CN104112282A - 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 method based on multiple moving targets in on-line study tracing and monitoring video

Technical field

The invention belongs to mode identification technology, more specifically, relate to a kind of method of following the tracks of multiple moving targets in video monitoring based on on-line study.

Background technology

Video frequency object tracking is focus and the difficulties in computer vision research field always, object is in order to retrieve Moving Object in Video Sequences track, and then for the more senior recognition performance of subsequent calculations machine vision system provides effective guarantee, automatic traffic management, the intellectuality of accelerating based target tracking technique are had to very high practical value.The problem facing in traditional monotrack comprises: moving target is non-rigid, the deformation problems that visual angle, illumination variation cause, the randomness of target travel, target disappears and reappears, similar target jamming in complex background, background is blocked, and target is from blocking etc., and above these problems make the research of monocular track algorithm have very large challenge.Except traditional monocular is followed the tracks of the problem facing (such as, detection after target is reappeared after disappearing, target jamming, the occlusion issue etc. under complex background environment), the scene of multiple target tracking processing is more complicated, and uncertain factor is more.Many orders are followed the tracks of and are also needed to solve following problem: between the phase mutual interference of moving target number uncertainty, same target, similar target, block, the little target of distinguishing similar etc. under complex background or crowd scene condition how.As following the tracks of application, pedestrian under crowded street scene enjoys vast focus of attention.

In pedestrian's tracking problem in crowded street, pedestrian's external appearance characteristic is very similar, and motor pattern is very close, but randomness is large especially, blocking of person to person will cause being easy to occurring the wrong phenomenon of following, and still not have the method for tracking target that solves this mistake root phenomenon at present.

Summary of the invention

For above defect or the Improvement requirement of prior art, the invention provides a kind of method of following the tracks of multiple moving targets in video monitoring based on on-line study, its object is, solve the track problem of calibrating occurring under existing congested in traffic condition, as crowded pedestrian's tracking problem under streetscape environment, traffic intersection vehicle tracking problems etc., multiple target tracking problem is converted into track sheet related question step by step under maximum posteriori criterion (Maximum Posterior Probability) problem by it.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, provide a kind of method of following the tracks of multiple moving targets in video monitoring based on on-line study, comprised the following steps:

(1) receive input video sequence, utilize the pedestrian detector of off-line training and adopt the method for multiple dimensioned traversal search frame to demarcate the position of target in input video sequence;

(2) adopt the conservative correlating method based on color characteristics to carry out data correlation to the detection target between two continuous frames in input video sequence, to obtain multiple reliable conservative sheets of pursuit path in short-term;

(3) according to pursuit path sheet in short-term and utilize the distribution of movement characteristic of same target and color similarity characteristic to build 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 movement locus of different target;

(4) utilize the positive and negative samples training building to practice Hough random forest, the leaf node of this Hough random forest kind has been stored classification statistical property and the orbiting motion characteristic of positive and negative samples collection;

(5) utilize time-domain constraints characteristic will follow the tracks of in short-term head and the tail interval difference in sheet and be less than threshold value T _pany two follow the tracks of in short-term that sheet is built into may associated track pair, all possible associated tracks are to forming the associated track pair set of possibility, by each may associated track to carrying out feature description, to generate the feature set that detects sample, and the feature set that detects sample is sent in Hough forest, to obtain leaf node corresponding to this feature set;

(6) obtain the classification statistical property of leaf node with movement statistics characteristic thereby obtain the right association probability of relevant track all association probabilities form association probability matrix;

(7) judge whether horizontal in association probability matrix, vertical two maximum elements are greater than a threshold value, if it is tentatively judge these two elements respectively corresponding track sheet belong to same track, then proceed to step (8), otherwise represent not belong to same track, then proceed to step (8);

(8) with Hungary Algorithm, two elements are judged again, finally to determine that the two corresponding track sheet belongs to same track; Only have the track sheet that Hungary Algorithm judgement belongs to same track and judgement belongs to same track in step (7) just to think real associated track pair;

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

(10) increase threshold value T _pvalue, and repeat above-mentioned steps (5) to (9), until can not regeneration can be associated track to.

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

(3-1) extract at random on pursuit path sheet in short-term two colouring information and positional informations that detect targets, for generating positive sample, the positive sample set of multiple positive compositions of sample;

(3-2) on the different a pair of sheets of pursuit path in short-term, extract respectively colouring information and a positional information that detects target, for generating negative sample, multiple negative samples form negative sample collection.

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

(4-1) generate according to positive and negative samples collection the feature set A={F that inputs training sample ₁=(x ₁, y ₁), F ₂=(x ₂, y ₂) ... F _n=(x _n, y _n), wherein y _i∈ 0,1, i=1,2...n, n represents the number of the concentrated sample of positive and negative samples, x _ithe proper vector of i input training sample, and x _i={ f _color,f _ent,f _motion, f _color,represent the color histogram similarity of i input training sample, f _ent,represent the local gray level entropy partial binary similarity of i input training sample, f _motionrepresent the motion excursion amount of i input training sample, and the center position of two detection targets that p1 and p2 randomly draw while representing respectively i input training sample structure in its place frame, p1 '=p2-v2 (t2-t1), p2 '=p1+v1 (t2-t1), the wherein speed of v1 and v2 two corresponding detection targets when corresponding i input training sample builds respectively, the time frame of t1 and t2 two corresponding detection targets when corresponding i input training sample builds respectively, y _iit is the classification scalar of i input training sample;

(4-2) utilize recurrence partitioning constantly to divide the feature set A of input training sample, to generate Hough random forest, until the leaf node of Hough random forest meets end condition, finally obtain classification statistical property and the movement statistics characteristic of leaf node.

Preferably, work as y _i, represent that this sample is negative sample, i.e. proper vector x at=0 o'clock _ifrom different track sheets, work as y _i, represent that this sample is positive sample, i.e. proper vector x at=1 o'clock _ifrom same track sheet.

Preferably, the end condition of leaf node is: the sample set quantity of preserving in (1) leaf node 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) degree of depth of Hough random forest is less than Second Threshold, and the size of this threshold value is determined by the feature set information of inputting training sample.

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

(4-2-1) random choose proper vector x _i={ f _color,f _ent, and random choose f therefrom _color,or f _entas node division threshold value, the f selecting _color,or f _entand proper vector x _iform parameter pond { τ _k, taking the node division threshold value selected as example, the training sample that is less than or equal to this node division threshold value is left child node, and the training sample that is greater than this node division threshold value is right child node, and 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) obtain classification uncertainty measure U ₁(S)=| S|H (Y) or orbiting motion skew consistance uncertainty measure wherein | S| represents to arrive the sample set S number of node k, S ₊represent positive sample set number in S, for motion excursion average corresponding to all positive samples in node k, H (Y) is the classification entropy of sample set S;

(4-2-3) select to make classification uncertainty measure U ₁or orbiting motion skew consistance uncertainty measure U (S) ₂(S) maximum optimized parameter τ _k*, so that two child nodes are divided front nodal point uncertainty decline maximum after dividing;

(4-2-4) at sample set S corresponding to left and right child node _l, S _ron basis, continue to divide left and right child node, and repeat above-mentioned steps (4-2-1) to (4-2-3) to obtain optimized parameter, until meet end condition, the terminator node finally obtaining is exactly leaf node;

(4-2-5) the classification statistical property of calculating leaf node with movement statistics characteristic it is respectively

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

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 adopt to obtain based on 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, the above technical scheme of conceiving by the present invention compared with prior art, can obtain following beneficial effect:

1, the present invention obtains training sample set effectively accurately by the conservative sheet of tracking in short-term, thereby effectively ensure the accuracy in tree learning process, in to the learning process of tree, classification information and movable information are taken into full account, this conforms to the foundation of trace model, by using Hungary Algorithm further to ensure reliable association results, adopt Increment Learning Algorithm to improve the study accuracy of tree, thereby solved the mistake existing in existing method with problem.

2, the method applied in the present invention can be for the demarcation of crowded street pedestrian's movement locus.Especially can alleviate similar phase close-target and easily occur the wrong bottleneck problem of following the tracks of of track identity, also have higher robustness with regard to track sheet disruption simultaneously.

3, provide effective guarantee to the more senior recognition performance of subsequent calculations machine vision system, automatic traffic management, the intellectuality accelerated based on pedestrian's tracking technique are had to very high practical value, concrete application can relate to hand over regulates the application scenarioss such as reason, robot navigation.。

4, the present invention is not only confined to pedestrian's tracking, and the movement objective orbit that is applicable to any known class is followed the tracks of application demand.

Brief description of the drawings

Fig. 1 the present invention is based on on-line study to follow the tracks of the process flow diagram of the method for multiple moving targets in video monitoring;

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

Fig. 3 is the invention process video frame images testing result used;

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

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

Fig. 6 is differentiation stage orbiting motion side-play amount definition schematic diagram;

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

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

Embodiment

In order to make object of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.In addition,, in each embodiment of described the present invention, involved technical characterictic just can combine mutually as long as do not form each other conflict.

Integral Thought of the present invention is, it comprises and moves frame by frame pedestrian detector, demarcate testing result, between two continuous frames, conservative association obtains reliably following the tracks of in short-term sheet, utilize space-time restriction, (feature set builds and comprises color histogram similarity to build positive and negative sample characteristics collection, histograms of oriented gradients similarity, partial binary proper vector similarity under local gray level maximum entropy principle, and be offset homogeneity measure based on the movement locus under movement locus continuously smooth assumed condition), based on Hough forest training under classification uncertainty and motion excursion consistance uncertainty measure, definition can associated track pair, formalization can the right association probability of associated track be outward appearance and motion associating posterior probability, under MAP criterion, obtain association results, adopt Hungary's allocation algorithm to correct associated error, on reliably longer track sheet, rebuild training sample again, upgrade forest leaf node statistical property, and using reliably longer track sheet as the associated flow process of new input iteration, until without any can associated track sheet.

As shown in Figure 1, the present invention is based on the on-line study method of multiple moving targets in video monitoring of following the tracks of comprises the following steps:

(1) receive input video sequence, utilize the pedestrian detector of off-line training and adopt the method for multiple dimensioned traversal search frame to demarcate the position of target in input video sequence;

(2) adopt the conservative correlating method based on color characteristics to carry out data correlation to the detection target between two continuous frames in input video sequence, to obtain multiple reliable conservative sheets of pursuit path in short-term;

(3) according to pursuit path sheet in short-term and utilize the distribution of movement characteristic of same target and color similarity characteristic to build 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 movement locus of different target; This step comprises following sub-step:

(3-1) extract at random on pursuit path sheet in short-term two colouring information and positional informations that detect targets, for generating positive sample, the positive sample set of multiple positive compositions of sample;

(3-2) on the different a pair of sheets of pursuit path in short-term, extract respectively colouring information and a positional information that detects target, for generating negative sample, multiple negative samples form negative sample collection;

(4) utilize the positive and negative samples training building to practice Hough random forest (Hough Forest), the leaf node of this Hough random forest kind has been stored classification statistical property and the orbiting motion characteristic of positive and negative samples collection; This step specifically comprises following sub-step:

(4-1) generate according to positive and negative samples collection the feature set A={F that inputs training sample ₁=(x ₁, y ₁), F ₂=(x ₂, y ₂) ... F _n=(x _n, y _n), wherein y _i∈ 0,1, i=1,2...n, n represents the number of the concentrated sample of positive and negative samples, x _ithe proper vector of i input training sample, and x _i={ f _color,f _ent,f _motion, f _color,represent the color histogram similarity of i input training sample, f _ent,represent the local gray level entropy partial binary similarity of i input training sample, f _motionrepresent the motion excursion amount of i input training sample, and as shown in Figure 5, the center position of two detection targets that p1 and p2 randomly draw while representing respectively i input training sample structure in its place frame, p1 '=p2-v2 (t2-t1), p2 '=p1+v1 (t2-t1), the wherein speed of v1 and v2 two corresponding detection targets when corresponding i input training sample builds respectively, the time frame of t1 and t2 two corresponding detection targets when corresponding i input training sample builds respectively, y _iit is the classification scalar of i input training sample; Work as y _i=0, represent that this sample is negative sample, i.e. proper vector x _ifrom different track sheets, work as y _i, represent that this sample is positive sample, i.e. proper vector x at=1 o'clock _ifrom same track sheet;

(4-2) utilize recurrence partitioning constantly to divide the feature set A of input training sample, to generate Hough random forest, until the leaf node of Hough random forest meets end condition, final classification statistical property and the movement statistics characteristic that obtains leaf node, wherein the end condition of leaf node is: the sample set quantity of preserving in (1) leaf node is less than first threshold, the size of this threshold value is by just, the category Properties of negative sample collection determines, if just, the similarity of negative sample collection is lower, this threshold value is larger, otherwise less, in the present embodiment, threshold value is 20, (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 (classification and movable information) of inputting training sample, the more complicated threshold value of information is larger, otherwise less, and in present embodiment, this threshold value is 15, the partiting step of sample set S that wherein arrives certain node k is as follows:

(4-2-1) random choose proper vector x _i={ f _color,f _ent, and random choose f therefrom _color,or f _entas node division threshold value, the f selecting _color,or f _entand proper vector x _iform parameter pond { τ _k, taking the node division threshold value selected as example, the training sample that is less than or equal to this node division threshold value is left child node, and the training sample that is greater than this node division threshold value is right child node, and 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) obtain classification uncertainty measure U ₁(S)=| S|H (Y) or orbiting motion skew consistance uncertainty measure wherein | S| represents to arrive the sample set S number of node k, S ₊represent positive sample set number in S, for motion excursion average corresponding to all positive samples in node k, H (Y) is the classification entropy of sample set S;

(4-2-3) select to make classification uncertainty measure U ₁or orbiting motion skew consistance uncertainty measure U (S) ₂(S) maximum optimized parameter τ _k*, so that two child nodes are divided front nodal point uncertainty decline maximum after dividing;

(4-2-4) at sample set S corresponding to left and right child node _l, S _ron basis, continue to divide left and right child node, and repeat above-mentioned steps (4-2-1) to (4-2-3) to obtain optimized parameter, until meet end condition, the terminator node finally obtaining is exactly leaf node;

(4-2-5) the classification statistical property of calculating leaf node with movement statistics characteristic it is respectively

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

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 adopt to obtain based on 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 in short-term head and the tail interval difference in sheet and be less than threshold value T _p(its value and the length of following the tracks of in short-term sheet are directly proportional, be 8 frames in the present embodiment) any two follow the tracks of in short-term that sheet is built into may associated track pair, all possible associated tracks are to forming the associated track pair set of possibility, by each may associated track to carrying out feature description, to generate the feature set that detects sample, and the feature set that detects sample is sent in Hough forest, to obtain leaf node corresponding to this feature set;

Carrying out feature describes and is specially: associated track on each track on extract respectively one and detect colouring information and the positional information (as shown in Figure 6) of target, be used for generating detection sample, multiple detection compositions of sample detect sample set, generate according to detecting sample set the feature set that detects sample, (4-1) is basic identical for its process and above-mentioned steps, and unique difference is there is no classification scalar y _i;

(6) according to the classification statistical property of the formula calculating leaf node in above step (4-2-5) with movement statistics characteristic thereby obtain the right association probability of relevant track be defined as: all association probabilities form association probability matrix.

(7) judge whether horizontal in association probability matrix, vertical two maximum elements are greater than a threshold value (being 0.5 in present embodiment), if it is tentatively judge these two elements respectively corresponding track sheet belong to same track, then proceed to step (8), otherwise represent not belong to same track, then proceed to step (8);

(8) with Hungary Algorithm, two elements are judged again, finally to determine that the two corresponding track sheet belongs to same track; Only have the track sheet that Hungary Algorithm judgement belongs to same track and judgement belongs to same track in step (7) just to think real associated track pair;

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

(10) increase threshold value T _pvalue (increasing in the present embodiment 10), and repeat above-mentioned steps (5) to (9), until can not regeneration can be associated track to.

Provide an instantiation below:

(1) receive input video sequence, utilize the pedestrian detector of off-line training and adopt the method for multiple dimensioned traversal search frame to demarcate the position of target in input video sequence, if Fig. 2 is certain frame video sequence, it is classical HOG+SVM pedestrian detection algorithm that this project adopts detecting device, and its testing result is with reference to figure 3.

(2) obtain frame by frame after testing result, adopt the conservative correlating method based on color characteristics to carry out data correlation to the detection target between two continuous frames in input video sequence, to obtain multiple reliable conservative sheets of pursuit path in short-term.Concrete steps comprise same testing result normalization size, extract testing result between adjacent two frames of color histogram proper vector and set up adjacent two frame incidence matrix, two detect color histogram similarity corresponding to target area characterizes associated confidence, also be that color similarity is the highest compared with other coupling combinations, and be greater than in the time domain of threshold value adjacent two and detect targets and be understood to from same track, and then can obtain pursuit path sheet more in short-term.

(3) according to pursuit path sheet in short-term and utilize the distribution of movement characteristic of same target and color similarity characteristic to build 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 movement locus of different target.Specific implementation is:

(3-1) extract at random on pursuit path sheet in short-term two colouring information and positional informations that detect targets, for generating positive sample, the positive sample set of multiple positive compositions of sample;

(3-2) on the different a pair of sheets of pursuit path in short-term, extract respectively colouring information and a positional information that detects target, for generating negative sample, multiple negative samples form negative sample collection;

(4) the concrete implementation step of positive and negative samples training white silk Hough random forest (Hough Forest) of utilizing (3) to build is as follows:

(4-1) first extract positive and negative sample characteristics collection, positive and negative sampling feature vectors is made up of three parts, detects the color histogram f in target respective pixel region _color,, the effective coverage partial binary descriptor f based on selecting under local gray level entropy principle _ent,(referring to Fig. 4), and same orbiting motion information (orbiting motion skew uncertainty) the center position of two detection targets that p1 and p2 randomly draw while representing respectively i input training sample structure in its place frame, p1 '=p2-v2 (t2-t1), p2 '=p1+v1 (t2-t1), the wherein speed of v1 and v2 two corresponding detection targets when corresponding i input training sample builds respectively, the time frame of t1 and t2 two corresponding detection targets when corresponding i input training sample builds respectively, schematic diagram is referring to Fig. 5.Obviously, positive sample is from same track two surveyed areas, and color histogram similarity should be higher, and effective coverage is selected scale-of-two descriptor and also answered similarity higher, same rail motion has higher flatness, and Fixed Time Interval bias internal distributes and should have very strong regularity.Negative sample collection is from different tracks two surveyed area pixel corresponding color histograms, select the partial binary of maximum region with local entropy and despise feature, obviously because from different targets, the more same track of similarity is low, and negative sample concentrates the motion excursion amount regularity of solution track not obvious.

(4-2) after above-mentioned processing, build a large amount of positive and negative sample characteristics Ji Ji, start to train Hough forestry below.Hough forest is cascade decision tree, and the core concept of learning algorithm is to find optimal dividing function for each leaf node that flies, by the sample set of this non-leaf node be divided into left and right can not two child nodes.Estimating of optimization function adopts respectively classification uncertainty and orbiting motion to be offset uncertain consistance, requires partition function as much as possible generic sample to be divided into same node, or positive sample similar motion excursion is gathered to same node.Concrete steps are as follows:

(4-2-1) random choose proper vector x _i={ f _color,f _ent, and random choose f therefrom _color,or f _entas node division threshold value, the f selecting _color,or f _entand proper vector x _iform parameter pond { τ _k, taking the node division threshold value selected as example, the training sample that is less than or equal to this node division threshold value is left child node, and the training sample that is greater than this node division threshold value is right child node, and sample set corresponding to left and right child node is S _l, S _r:

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

(4-2-3) select to make classification uncertainty measure U ₁or orbiting motion skew consistance uncertainty measure U (S) ₂(S) maximum optimized parameter τ _k*, so that two child nodes are divided front nodal point uncertainty decline maximum after dividing;

(4-2-4) at sample set S corresponding to left and right child node _l, S _ron basis, continue to divide left and right child node, and repeat above-mentioned steps (4-2-1) to (4-2-3) to obtain optimized parameter, until meet end condition, the terminator node finally obtaining is exactly leaf node;

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

So far, the study of Hough forest is complete.

(5) utilize time-domain constraints characteristic will follow the tracks of in short-term head and the tail interval difference in sheet and be less than threshold value T _pany two follow the tracks of in short-term that sheet is built into may associated track pair, all possible associated tracks are to forming the associated track pair set of possibility, by each may associated track to carrying out feature description, to generate the feature set that detects sample, and the feature set that detects sample is sent in Hough forest, to obtain leaf node corresponding to this feature set;

Detecting sample characteristics describes and is specially: first associated track on each track on extract respectively one and detect colouring information and the positional information (as shown in Figure 6) of target, be used for generating detection sample, multiple detection compositions of sample detect sample set, (4-1) method of employing obtains the characteristic descriptor set of sample set, and unique difference is there is no classification scalar y _i;

(6) the classification statistical property of the leaf node having obtained according to above step (4-2-5) with movement statistics characteristic calculate the right association probability of the relevant track of institute calmly all association probabilities form association probability matrix.

(7) judge whether horizontal in association probability matrix, vertical two maximum elements are greater than a threshold value (being 0.5 in present embodiment), if it is tentatively judge these two elements respectively corresponding track sheet belong to same track, otherwise represent not belong to same track, then proceed to step (8);

(8) with Hungary Algorithm, two elements are judged again, finally to determine that the two corresponding track sheet belongs to same track; Only have the track sheet that Hungary Algorithm judgement belongs to same track and judgement belongs to same track in step (7) just to think real associated track pair;

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

(10) increase threshold value T _pvalue (increasing in the present embodiment 10), and repeat above-mentioned steps (5) to (9), until can not regeneration can be associated track to.

Generally speaking, the invention discloses motion target tracking method in a kind of monitor video based on on-line study.Be applicable to all safety check Moving Object in Video Sequences and follow the tracks of, as pedestrian's tracing of the movement in user and streetscape scene, track of vehicle is demarcated etc.Multiple target tracking problem is converted into track sheet related question step by step under MAP problem by this patent.First utilize the particular category survey device of off-line training to detect target area in video sequence; Then in conjunction with appearance characteristics, adopt the target between associated adjacent two frames of the conservative associated thinking of dual threshold, obtain reliable conservative short tracking sheet; Again on the tracking sheet obtaining, utilize time-space domain distribution constraint information, define positive and negative sample set, extract respectively color, texture appearance characteristic similarity and movable information, as on-line study device training characteristics collection, by the training process of machine learning on-line learning algorithm, obtain the probabilistic statistical characteristics based on motion and appearance characteristics in the track sheet regularity of distribution; Finally two track sheet correlation forms are turned to the associating posterior probability greatest problem solving based on motion and outward appearance; The associated association input being embodied in last association results as next collection, rebuilds sample training collection step by step, and the probability nature of incremental learning renewal learning algorithm model, by increasing gradually track interval correlation time, obtains the track sheet of longer time.This construction framework based on on-line study has better adaptive ability compared with off-line training template, be more suitable for track sheet study instantly, but on-line study mechanism, because the training sample set of next stage is from the Output rusults of upper level, if upper level Output rusults has error, can produce the accumulation of error.For fear of this situation, further introduce the Hungary's allocation algorithm based on appearance characteristics, correct every grade of learning algorithm association results, greatly improve the robustness of algorithm, especially alleviated the track identity that under the similar crowd scene of outward appearance closely, multiple mobile object tracking problem occurs and demarcated frequent switching problem.This project on-line learning algorithm build mechanism can be conventional arbitrary learning algorithm: the boosting of computer vision neighborhood, SVM, decision tree etc.

Those skilled in the art will readily understand; the foregoing is only preferred embodiment of the present invention; not in order to limit the present invention, all any amendments of doing within the spirit and principles in the present invention, be equal to and replace and improvement etc., within all should being included in protection scope of the present invention.

Claims (6)

1. a method of following the tracks of multiple moving targets in video monitoring based on on-line study, is characterized in that, comprises the following steps:

(1) receive input video sequence, utilize the pedestrian detector of off-line training and adopt the method for multiple dimensioned traversal search frame to demarcate the position of target in input video sequence;

(2) adopt the conservative correlating method based on color characteristics to carry out data correlation to the detection target between two continuous frames in input video sequence, to obtain multiple reliable conservative sheets of pursuit path in short-term;

(3) according to pursuit path sheet in short-term and utilize the distribution of movement characteristic of same target and color similarity characteristic to build 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 movement locus of different target;

(4) utilize the positive and negative samples training building to practice Hough random forest, the leaf node of this Hough random forest kind has been stored classification statistical property and the orbiting motion characteristic of positive and negative samples collection;

(5) utilize time-domain constraints characteristic will follow the tracks of in short-term head and the tail interval difference in sheet and be less than threshold value T _pany two follow the tracks of in short-term that sheet is built into may associated track pair, all possible associated tracks are to forming the associated track pair set of possibility, by each may associated track to carrying out feature description, to generate the feature set that detects sample, and the feature set that detects sample is sent in Hough forest, to obtain leaf node corresponding to this feature set;

(6) obtain the classification statistical property of leaf node with movement statistics characteristic thereby obtain the right association probability of relevant track all association probabilities form association probability matrix;

(7) judge whether horizontal in association probability matrix, vertical two maximum elements are greater than a threshold value, if it is tentatively judge these two elements respectively corresponding track sheet belong to same track, then proceed to step (8), otherwise represent not belong to same track, then proceed to step (8);

(8) with Hungary Algorithm, two elements are judged again, finally to determine that the two corresponding track sheet belongs to same track; Only have the track sheet that Hungary Algorithm judgement belongs to same track and judgement belongs to same track in step (7) just to think real associated track pair;

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

(10) increase threshold value T _pvalue, and repeat above-mentioned steps (5) to (9), until can not regeneration can be associated track to.

2. method according to claim 1, is characterized in that, step (3) comprises following sub-step:

(3-1) extract at random on pursuit path sheet in short-term two colouring information and positional informations that detect targets, for generating positive sample, the positive sample set of multiple positive compositions of sample;

(3-2) on the different a pair of sheets of pursuit path in short-term, extract respectively colouring information and a positional information that detects target, for generating negative sample, multiple negative samples form negative sample collection.

3. method according to claim 2, is characterized in that, step (4) comprises following sub-step:

(4-1) generate according to positive and negative samples collection the feature set A={F that inputs training sample ₁=(x ₁, y ₁), F ₂=(x ₂, y ₂) ... F _n=(x _n, y _n), wherein y _i∈ 0,1, i=1,2...n, n represents the number of the concentrated sample of positive and negative samples, x _ithe proper vector of i input training sample, and x _i={ f _color,f _ent,f _motion, f _color,represent the color histogram similarity of i input training sample, f _ent,represent the local gray level entropy partial binary similarity of i input training sample, f _motionrepresent the motion excursion amount of i input training sample, and the center position of two detection targets that p1 and p2 randomly draw while representing respectively i input training sample structure in its place frame, p1 '=p2-v2 (t2-t1), p2 '=p1+v1 (t2-t1), the wherein speed of v1 and v2 two corresponding detection targets when corresponding i input training sample builds respectively, the time frame of t1 and t2 two corresponding detection targets when corresponding i input training sample builds respectively, y _iit is the classification scalar of i input training sample;

(4-2) utilize recurrence partitioning constantly to divide the feature set A of input training sample, to generate Hough random forest, until the leaf node of Hough random forest meets end condition, finally obtain classification statistical property and the movement statistics characteristic of leaf node.

4. method according to claim 3, is characterized in that, works as y _i, represent that this sample is negative sample, i.e. proper vector x at=0 o'clock _ifrom different track sheets, work as y _i, represent that this sample is positive sample, i.e. proper vector x at=1 o'clock _ifrom same track sheet.

5. method according to claim 3, is characterized in that, the end condition of leaf node is: the sample set quantity of preserving in (1) leaf node 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) degree of depth of Hough random forest is less than Second Threshold, and the size of this threshold value is determined by the feature set information of inputting training sample.

6. method according to claim 3, is characterized in that, the partiting step of sample set S that arrives certain node k is as follows:

(4-2-1) random choose proper vector x _i={ f _color,f _ent, and random choose f therefrom _color,or f _entas node division threshold value, the f selecting _color,or f _entand proper vector x _iform parameter pond { τ _k, taking the node division threshold value selected as example, the training sample that is less than or equal to this node division threshold value is left child node, and the training sample that is greater than this node division threshold value is right child node, and 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) obtain classification uncertainty measure U ₁(S)=| S|H (Y) or orbiting motion skew consistance uncertainty measure wherein | S| represents to arrive the sample set S number of node k, S ₊represent positive sample set number in S, for motion excursion average corresponding to all positive samples in node k, H (Y) is the classification entropy of sample set S;

(4-2-3) select to make classification uncertainty measure U ₁or orbiting motion skew consistance uncertainty measure U (S) ₂(S) maximum optimized parameter τ _k*, so that two child nodes are divided front nodal point uncertainty decline maximum after dividing;

(4-2-4) at sample set S corresponding to left and right child node _l, S _ron basis, continue to divide left and right child node, and repeat above-mentioned steps (4-2-1) to (4-2-3) to obtain optimized parameter, until meet end condition, the terminator node finally obtaining is exactly leaf node;

(4-2-5) the classification statistical property of calculating leaf node with movement statistics characteristic it is respectively

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

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 adopt to obtain based on 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.