CN104112282B - A method for tracking a plurality of moving objects in a monitor video based on on-line study - Google Patents

Abstract

The invention discloses a method for tracking a plurality of moving objects in a monitor video based on on-line study. The method is characterized by, to begin with, detecting a object region in a video sequence by utilizing an off-line training specific type detector; with the appearance characteristics being combined, associating the object between adjacent two frames by utilizing a dual-threshold conservative association concept to obtain a reliable conservative short tracking piece; defining positive and negative sample sets on the obtained tracking piece by utilizing time-space domain distribution constraint information, extracting colour, texture appearance characteristic similarity and motion information respectively to serve as an on-line study device training feature set, studying an on-line study algorithm through a machine and obtaining probability statistical features based on motion and appearance characteristics on the track piece distribution rules; and finally, converting the associating mode of the two track pieces into the problem of finding maximum posterior probability based on combination of the motion and the appearance. The method helps to solve the problem of track identity calibration mis-switching in the multi-moving-object tracking in a close-distance appearance-similar crowd scene.

Description

A kind of based on the method for multiple moving targets in on-line study tracing and monitoring video

Technical field

The invention belongs to mode identification technology, follow the tracks of video monitoring more particularly, to one based on on-line study In the method for multiple moving targets.

Background technology

The focus in video frequency object tracking always computer vision research field and difficulties are in order to retrieve Moving Object in Video Sequences track, and then provide effective guarantee for subsequent calculations machine visual system higher level recognition performance, right Accelerate automatic traffic management based on target following technology, intellectuality has the highest practical value.Traditional single goal with In track, problems faced includes: moving target is non-rigid, the deformation problems that visual angle, illumination variation cause, target travel random Property, target disappears and reappears, similar purpose interference in complex background, and background is blocked, and target is from blocking, and problems above makes Obtain the research of monocular track algorithm and there is the biggest challenge.(such as mesh in addition to traditional monocular follows the tracks of problem encountered Detection after reappearing after mark disappearance, the target jamming under complex background environment, occlusion issue etc.), the field that multiple target tracking processes Scape is more complicated, and uncertain factor is more.Many mesh are followed the tracks of also to be needed to solve problems with: moving target number is uncertain, identical Block between the interfering of target, similar target, how distinguishing similar little under the conditions of complex background or crowd scene Target etc..As under crowded street scene, pedestrian tracking application enjoys vast focus of attention.

In the pedestrian tracking problem in crowded street, pedestrian's external appearance characteristic is much like, and motor pattern is the most close, but with Meaning property is big especially, and blocking of person to person will cause being easy to mistake occur with phenomenon, and it is existing to have there presently does not exist this mistake root of solution The method for tracking target of elephant.

Summary of the invention

For disadvantages described above or the Improvement requirement of prior art, the invention provides a kind of based on on-line study tracking video The method of multiple moving targets in monitoring, it is intended that the track problem of calibrating occurred under the conditions of solving existing traffic congestion, Such as pedestrian tracking problem crowded under streetscape environment, traffic intersection vehicle tracking problem etc., multiple target tracking problem is converted into by it Track sheet related question step by step under maximum posteriori criterion (Maximum Posterior Probability) problem.Any Known class moving target and feasible on-line study mechanism are applicable to framework of the present invention.

For achieving the above object, according to one aspect of the present invention, it is provided that a kind of based on on-line study tracking video prison The method of multiple moving targets in control, comprises the following steps:

(1) receive input video sequence, utilize the pedestrian detector of off-line training and use multiple dimensioned traversal search frame Method demarcates the position of target in input video sequence；

(2) use conservative correlating method based on color characteristics to detection mesh between two continuous frames in input video sequence Mark carries out data association, to obtain multiple reliable conservative pursuit path sheet in short-term；

(3) motion profile and the color similarity characteristic according to pursuit path sheet in short-term and utilizing same target builds Positive and negative samples collection, wherein positive sample set is all from the movement locus of same target, and negative sample collection is from the motion rail of different target Mark；

(4) utilizing the positive and negative samples collection training Hough random forest built, the leaf node of this Hough random forest kind is deposited Classification statistical property and the orbiting motion characteristic of positive and negative samples collection are stored up；

(5) utilize time-domain constraints characteristic will follow the tracks of head and the tail interval difference in sheet in short-term and be less than threshold value T_pAny two in short-term with Track sheet be built into may association track pair, all of may association track to constitute may association track to set, by each can Track can be associated to carrying out feature description, to generate the feature set of detection sample, and the feature set of detection sample is sent into Hough In forest, the leaf node corresponding to obtain this feature collection；

(6) the classification statistical property of leaf node is obtainedWith movement statistics characteristicThus obtain The association probability of relevant track pairAll association probabilities constitute association probability square Battle array；

(7) judge that two maximum elements horizontal in association probability matrix, vertical, whether more than a threshold value, are the most tentatively sentenced The track sheet that disconnected the two element is corresponding respectively belongs to same track, then proceeds to step (8), otherwise represents and be not belonging to same rail Mark, then proceeds to step (8)；

(8) with Hungary Algorithm, two elements are judged again, finally to determine that the track sheet of the two correspondence belongs to Same track；Only Hungary Algorithm judges belong to same track and judge to belong to the track sheet of same track in step (7) Just it is considered real association track pair；

(9) two the element repeated execution of steps (3) belonging to same track obtained step (8), to obtain new instruction Practice sample set, and this training sample set is performed the learning procedure of step (4)；

(10) threshold value T is increased_pValue, and repeat the above steps (5) to (9), until can not the track that can associate of regeneration Till to.

Preferably, step (3) includes following sub-step:

(3-1) extract colouring informations and the positional information of two detection targets on pursuit path sheet in short-term at random, be used for giving birth to Becoming positive sample, multiple positive samples constitute positive sample set；

(3-2) in different a pair in short-term pursuit path sheet, extract colouring information and the position of a detection target respectively Information, is used for generating negative sample, and multiple negative samples constitute negative sample collection.

Preferably, step (4) includes following sub-step:

(4-1) feature set A={F of input training sample is generated according to positive and negative samples collection₁=(x₁, y₁), F₂=(x₂, y₂)......F_n=(x_n, y_n), wherein y_i∈ 0,1, i=1,2...n, n represent that positive and negative samples concentrates the number of sample, x_iIt is The characteristic vector of i-th input training sample, and x_i={ f_color,f_ent,f_motion, f_color,Represent i-th input training sample Color histogram similarity, f_ent,Represent the local gray level entropy partial binary similarity of i-th input training sample, f_motionRepresent the motion excursion amount of i-th input training sample, andP1 and p2 Represent two detection target central point positions in its place frame that i-th input training sample is randomly drawed when building respectively Put, p1 '=p2-v2 (t2-t1), p2 '=p1+v1 (t2-t1), wherein v1 with v2 corresponding i-th input training sample respectively The speed of two detection targets corresponding during this structure, t1 with t2 corresponding i-th input training sample respectively is corresponding when building The time frame of two detection targets, y_iIt it is the classification scalar of i-th input training sample；

(4-2) utilize recurrence partitioning that feature set A of input training sample is constantly divided, random to generate Hough Forest, until the leaf node of Hough random forest meets end condition, the final classification statistics spy obtaining leaf node Property and movement statistics characteristic.

Preferably, y is worked as_iWhen=0, represent that this sample is negative sample, i.e. characteristic vector x_iFrom different track sheets, work as y_i When=1, represent that this sample is positive sample, i.e. characteristic vector x_iFrom same track sheet.

Preferably, the end condition of leaf node is: the sample set quantity preserved in (1) leaf node is less than the first threshold Value, the size of this threshold value is determined by the category Properties of positive and negative samples collection；(2) degree of depth of Hough random forest is less than Second Threshold, The size of this threshold value is determined by the feature set information inputting training sample.

Preferably, the partiting step of the sample set S arriving certain node k is as follows:

(4-2-1) random choose characteristic vector x_i={ f_color,f_ent, and therefrom random choose f_color,Or f_entAs joint Point divides threshold value, the f selected_color,Or f_entAnd characteristic vector x_iConstitute parameter pond { τ_k, with the node division threshold value selected it is Example, is left child node less than or equal to the training sample of this node division threshold value, more than the training sample of this node division threshold value For right child node, sample set corresponding to left and right child node is S_L, S_R:

$S_L\left({\mathrm{\φ}}_{{\mathrm{\τ}}_k}\right)=\{F_i\∈S|f_{{\mathrm{\φ}}_{{\mathrm{\τ}}_k}}\left(F_i\right)=0\}$

$S_R\left({\mathrm{\φ}}_{{\mathrm{\τ}}_k}\right)=\{F_i\∈S|f_{{\mathrm{\φ}}_{{\mathrm{\τ}}_k}}\left(F_i\right)=1\};$

(4-2-2) classification uncertainty measure U is obtained₁(S)=| S | H (Y) or the orbiting motion skew uncertain survey of concordance DegreeWherein | S | represents the sample set S number arriving node k, S₊In expression S just Sample set number,For the motion excursion average that positive samples all in node k are corresponding, H (Y) is the classification entropy of sample set S；

(4-2-3) select to make classification uncertainty measure U₁(S) or orbiting motion skew concordance uncertainty measure U₂(S) Big optimized parameter τ_k*, so that two child nodes relatively divide front nodal point uncertainty decline maximum after Hua Fening；

(4-2-4) at the sample set S that left and right child node is corresponding_L, S_ROn the basis of continue to divide left and right child node, and repeat Above-mentioned steps (4-2-1) to (4-2-3) is to obtain optimized parameter, until meeting end condition, and the terminator finally given Node is exactly leaf node；

(4-2-5) the classification statistical property of leaf node is calculatedWith movement statistics characteristicIts point It is not

$p\left(y_{\mathrm{app}}^{+}\right|L)={\mathrm{\ψ}}_L\left(S_L\right)/N_L$

Wherein N_LRepresent sample set number in leaf node L, ψ_L(S_L) represent positive number of samples in leaf node L；

Movement statistics characteristic is to use to obtain based on the gaussian kernel-Parzen window estimation technique:

$p\left(y_{\mathrm{mot}}^{+}\right|L)=\frac{1}{{\mathrm{\ψ}}_L\left(S_L\right)}\left(\underset{f_{\mathrm{motion}}^i\∈L}{\mathrm{\Σ}}\frac{1}{{2\mathrm{\π\σ}}^2}\exp (-\frac{{\left|\right|f_{\mathrm{motion}}^i-{\stackrel{\‾}{f}}_{\mathrm{motion}}\left|\right|}^2}{{2\mathrm{\σ}}^2})\right)$

Wherein σ is variance, and its value is 5.

In general, by the contemplated above technical scheme of the present invention compared with prior art, it is possible to show under acquirement Benefit effect:

1, the present invention obtains training sample set the most accurately by the conservative sheet of tracking in short-term, thus tree is effectively ensured and learns Accuracy during habit, has taken into full account classification information and movable information, this and trace model in the learning process to tree Foundation be consistent, by using Hungary Algorithm to be further ensured that reliable association results, use Increment Learning Algorithm to improve tree Study accuracy, thus solve present in existing method wrong with problem.

2, the method applied in the present invention may be used for the demarcation of crowded street pedestrian movement's track.Especially can alleviate phase Easily there is the bottleneck problem that track identity mistake is followed the tracks of in patibhaga-nimitta close-target, and also has higher robust with regard to track sheet disruption simultaneously Property.

3, effective guarantee is provided to subsequent calculations machine visual system higher level recognition performance, to accelerating based on pedestrian tracking skill The automatic traffic management of art, intellectuality have the highest practical value, and concrete application can relate to friendship and regulate reason, robot The application scenarios such as navigation..

4, the present invention is not only limited to pedestrian tracking, it is adaptable to the movement objective orbit of any known class is followed the tracks of application and needed Ask.

Accompanying drawing explanation

Fig. 1 is that the present invention follows the tracks of the flow chart of the method for multiple moving targets in video monitoring based on on-line study；

Fig. 2 is the video frame images example used in the embodiment of the present invention；

Fig. 3 is that the present invention implements video frame images testing result used；

Fig. 4 (a) is some detection example of the image to be detected used in the embodiment of the present invention, and (b) is that this detection image is corresponding The effective coverage picked out under local gray level maximum entropy principle, (c) and local gray level entropy schematic diagram；

Fig. 5 is that training stage orbiting motion side-play amount defines schematic diagram；

Fig. 6 is to differentiate phase trajectory motion excursion amount definition schematic diagram；

Fig. 7 (a) and (b) be the video frame images sample used in the embodiment of the present invention in TUD frame of video corresponding Final orbit mark calibration result schematic diagram (track result is rectangle dashed box region in figure)；

Fig. 8 (a) and (b) be the video frame images sample used in the embodiment of the present invention in ETH frame of video corresponding Final orbit mark calibration result schematic diagram (track result is rectangle dashed box region in figure).

Detailed description of the invention

In order to make the purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, right The present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, and It is not used in the restriction present invention.If additionally, technical characteristic involved in each embodiment of invention described below The conflict of not constituting each other just can be mutually combined.

The Integral Thought of the present invention is, it includes running frame by frame pedestrian detector, demarcate testing result, two continuous frames it Between conservative association obtain reliably following the tracks of sheet in short-term, utilizing space-time restriction, (feature set structure includes to build positive and negative sample characteristics collection Color histogram similarity, histograms of oriented gradients similarity, partial binary characteristic vector under local gray level maximum entropy principle Similarity, and offset homogeneity measure based on the movement locus under movement locus continuously smooth assumed condition), based on classification not Hough forest training under degree of certainty and motion excursion concordance uncertainty measure, definition can associate track pair, and formalization can close The association probability of connection track pair is outward appearance and Motion-Joint posterior probability, obtains association results under MAP criterion, uses Hungary to divide Join algorithm and correct association error, then on the longest track sheet, rebuild training sample, update forest leaf node Statistical property, and as newly inputted iteration, reliably the longest track sheet is associated flow process, until not having any to associate track Till sheet.

Include following as it is shown in figure 1, the present invention follows the tracks of the method for multiple moving targets in video monitoring based on on-line study Step:

(1) receive input video sequence, utilize the pedestrian detector of off-line training and use multiple dimensioned traversal search frame Method demarcates the position of target in input video sequence；

(2) use conservative correlating method based on color characteristics to detection mesh between two continuous frames in input video sequence Mark carries out data association, to obtain multiple reliable conservative pursuit path sheet in short-term；

(3) motion profile and the color similarity characteristic according to pursuit path sheet in short-term and utilizing same target builds Positive and negative samples collection, wherein positive sample set is all from the movement locus of same target, and negative sample collection is from the motion rail of different target Mark；This step includes following sub-step:

(3-1) extract colouring informations and the positional information of two detection targets on pursuit path sheet in short-term at random, be used for giving birth to Becoming positive sample, multiple positive samples constitute positive sample set；

(3-2) in different a pair in short-term pursuit path sheet, extract colouring information and the position of a detection target respectively Information, is used for generating negative sample, and multiple negative samples constitute negative sample collection；

(4) positive and negative samples collection training Hough random forest (Hough Forest) built, this Hough random forest are utilized The leaf node planted stores classification statistical property and the orbiting motion characteristic of positive and negative samples collection；This step specifically includes following Sub-step:

(4-1) feature set A={F of input training sample is generated according to positive and negative samples collection₁=(x₁, y₁), F₂=(x₂, y₂)......F_n=(x_n, y_n), wherein y_i∈ 0,1, i=1,2...n, n represent that positive and negative samples concentrates the number of sample, x_iIt is The characteristic vector of i-th input training sample, and x_i={ f_color,f_ent,f_motion, f_color,Represent i-th input training sample Color histogram similarity, f_ent,Represent the local gray level entropy partial binary similarity of i-th input training sample, f_motionRepresent the motion excursion amount of i-th input training sample, andSuch as Fig. 5 institute Showing, two detection targets that p1 and p2 randomly draws when representing i-th input training sample structure respectively are in its place frame Center position, p1 '=p2-v2 (t2-t1), p2 '=p1+v1 (t2-t1), wherein v1 with v2 corresponding i-th respectively is defeated Entering the speed of two detection targets corresponding when training sample builds, t1 with t2 corresponding i-th input training sample respectively builds Time corresponding two detection targets time frames, y_iIt it is the classification scalar of i-th input training sample；Work as y_i=0, represent this sample Originally it is negative sample, i.e. characteristic vector x_iFrom different track sheets, work as y_iWhen=1, represent this sample be positive sample, i.e. feature to Amount x_iFrom same track sheet；

(4-2) utilize recurrence partitioning that feature set A of input training sample is constantly divided, random to generate Hough Forest, until the leaf node of Hough random forest meets end condition, the final classification statistics spy obtaining leaf node Property and movement statistics characteristic, wherein the end condition of leaf node is: the sample set quantity preserved in (1) leaf node is less than the One threshold value, the size of this threshold value is determined by the category Properties of positive and negative samples collection, if the similarity of positive and negative samples collection is the lowest, should Threshold value is the biggest, otherwise the least, and in the present embodiment, threshold value is 20；(2) degree of depth of Hough random forest is less than the second threshold Value, the size of this threshold value is determined by the feature set information (classification and movable information) inputting training sample, the most complicated then threshold of information It is worth the biggest, on the contrary the least, and in present embodiment, this threshold value is 15；Wherein arrive the partiting step of the sample set S of certain node k As follows:

(4-2-1) random choose characteristic vector x_i={ f_color,f_ent, and therefrom random choose f_color,Or f_entAs joint Point divides threshold value, the f selected_color,Or f_entAnd characteristic vector x_iConstitute parameter pond { τ_k, with the node division threshold value selected it is Example, is left child node less than or equal to the training sample of this node division threshold value, more than the training sample of this node division threshold value For right child node, sample set corresponding to left and right child node is S_L, S_R:

$S_L\left({\mathrm{\φ}}_{{\mathrm{\τ}}_k}\right)=\{F_i\∈S|f_{{\mathrm{\φ}}_{{\mathrm{\τ}}_k}}\left(F_i\right)=0\}$

$S_R\left({\mathrm{\φ}}_{{\mathrm{\τ}}_k}\right)=\{F_i\∈S|f_{{\mathrm{\φ}}_{{\mathrm{\τ}}_k}}\left(F_i\right)=1\};$

(4-2-2) classification uncertainty measure U is obtained₁(S)=| S | H (Y) or the orbiting motion skew uncertain survey of concordance DegreeWherein | S | represents the sample set S number arriving node k, S₊In expression S just Sample set number,For the motion excursion average that positive samples all in node k are corresponding, H (Y) is the classification entropy of sample set S；

(4-2-3) select to make classification uncertainty measure U₁(S) or orbiting motion skew concordance uncertainty measure U₂(S) Big optimized parameter τ_k*, so that two child nodes relatively divide front nodal point uncertainty decline maximum after Hua Fening；

(4-2-4) at the sample set S that left and right child node is corresponding_L, S_ROn the basis of continue to divide left and right child node, and repeat Above-mentioned steps (4-2-1) to (4-2-3) is to obtain optimized parameter, until meeting end condition, and the terminator finally given Node is exactly leaf node；

(4-2-5) the classification statistical property of leaf node is calculatedWith movement statistics characteristicIts point It is not

$p\left(y_{\mathrm{app}}^{+}\right|L)={\mathrm{\ψ}}_L\left(S_L\right)/N_L$

Wherein N_LRepresent sample set number in leaf node L, ψ_L(S_L) represent positive number of samples in leaf node L.

Movement statistics characteristic is to use to obtain based on the gaussian kernel-Parzen window estimation technique:

$p\left(y_{\mathrm{mot}}^{+}\right|L)=\frac{1}{{\mathrm{\ψ}}_L\left(S_L\right)}\left(\underset{f_{\mathrm{motion}}^i\∈L}{\mathrm{\Σ}}\frac{1}{{2\mathrm{\π\σ}}^2}\exp (-\frac{{\left|\right|f_{\mathrm{motion}}^i-{\stackrel{\‾}{f}}_{\mathrm{motion}}\left|\right|}^2}{{2\mathrm{\σ}}^2})\right)$

Wherein σ is variance, and its value is 5.

(5) utilize time-domain constraints characteristic will follow the tracks of head and the tail interval difference in sheet in short-term and be less than threshold value T_p(its value and in short-term with The length of track sheet is directly proportional, and is 8 frames in the present embodiment) any two follow the tracks of in short-term sheet be built into may association track Right, each track that may associate, to set, is retouched by all of track the associate track possible to composition that may associate to carrying out feature State, to generate the feature set of detection sample, and the feature set of detection sample is sent in Hough forest, to obtain this feature set pair The leaf node answered；

Carry out feature description particularly as follows: association track on each track on extract respectively one detection target face Color information and positional information (as shown in Figure 6), be used for generating detection sample, and multiple detection samples constitute detection sample set, according to Detection sample set generates the feature set of detection sample, and its process and above-mentioned steps (4-1) are essentially identical, and unique difference is not have class Other scalar y_i；

(6) the classification statistical property of leaf node is calculated according to the formula in above step (4-2-5)And fortune Dynamic statistical propertyThus obtain the association probability of relevant track pair be defined as:All association probabilities constitute association probability matrix.

(7) judge whether two maximum elements horizontal in association probability matrix, vertical are more than a threshold value and (in present embodiment are 0.5), the most tentatively judge that the track sheet that the two element is corresponding respectively belongs to same track, then proceed to step (8), Otherwise represent and be not belonging to same track, then proceed to step (8)；

(8) with Hungary Algorithm, two elements are judged again, finally to determine that the track sheet of the two correspondence belongs to Same track；Only Hungary Algorithm judges belong to same track and judge to belong to the track sheet of same track in step (7) Just it is considered real association track pair；

(9) two the element repeated execution of steps (3) belonging to same track obtained step (8), to obtain new instruction Practice sample set, and this training sample set is performed the learning procedure of step (4)；

(10) threshold value T is increased_pValue (increasing by 10 in the present embodiment), and repeat the above steps (5) is to (9), until Can not the track that can associate of regeneration to till.

An instantiation be given below:

(1) receive input video sequence, utilize the pedestrian detector of off-line training and use multiple dimensioned traversal search frame Method demarcates the position of target in input video sequence, and if Fig. 2 is certain frame video sequence, this project uses detector to be classical HOG+SVM pedestrian detection algorithm, its testing result is with reference to Fig. 3.

(2) obtain, frame by frame after testing result, using conservative correlating method based on color characteristics in input video sequence Detection target between two continuous frames carries out data association, to obtain multiple reliable conservative pursuit path sheet in short-term.Concrete step Suddenly include, by same for testing result normalization size, extracting testing result between adjacent two frames of color histogram characteristic vector Setting up adjacent two frame incidence matrix, two detect the color histogram similarity characterization associated confidence that target areas are corresponding, namely Color similarity is the highest compared with other coupling combinations, and adjacent two detection targets are understood to from same more than in the time domain of threshold value Track, and then pursuit path sheet the most in short-term can be obtained.

(3) motion profile and the color similarity characteristic according to pursuit path sheet in short-term and utilizing same target builds Positive and negative samples collection, wherein positive sample set is all from the movement locus of same target, and negative sample collection is from the motion rail of different target Mark.It is implemented as:

(3-1) extract colouring informations and the positional information of two detection targets on pursuit path sheet in short-term at random, be used for giving birth to Becoming positive sample, multiple positive samples constitute positive sample set；

(3-2) in different a pair in short-term pursuit path sheet, extract colouring information and the position of a detection target respectively Information, is used for generating negative sample, and multiple negative samples constitute negative sample collection；

(4) positive and negative samples collection training Hough random forest (Hough Forest) utilizing (3) to build is embodied as step As follows:

(4-1) first extracting positive and negative sample characteristics collection, positive and negative sampling feature vectors is made up of three parts, and detection target is corresponding The color histogram f of pixel region_color,, based on the effective coverage partial binary selected under local gray level entropy principle, son is described f_ent,(seeing Fig. 4), and same orbiting motion information (orbiting motion skew uncertainty)P1 and p2 represents what i-th input training sample was randomly drawed when building respectively Two detection target center position in its place frame, p1 '=p2-v2 (t2-t1), p2 '=p1+v1 (t2-t1), The wherein speed of two detection targets that v1 with v2 corresponding i-th input training sample respectively is corresponding when building, t1 and t2 is respectively The time frame of two detection targets that corresponding i-th input training sample is corresponding when building, schematic diagram sees Fig. 5.Obviously, positive sample This detects region from same track two, and color histogram similarity should be higher, and effective coverage is selected binary system and described son and also should Similarity is higher, and the motion of same rail has higher flatness, and the distribution of Fixed Time Interval bias internal should have very strong regular.Negative Sample set detects area pixel corresponding color rectangular histogram from different tracks two, and local entropy is selected the local two of maximum region and entered System despises feature, it is clear that because from different targets, the more same track of similarity is low, and the motion of negative sample concentration solution locus is inclined Shifting amount regularity is inconspicuous.

(4-2) after above-mentioned process, construct a large amount of positive and negative sample characteristics Ji Ji, start below to train Hough forest Learn.Hough forest is cascade decision tree, and the core concept of learning algorithm is to find optimal dividing letter for each leaf node that flies Number, the sample set of this non-leaf nodes is divided into left and right can not two child nodes.Estimating of optimization function is adopted respectively Offset uncertain concordance with classification uncertainty and orbiting motion, i.e. require partition function as far as possible by generic sample It is divided into same node, or positive sample similar for motion excursion is gathered same node.Specifically comprise the following steps that

(4-2-1) random choose characteristic vector x_i={ f_color,f_ent, and therefrom random choose f_color,Or f_entAs joint Point divides threshold value, the f selected_color,Or f_entAnd characteristic vector x_iConstitute parameter pond { τ_k, with the node division threshold value selected it is Example, is left child node less than or equal to the training sample of this node division threshold value, more than the training sample of this node division threshold value For right child node, sample set corresponding to left and right child node is S_L, S_R:

(4-2-2) classification uncertainty measure U is obtained₁(S)=| S | H (Y) or the orbiting motion skew uncertain survey of concordance Degree $U_2\left(S\right)=\frac{1}{\left|S_{+}\right|}{\mathrm{\Σ}}_{i\∈S_{+}}{\left|\right|f_{\mathrm{motion}}^i-{\stackrel{\‾}{f}}_{\mathrm{motion}}\left|\right|}^2,$

(4-2-3) select to make classification uncertainty measure U₁(S) or orbiting motion skew concordance uncertainty measure U₂(S) Big optimized parameter τ_k*, so that two child nodes relatively divide front nodal point uncertainty decline maximum after Hua Fening；

(4-2-4) at the sample set S that left and right child node is corresponding_L, S_ROn the basis of continue to divide left and right child node, and repeat Above-mentioned steps (4-2-1) to (4-2-3) is to obtain optimized parameter, until meeting end condition, and the terminator finally given Node is exactly leaf node；

(4-2-5) the classification statistical property of leaf node is calculatedWith movement statistics characteristic

So far, the study of Hough forest is complete.

(5) utilize time-domain constraints characteristic will follow the tracks of head and the tail interval difference in sheet in short-term and be less than threshold value T_pAny two in short-term with Track sheet be built into may association track pair, all of may association track to constitute may association track to set, by each can Track can be associated to carrying out feature description, to generate the feature set of detection sample, and the feature set of detection sample is sent into Hough In forest, the leaf node corresponding to obtain this feature collection；

Detection sample characteristics describe particularly as follows: first association track on each track on extract a detection respectively The colouring information of target and positional information (as shown in Figure 6), be used for generating detection sample, and multiple detection samples constitute detection sample Collection, (4-1) method of employing obtains the characteristic descriptor set of sample set, and unique difference is not have classification scalar y_i；

(6) the classification statistical property of the leaf node obtained according to above step (4-2-5)And motion Statistical propertyCalculate the association probability of the relevant track pair of institute calmlyInstitute Relevant probability constitutes association probability matrix.

(7) judge whether two maximum elements horizontal in association probability matrix, vertical are more than a threshold value and (in present embodiment are 0.5), the most tentatively judging that the track sheet that the two element is corresponding respectively belongs to same track, otherwise expression is not belonging to same One track, then proceeds to step (8)；

(8) with Hungary Algorithm, two elements are judged again, finally to determine that the track sheet of the two correspondence belongs to Same track；Only Hungary Algorithm judges belong to same track and judge to belong to the track sheet of same track in step (7) Just it is considered real association track pair；

(9) two the element repeated execution of steps (3) belonging to same track obtained step (8), to obtain new instruction Practice sample set, and this training sample set is performed the learning procedure incremental learning Hough forest of step (4) to improve leaf node The precision of statistical property；

(10) threshold value T is increased_pValue (increasing by 10 in the present embodiment), and repeat the above steps (5) is to (9), until Can not the track that can associate of regeneration to till.

Sum it up, the invention discloses motion target tracking method in a kind of monitor video based on on-line study.Suitable Follow the tracks of for all safety check Moving Object in Video Sequences, such as pedestrian movement's track following in user and streetscape scene, vehicle rail Mark demarcation etc..Multiple target tracking problem is converted into track sheet related question step by step under MAP problem by this patent.First with from The particular category of line training is surveyed device and is detected target area in video sequence；Then in conjunction with appearance characteristics, dual threshold is used to guard Association thinking associates the target between adjacent two frames, obtains reliable conservative short tracking sheet；Again on the tracking sheet obtained, utilize Time-space domain distribution constraint information, defines positive and negative sample set, extracts color, texture appearance characteristic similarity and motion letter respectively Breath, as on-line study device training characteristics collection, by the training process of machine learning on-line learning algorithm, obtains the distribution of track sheet Based on motion and the probabilistic statistical characteristics of appearance characteristics in rule；Finally turn to solve based on motion by two track sheet correlation forms Associating posterior probability greatest problem with outward appearance；Association is embodied in the association last association results collected as next step by step Input, rebuilds sample training collection, the probability nature of incremental learning renewal learning algorithm model, closes by being gradually increased track Connection time interval, obtains the track sheet of longer time.For this structure framework relatively off-line training template based on on-line study There is more preferable adaptive ability, be more suitable for track sheet study instantly, but on-line study is machine-processed, due to the training sample of next stage This collection, from the output result of upper level, if upper level output result has error, can produce the accumulation of error.In order to avoid this Situation, is further introduced into Hungary based on appearance characteristics allocation algorithm, corrects every grade of learning algorithm association results, be greatly improved The robustness of algorithm, especially alleviates what multiple mobile object tracking problem under the crowd scene that closely outward appearance is similar occurred Track identity demarcates frequent switching problem.This project on-line learning algorithm build mechanism can be that computer vision neighborhood is conventional appoints One learning algorithm: boosting, SVM, decision tree etc..

As it will be easily appreciated by one skilled in the art that and the foregoing is only presently preferred embodiments of the present invention, not in order to Limit the present invention, all any amendment, equivalent and improvement etc. made within the spirit and principles in the present invention, all should comprise Within protection scope of the present invention.

Claims (6)

1. follow the tracks of the method for multiple moving targets in video monitoring based on on-line study for one kind, it is characterised in that include following step Rapid:

(1) input video sequence, the pedestrian detector utilizing off-line training the method using multiple dimensioned traversal search frame are received Demarcate the position of target in input video sequence；

(2) use conservative correlating method based on color characteristics that detection target between two continuous frames in input video sequence is entered Row data association, to obtain multiple reliable conservative pursuit path sheet in short-term；

(3) motion profile and the color similarity characteristic according to pursuit path sheet in short-term and utilizing same target builds positive and negative Sample set, wherein positive sample set is all from the movement locus of same target, and negative sample collection is from the movement locus of different target；

(4) utilizing the positive and negative samples collection training Hough random forest built, the leaf node of this Hough random forest kind stores The classification statistical property of positive and negative samples collection and orbiting motion characteristic；

(5) utilize time-domain constraints characteristic will follow the tracks of head and the tail interval difference in sheet in short-term and be less than threshold value T_pAny two follow the tracks of sheet in short-term Being built into the possible track pair that associates, all of possible track the associate track possible to composition that associate is to set, by each possible pass The feature set of detection sample, to carrying out feature description, to generate the feature set of detection sample, and is sent into Hough forest by connection track In, the leaf node corresponding to obtain this feature collection；

(6) the classification statistical property of leaf node L is obtainedWith movement statistics characteristicThus owned The association probability of association track pairAll association probabilities constitute association probability matrix；

(7) judge that two maximum elements horizontal in association probability matrix, vertical, whether more than a threshold value, the most tentatively judge this The track sheet of two element correspondences respectively belongs to same track, then proceeds to step (8), otherwise represents and be not belonging to same track, Then step (8) is proceeded to；

(8) with Hungary Algorithm, two elements are judged again, finally to determine that the track sheet of the two correspondence belongs to same Track；Only Hungary Algorithm judges belong to same track and judge that in step (7) the track sheet belonging to same track is just recognized For being real association track pair；

(9) two the element repeated execution of steps (3) belonging to same track obtained step (8), to obtain new training sample This collection, and this training sample set is performed the learning procedure of step (4)；

(10) threshold value T is increased_pValue, and repeat the above steps (5) to (9), until can not the track that can associate of regeneration to for Only.

Method the most according to claim 1, it is characterised in that step (3) includes following sub-step:

(3-1) colouring informations and the positional information of two detection targets on pursuit path sheet in short-term are extracted at random, for just generating Sample, multiple positive samples constitute positive sample set；

(3-2) in different a pair in short-term pursuit path sheet, extract colouring information and the position letter of a detection target respectively Breath, is used for generating negative sample, and multiple negative samples constitute negative sample collection.

Method the most according to claim 2, it is characterised in that step (4) includes following sub-step:

(4-1) feature set of input training sample is generated according to positive and negative samples collection Wherein, y_i∈) 0}, 1, i=1,2...n, n represent that positive and negative samples concentrates the number of sample, x_iIt is that i-th inputs training sample Characteristic vector, and x_i={ f_color,f_ent,f_motion, f_color, represent the color histogram similarity of i-th input training sample, f_ent, represent the local gray level entropy partial binary similarity of i-th input training sample, f_motionRepresent i-th input training The motion excursion amount of sample, andP1 and p2 represents i-th input training sample respectively Two the detection target center position in its place frame randomly drawed during this structure, p1 '=p2-v2 (t2-t1), P2 '=p1+v1 (t2-t1), two detection mesh that wherein v1 with v2 corresponding i-th input training sample respectively is corresponding when building Target speed, the time frame of two detection targets that t1 with t2 corresponding i-th input training sample respectively is corresponding when building, y_iIt is The classification scalar of i-th input training sample；

(4-2) utilize recurrence partitioning that feature set A of input training sample is constantly divided, the most gloomy to generate Hough Woods, until the leaf node of Hough random forest meets end condition, the final classification statistical property obtaining leaf node With movement statistics characteristic.

Method the most according to claim 3, it is characterised in that work as y_iWhen=0, represent that this sample is negative sample, i.e. feature to Amount x_iFrom different track sheets, work as y_iWhen=1, represent that this sample is positive sample, i.e. characteristic vector x_iFrom same track sheet.

Method the most according to claim 3, it is characterised in that the end condition of leaf node is: protect in (1) leaf node The sample set quantity deposited is less than first threshold, and the size of this threshold value is determined by the category Properties of positive and negative samples collection；(2) Hough with The degree of depth of machine forest is less than Second Threshold, and the size of this threshold value is determined by the feature set information inputting training sample.

Method the most according to claim 3, it is characterised in that arrive the partiting step of sample set S of certain node k such as Under:

(4-2-1) random choose characteristic vector x_i={ f_color,f_ent, and therefrom random choose f_color, or f_entDraw as node Divide threshold value, the f selected_color, or f_entAnd characteristic vector x_iConstitute parameter pond { τ_k, as a example by the node division threshold value selected, It is left child node less than or equal to the training sample of this node division threshold value, is right more than the training sample of this node division threshold value Child node, sample set corresponding to left and right child node is S_L, S_R:

$S_L\left({\mathrm{\φ}}_{{\mathrm{\τ}}_k}\right)=\{F_i\∈S|f_{{\mathrm{\φ}}_{{\mathrm{\τ}}_k}}\left(F_i\right)=0\}$

$S_R\left({\mathrm{\φ}}_{{\mathrm{\τ}}_k}\right)=\{F_i\∈S|f_{{\mathrm{\φ}}_{{\mathrm{\τ}}_k}}\left(F_i\right)=1\};$

(4-2-2) classification uncertainty measure U is obtained₁(S)=| S | H (Y) or orbiting motion skew concordance uncertainty measureWherein | S | represents the sample set S number arriving node k, S₊Represent positive sample in S This collection number,For the motion excursion average that positive samples all in node k are corresponding, H (Y) is the classification entropy of sample set S；

(4-2-3) select to make classification uncertainty measure U₁(S) or orbiting motion skew concordance uncertainty measure U₂(S) maximum Optimized parameter τ_k*, so that two child nodes relatively divide front nodal point uncertainty decline maximum after Hua Fening；

(4-2-4) at the sample set S that left and right child node is corresponding_L, S_ROn the basis of continue to divide left and right child node, and repeat above-mentioned Step (4-2-1) to (4-2-3) is to obtain optimized parameter, until meeting end condition, and the terminator node finally given It it is exactly leaf node；

(4-2-5) the classification statistical property of leaf node is calculatedWith movement statistics characteristicIt is respectively

$p\left(y_{app}^{+}\right|L)={\mathrm{\ψ}}_L\left(S_L\right)/N_L$

Wherein N_LRepresent sample set number in leaf node L, ψ_L(S_L) represent positive number of samples in leaf node L；

Movement statistics characteristic is to use to obtain based on the gaussian kernel-Parzen window estimation technique:

$p\left(y_{mot}^{+}\right|L)=\frac{1}{{\mathrm{\ψ}}_L\left(S_L\right)}\left(\underset{f_{motion}^i\∈L}{\Σ}\frac{1}{2{\mathrm{\π\σ}}^2}\exp \right(\frac{{||f_{motion}^i-{\stackrel{\‾}{f}}_{motion}||}^2}{2{\mathrm{\σ}}^2}\left)\right)$

Wherein σ is variance, and its value is 5.