feature, give this fresh target with numbering; If be less than the scope of minimum vehicle, judge that it is not a complete car, now do not award numbering, but count 0, wait for several frames after more clear testing result accurately numbering; If this target is greater than the scope of minimum vehicle, so think that this target is that several cars have occurred that adhesion phenomenon has caused wrong testing result, now gives this target with numbering, and give this target with mark, to represent that this target is as undesired target.

7. when many vehicle trackings according to claim 5, the disposal route of target adhesion and division situation, is characterized in that, step S222 further comprises the steps:

If in coupling time, is found not numbered by matcher, but there is numeration number, it is described vehicle in as fresh target is imperfect to stay, now, if having, it is matched, for its numeration adds 1 operation;

In the time of coupling, find there is mark by matcher, whether the total area that contrasts target in its two frame varies widely, if had, mark is removed, the target matching is with numbering, toply give original numbering, remaining give successively new numbering, if area does not have that great changes will take place, award corresponding numbering and mark;

In the time of coupling, target is one to one, directly give its corresponding numbering, in the time that coupling is unsuccessful, if there is the multiple object matchings that detect sequence in k frame to the same target that detects sequence in k+1 frame, now show in k frame it is correct result, but during to k+1 frame, there is the situation that several cars merge, now with Mean shift algorithm, those the several targets in k frame have been departed from the tracking of k+1 frame detection sequence;

If there is an object matching of sequence in k frame to the multiple targets that detect sequence in k+1 frame, correct result while now showing in k frame, but to k+1 frame there is the situation of a car division, now with Mean shift algorithm, that target in k frame is departed from the tracking of k+1 frame detection sequence.