CN103400391A - Multiple-target tracking method and device based on improved random forest - Google Patents
Multiple-target tracking method and device based on improved random forest Download PDFInfo
- Publication number
- CN103400391A CN103400391A CN2013103535117A CN201310353511A CN103400391A CN 103400391 A CN103400391 A CN 103400391A CN 2013103535117 A CN2013103535117 A CN 2013103535117A CN 201310353511 A CN201310353511 A CN 201310353511A CN 103400391 A CN103400391 A CN 103400391A
- Authority
- CN
- China
- Prior art keywords
- target
- random forest
- piece
- pieces
- sorter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000007637 random forest analysis Methods 0.000 title claims abstract description 80
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 39
- 241001270131 Agaricus moelleri Species 0.000 claims description 5
- 230000015572 biosynthetic process Effects 0.000 claims description 2
- 230000008569 process Effects 0.000 abstract description 11
- 238000004422 calculation algorithm Methods 0.000 description 10
- 239000012535 impurity Substances 0.000 description 9
- 238000003066 decision tree Methods 0.000 description 7
- 238000012544 monitoring process Methods 0.000 description 6
- 230000008901 benefit Effects 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 3
- 230000000875 corresponding effect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000033001 locomotion Effects 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000013138 pruning Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- IUVCFHHAEHNCFT-INIZCTEOSA-N 2-[(1s)-1-[4-amino-3-(3-fluoro-4-propan-2-yloxyphenyl)pyrazolo[3,4-d]pyrimidin-1-yl]ethyl]-6-fluoro-3-(3-fluorophenyl)chromen-4-one Chemical compound C1=C(F)C(OC(C)C)=CC=C1C(C1=C(N)N=CN=C11)=NN1[C@@H](C)C1=C(C=2C=C(F)C=CC=2)C(=O)C2=CC(F)=CC=C2O1 IUVCFHHAEHNCFT-INIZCTEOSA-N 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013079 data visualisation Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000009017 pursuit movement Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a multiple-target tracking method and device based on an improved random forest. The device comprises a random forest training module, a random forest classification module and a target reposition module, wherein the random forest training module is used for training and learning the target before intersection and building a target classifier; the random forest classification module is used for classifying the intersected target region through the trained classifier in the next frame; the target reposition module is used for conducting clustering operation to the classified block, so as to form the target position region. According to the method and device provided by the invention, problems of error tracking and miss tracking caused by intersection in a multiple-target tracking process, for example, the intersection of a plurality of targets in pedestrian tracking and vehicle tracking, are effectively solved.
Description
Technical field
The present invention relates to computer vision field and artificial intelligence field, especially relate to a kind of multi-object tracking method based on improved random forest and device.
Background technology
In recent years, development along with science and technology, the public safety video monitoring system is the powerful measure that urban society's public security is initiatively grasped and hit, as AT STATION, the ground such as harbour, airport, harbour, urban transportation thoroughfare and gateway sets up the public safety video monitoring system, bring into play the advantage of its modern technologies risk prevention instruments, to safeguarding society and politics and public security, stablize significant.
Traditional video monitoring system only is equivalent to the effect of a video acquisition browser, and the monitor staff need to carry out artificial observation for a long time, only with subjectivity, determine whether take emergency measures, and the supervisor seldom can keep notice for a long time.Intelligent video monitoring system obtains flourishly in recent years, and the application of a lot of maturations has been arranged, as to forbidden zone, swarmed into, cross over half line, leave over the detection of behaviors such as removing.In general the principle of work of intelligent video monitoring system is divided into several steps: at first in the video sequence of input, extract interested moving target, then moving target is followed the tracks of, by the pursuit movement target, identify its motion feature, finally, when moving target is made the certain user and thought abnormal motion feature, send warning.Can find out, target following is an important component part of intelligent video monitoring system.
an important directions of intelligent image analysis field is exactly multiple target tracking, the difficult point of multiple target tracking and emphasis are the later coupling orientation problems of a plurality of targets intersection, traditional method has meanshift, optical flow method etc., these methods can not effectively solve the problem of reorientating after target is intersected, there will be a large amount of matching errors, the problems such as track rejection, the present invention proposes a kind of multiple-target system based on random forest, after can effectively reducing the target intersection, target is with losing with wrong probability, the random forest sorting algorithm is fast because of its execution time, have application in a lot of fields recent years, but main application is Images Classification, image recognition etc., the present invention is applied to it in image tracking algorithm, effect is compared traditional algorithm obvious lifting.
In multiple-target system, a major issue that faces at present is exactly how the later target trajectory of a plurality of targets intersections is consistent, effectively address this problem is subsequent treatment such as recognition of face, the important prerequisite of people stream counting etc., the present invention is directed to the cross-cutting issue of multiple target tracking process, as after in pedestrian's tracking and vehicle tracking, existing a plurality of targets to intersect, occur following mistake with the problem of losing, proposed the solution that a kind of multiple target tracking based on improved random forest intersects.
Summary of the invention
In order to address the above problem, the invention provides a kind of device of multiple target tracking based on improved random forest, this device comprises: the random forest training module, for the target to before intersecting, carry out training study, set up the target classification device; The random forest sort module, classify to the target area after intersecting with the sorter that trains at next frame; Target is reorientated module, to sorted, carries out cluster operation, formation zone, target location.
further, described random forest training module is configured to: for intersecting the former frame target, the model random forest, the grey scale difference that random point is right is as feature, at each fritter of target, choose at random the gray scale of two pixels, then calculate the difference of these two gray scales, if greater than 0 enter right branch, otherwise be left branch, target A1 before intersecting, the piece that A2 is divided into respectively several prescribed level that overlap each other is stored in respectively Blocks1 and Blocks2, then at first for target A1 training classifier, Blocks1 is used as to positive sample, Blocks2 is used as to negative sample, use random forest to train and obtain the posterior probability that each leaf node belongs to current goal, posterior probability thinks that more current block more likely belongs to current goal, after all pieces in Blocks1 and Blocks2 are trained, generate the random forest sorter C1 of target A1, can train equally the sorter C2 of A2.
Further, it is 13 that described random forest training module uses the height of the tree of random forest, and the number of tree is 10.
further, the random forest sort module is configured to: at first use the same piece of block size used when training, zone after present frame intersects is scanned by the moving window mode, in each position of scanning, use sorter C1 and C2 to classify, namely each window every one deck by sorter C1 or C2, the final leaf node leaf that arrives every one tree, the class probability that generates when if the posterior probability of this leaf node leaf is trained greater than sorter C1, think that this piece belongs to target A1, if the class probability that generates while training greater than sorter C2, belong to target A2, the set that belongs to the piece of A1 after classification is dividedBlocks1, the set that belongs to the piece of A2 is dividedBlocks2.
Further, described target is reorientated module and is configured to: to dividedBlocks1
With
Piece in dividedBlocks2 is carried out respectively cluster operation, while being specially degree of overlapping when two pieces greater than threshold value, be merged into a piece, when without any degree of overlapping each other greater than the piece of threshold value the time, just directly get the boundary rectangle of these pieces, namely comprise the rectangular area of maximum of these pieces as target location; When all pieces all intersect, all pieces are merged and become a piece, this piece is exactly the band of position of target; When having at least two pieces to merge, get the boundary rectangle of these two pieces as the zone, target location.
The present invention also provides a kind of multi-object tracking method based on improved random forest, and the method comprises the following steps:
Step (1) is carried out training study for the target to before intersecting, and sets up the target classification device;
Step (2) is classified to the target area after intersecting with the sorter that trains at next frame;
Step (3) is carried out cluster operation to sorted, forms the zone, target location.
further, step (1) comprising: for intersecting the former frame target, the model random forest, the grey scale difference that random point is right is as feature, at each fritter of target, choose at random the gray scale of two pixels, then calculate the difference of these two gray scales, if greater than 0 enter right branch, otherwise be left branch, target A1 before intersecting, the piece that A2 is divided into respectively several prescribed level that overlap each other is stored in respectively Blocks1 and Blocks2, then at first for target A1 training classifier, Blocks1 is used as to positive sample, Blocks2 is used as to negative sample, use random forest to train and obtain the posterior probability that each leaf node belongs to current goal, posterior probability thinks that more current block more likely belongs to current goal, after all pieces in Blocks1 and Blocks2 are trained, generate the random forest sorter C1 of target A1, can train equally the sorter C2 of A2.
Further, it is 13 that described random forest training module uses the height of the tree of random forest, and the number of tree is 10.
further, step (2) comprising: at first use the same piece of block size used when training, zone after present frame intersects is scanned by the moving window mode, in each position of scanning, use sorter C1 and C2 to classify, namely each window every one deck by sorter C1 or C2, the final leaf node leaf that arrives every one tree, the class probability that generates when if the posterior probability of this leaf node leaf is trained greater than sorter C1, think that this piece belongs to target A1, if the class probability that generates while training greater than sorter C2, belong to target A2, the set that belongs to the piece of A1 after classification is dividedBlocks1, the set that belongs to the piece of A2 is dividedBlocks2.
Further, step (3) comprising: the piece in dividedBlocks1 and dividedBlocks2 is carried out respectively to cluster operation, while being specially degree of overlapping when two pieces greater than threshold value, be merged into a piece, when without any degree of overlapping each other greater than the piece of threshold value the time, just directly get the boundary rectangle of these pieces, namely comprise the rectangular area of maximum of these pieces as target location; When all pieces all intersect, all pieces are merged and become a piece, this piece is exactly the band of position of target; When having at least two pieces to merge, get the boundary rectangle of these two pieces as the zone, target location.
The accompanying drawing explanation
Fig. 1 is the schematic diagram of the decision tree of the random forest that illustrates.
Fig. 2 is the process flow diagram that the method according to this invention is shown.
Embodiment
For above-mentioned purpose of the present invention, feature and advantage are become apparent more, the present invention is further detailed explanation below in conjunction with the drawings and specific embodiments:
Because solve the essence of target cross-cutting issue, be exactly how after intersecting, to distinguish prechiasmal target, how we distinguish the most important thing that prechiasmal target is research like this, this problem and one of the computer vision research field important direction problem of studying of namely classifying is the same, so the present invention takes full advantage of the sorting algorithm random forest of comparative maturity, practical problems in conjunction with the multiple goal intersection runs into, organically be attached to sorting algorithm in multiple-target system.
Based on the multiple target tracking device of improved random forest, mainly three large modules, consisting of in logic is the random forest training module, the random forest sort module, and target is reorientated module.
1. at the random forest training module, the target before intersecting is carried out to training study, set up the target classification device.
2. after establishing, at next frame, in the random forest sort module, with the sorter that trains, classified in the target area after intersecting.
3. after the piece of target intersection region being classified, in target, reorientate module and according to the target under it, carry out cluster to sorted, form real target.
Compared with prior art, the present invention has following advantage:
The present invention changes into a target classification problem to the multiple target tracking cross-cutting issue, then uses and changes
The random forest sorting algorithm of advancing is classified, the surface characteristics of intersecting and separating again target in whole process due to multiple goal in actual applications can not have greatly changed, so use random forest can obtain good classifying quality, reduce greatly target simultaneously with wrong probability with losing.
Can process simultaneously the tracking problem of a large amount of targets, this is that general traditional algorithm is not available, and this all has benefited from the nature that a plurality of targets are classified that just possess of random forest itself, with and algorithm construction succinct, the execution time is fast.Can effectively with multiple nucleus system, be combined simultaneously, make its execution speed faster.The solution of intersecting based on the multiple target tracking of improved random forest mainly contains three large modules compositions in logic, i.e. random forest training module, and the random forest sort module, target is reorientated module
1. random forest training module
Random forest is a sorter that comprises a plurality of decision trees, and the classification of its output is to be determined by the mode of the classification of indivedual trees output.Leo Breiman and Adele Cutler develop the algorithm of reasoning out random forest.And " Random Forests " is their trade mark.This term Stochastic Decision-making forest (random decision forests) that to be nineteen ninety-five proposed by the Tin Kam Ho of Bell Laboratory and coming.This method is in conjunction with " the random subspace method " of Breimans " Bootstrap aggregating " idea and Ho " to build the set of decision tree.
The random forest method is a kind of assembled classification method, and the basic composition of the method is exactly decision tree.Decision tree is a kind of hierarchical structure that is comprised of node and vector.Decision tree has comprised three kinds of nodes: root node (for example A1), interior node (for example A2), destination node (or being called leaf node).Wherein, root node has and only has one, and it is the overall set of training data.Decision tree is such tree (as shown in Figure 1):
(1) on each internal node, select an attribute (for example a11, a12, a21, a22) to cut apart;
(2) corresponding property value of each bifurcated;
(3) each leafy node represents a classification (for example C1, C2, C3, C4).
The advantage of random forest method has:
For a variety of data, it can produce the sorter of pin-point accuracy.
It can process a large amount of input variables.
It can be when determining classification, the importance of assessment variable.
When building forest, it can the error after inside is for vague generalization produce the not estimation of deviation.
It comprises a good method can estimate the data of losing, and, if data loss is greatly arranged, still can remain accuracy.
It provides an experimental technique, can remove to detect variable interactions.
For unbalanced grouped data collection, it can balance error.
It calculates the degree of getting close in each example, for data mining, the detecting person of departing from (outlier) and the data visualization is very useful.
It can be extended and be applied on unlabelled data, and this class data typically uses Unsupervised Clustering.Also can detect the person of departing from and watch data.
Learning process is very fast.
The introducing of two randomnesss, make random forest be not easy to be absorbed in over-fitting.
The introducing of two randomnesss, make random forest have good noise resisting ability.
Adaptable to data set: can process discrete data, also can process continuous data, data set is without standardization.
It can process the data of very high-dimensional (feature is a lot), and it goes without doing feature selecting.
The training process of random forest is as follows:
1. original training set is N, and application bootstrap method is randomly drawed k new self-service sample set, and constructs thus k classification tree, and each is that the sample that is extracted into has formed k bag outer data (out of bag).
2. be provided with M variable, each Nodes of each tree randomly draw m variable (m<<M) (being that m is much smaller than M), then in m, select impurity level (impurity level namely refers to the Geordie impurity level, sample of choosing at random of its expression in subset by the possibility of misclassification.The Geordie impurity level multiply by it by the probability of misclassification for the selected probability of this sample.When in a node, all samples were all a class, the Geordie impurity level was zero.The popular impurity level of saying is exactly the larger of purity, and impurity level is less, and it is exactly the correct ability of distinguishing different subclasses of this feature that purity can be understood like this.) minimum variable, then look for the optimal threshold of this variable to carry out node split.
3. every tree all can complete growth and can beta pruning (Pruning).
When using random forest to classify to data, allow every tree in random forest classify to data, then the classification results of all trees ballot is obtained to final classification results.
the present invention proposes the method for an improved random forest, when traditional random forest is selected feature, select the feature of impurity level minimum to carry out the node growth, calculate each node impurity level and increased algorithm time complexity, be not suitable for the target tracking algorism that requirement of real-time is higher, the present invention obtains at the random gray-scale value of 2 got of each node the feature (namely carry out following operation at each Nodes: the difference of choosing at random the gray-scale value of two pixels in current image block is used as the feature of this point) that difference is used as this node, greater than 0 right child who thinks node, otherwise think left child, saved so to a great extent the training time of random forest, the phenomenon of over-fitting that while has been extracted two promises at random, owing to just getting at random the gray scale difference value of 2 as feature, fairly simple, all need the effect of more feature guarantee classification.
Because affect the principal element of random forest classification performance, there are two:
In forest, the classical strength of single tree is that the classical strength of every tree is larger, and the classification performance of random forest is better.
The degree of correlation in forest between tree: the degree of correlation between tree is larger, and the classification performance of random forest is poorer.In order to increase the classical strength this patent, the height of every tree has been increased to 13, has can be good at keeping classifying quality.Simultaneously in order to improve the degree of correlation between tree in forest, when random each number feature of generation, whether identically with the feature of former spanning tree check, if the same regenerate, can guarantee as much as possible that like this degree of correlation of setting between forest diminishes, thereby improve the classifying quality of random forest.
The present invention is applied to random forest to solve on the cross-cutting issue of multiple target tracking, for intersecting the former frame target, the model random forest, the present invention's grey scale difference that random point is right is as feature, at each fritter of target, choose at random the gray scale of two points, then calculating the difference of these two gray scales, if greater than 0 enter right branch, otherwise is left branch.The height of finding tree by many experiments is that the 13(characteristic number is 13), the number of tree is that 10 random forest is best to multiple goal cross-cutting issue effect, and it is 13 that the present invention uses the height of tree, and the number of tree is 10.as shown in Figure 2, before supposing to intersect, target is A1, A2, the piece that A1 and A2 is divided into respectively to several 10*10 sizes that overlap each other (at first generates the piece of a 10*10 size in the upper left corner of target, then use step-length be 1 to the right or downwards this piece of translation (specific as follows: first to the right at every turn a mobile step-length until to the target right margin, and then move down step-length 1, and then from the left margin target right margin that starts to move right, by that analogy, until move to the lower boundary of target), so just can generate the piece of a lot of 10*10 sizes), it is stored in respectively Blocks1 and Blocks2, then at first for target A1 training classifier, Blocks1 is used as to positive sample, Blocks2 is used as to negative sample, use random forest to train and obtain the posterior probability that each leaf node belongs to current goal, posterior probability thinks that more current block more likely belongs to current goal, after all pieces in Blocks1 and Blocks2 are trained, generate the random forest sorter C1 of target A1, in like manner can train the sorter C2 of A2.
2. in present frame, use sort module, at first use the same piece of block size 10*10 used when training, zone after intersection by moving window (this window size is 10*10) mode scan (concrete scan mode for first to the right at every turn a mobile step-length until to the target right margin, and then move down step-length 1, and then from the left margin target right margin that starts to move right, by that analogy, until move to the lower boundary of target), in each position of scanning, use sorter C1 and C2 to classify, namely each 10*10 window every one deck by sorter C1 or C2, the final leaf node leaf that arrives every one tree, the class probability that generates when if the posterior probability of this leaf is trained greater than sorter C1, think that this piece belongs to target A1, if the class probability that generates while training greater than sorter C2, belong to target A2, the set that belongs to the piece of A1 after classification is dividedBlocks1, the set that belongs to the piece of A2 is dividedBlocks2.
3. target is reorientated in module the piece in dividedBlocks1 and dividedBlocks2 is used respectively to cluster operation, the degree of overlapping that is specially when two pieces is merged into a piece greater than 0.3, A. when without any degree of overlapping each other greater than 0.3 piece the time, just directly get the boundary rectangle of these pieces, namely comprise the rectangular area of maximum of these pieces as target location; B. when all pieces all intersect, all pieces are merged and become a piece, this piece is exactly the band of position of target; C. when having at least two pieces to merge, get the boundary rectangle of these two pieces as the zone, target location.Piece after merging is respectively Box1, Box2, and target A1 is Box1 in the position of present frame, A2 is Box2 in position corresponding to present frame.
Be more than the detailed description that the preferred embodiments of the present invention are carried out, but those of ordinary skill in the art should be appreciated that within the scope of the present invention, and guided by the spirit, various improvement, interpolation and replacement are all possible.These are all in the protection domain that claim of the present invention limits.
Claims (10)
1. device of the multiple target tracking based on improved random forest is characterized in that this device comprises:
The random forest training module, carry out training study for the target to before intersecting, and sets up the target classification device;
The random forest sort module, classify to the target area after intersecting with the sorter that trains at next frame;
Target is reorientated module, to sorted, carries out cluster operation, formation zone, target location.
2. device according to claim 1 is characterized in that described random forest training module further is configured to:
for intersecting the former frame target, the model random forest, the grey scale difference that random point is right is as feature, at each fritter of target, choose at random the gray scale of two pixels, then calculate the difference of these two gray scales, if greater than 0 enter right branch, otherwise be left branch, target A1 before intersecting, the piece that A2 is divided into respectively several prescribed level that overlap each other is stored in respectively Blocks1 and Blocks2, then at first for target A1 training classifier, Blocks1 is used as to positive sample, Blocks2 is used as to negative sample, use random forest to train and obtain the posterior probability that each leaf node belongs to current goal, posterior probability thinks that more current block more likely belongs to current goal, after all pieces in Blocks1 and Blocks2 are trained, generate the random forest sorter C1 of target A1, can train equally the sorter C2 of A2.
3. device according to claim 2 is characterized in that:
It is 13 that described random forest training module uses the height of the tree of random forest, and the number of tree is 10.
4. device according to claim 1 is characterized in that the random forest sort module further is configured to:
at first use the same piece of block size used when training, zone after present frame intersects is scanned by the moving window mode, in each position of scanning, use sorter C1 and C2 to classify, namely each window every one deck by sorter C1 or C2, the final leaf node leaf that arrives every one tree, the class probability that generates when if the posterior probability of this leaf node leaf is trained greater than sorter C1, think that this piece belongs to target A1, if the class probability that generates while training greater than sorter C2, belong to target A2, the set that belongs to the piece of A1 after classification is dividedBlocks1, the set that belongs to the piece of A2 is dividedBlocks2.
5. device according to claim 1, described target is reorientated module and further is configured to:
Piece in dividedBlocks1 and dividedBlocks2 is carried out respectively to cluster operation, while being specially degree of overlapping when two pieces greater than threshold value, be merged into a piece, when without any degree of overlapping each other greater than the piece of threshold value the time, just directly get the boundary rectangle of these pieces, namely comprise the rectangular area of maximum of these pieces as target location; When all pieces all intersect, all pieces are merged and become a piece, this piece is exactly the band of position of target; When having at least two pieces to merge, get the boundary rectangle of these two pieces as the zone, target location.
6. multi-object tracking method based on improved random forest is characterized in that the method comprises the following steps:
Step (1) is carried out training study for the target to before intersecting, and sets up the target classification device;
Step (2) is classified to the target area after intersecting with the sorter that trains at next frame;
Step (3) is carried out cluster operation to sorted, forms the zone, target location.
7. method according to claim 6 is characterized in that step (1) further comprises:
for intersecting the former frame target, the model random forest, the grey scale difference that random point is right is as feature, at each fritter of target, choose at random the gray scale of two pixels, then calculate the difference of these two gray scales, if greater than 0 enter right branch, otherwise be left branch, target A1 before intersecting, the piece that A2 is divided into respectively several prescribed level that overlap each other is stored in respectively Blocks1 and Blocks2, then at first for target A1 training classifier, Blocks1 is used as to positive sample, Blocks2 is used as to negative sample, use random forest to train and obtain the posterior probability that each leaf node belongs to current goal, posterior probability thinks that more current block more likely belongs to current goal, after all pieces in Blocks1 and Blocks2 are trained, generate the random forest sorter C1 of target A1, can train equally the sorter C2 of A2.
8. method according to claim 7 is characterized in that:
It is 13 that described random forest training module uses the height of the tree of random forest, and the number of tree is 10.
9. method according to claim 6 is characterized in that step (2) further comprises:
at first use the same piece of block size used when training, zone after present frame intersects is scanned by the moving window mode, in each position of scanning, use sorter C1 and C2 to classify, namely each window every one deck by sorter C1 or C2, the final leaf node leaf that arrives every one tree, the class probability that generates when if the posterior probability of this leaf node leaf is trained greater than sorter C1, think that this piece belongs to target A1, if the class probability that generates while training greater than sorter C2, belong to target A2, the set that belongs to the piece of A1 after classification is dividedBlocks1, the set that belongs to the piece of A2 is dividedBlocks2.
10. method according to claim 6 is characterized in that step (3) further comprises:
Piece in dividedBlocks1 and dividedBlocks2 is carried out respectively to cluster operation, while being specially degree of overlapping when two pieces greater than threshold value, be merged into a piece, when without any degree of overlapping each other greater than the piece of threshold value the time, just directly get the boundary rectangle of these pieces, namely comprise the rectangular area of maximum of these pieces as target location; When all pieces all intersect, all pieces are merged and become a piece, this piece is exactly the band of position of target; When having at least two pieces to merge, get the boundary rectangle of these two pieces as the zone, target location.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310353511.7A CN103400391B (en) | 2013-08-09 | 2013-08-09 | A kind of multi-object tracking method based on the random forest improved and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310353511.7A CN103400391B (en) | 2013-08-09 | 2013-08-09 | A kind of multi-object tracking method based on the random forest improved and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103400391A true CN103400391A (en) | 2013-11-20 |
| CN103400391B CN103400391B (en) | 2016-08-10 |
Family
ID=49564003
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310353511.7A Expired - Fee Related CN103400391B (en) | 2013-08-09 | 2013-08-09 | A kind of multi-object tracking method based on the random forest improved and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103400391B (en) |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104299243A (en) * | 2014-09-28 | 2015-01-21 | 南京邮电大学 | Target tracking method based on Hough forests |
| CN105389589A (en) * | 2015-11-06 | 2016-03-09 | 北京航空航天大学 | Random-forest-regression-based rib detection method of chest X-ray film |
| CN105809707A (en) * | 2014-12-30 | 2016-07-27 | 江苏慧眼数据科技股份有限公司 | Pedestrian tracking method based on random forest algorithm |
| CN106295532A (en) * | 2016-08-01 | 2017-01-04 | 河海大学 | A kind of human motion recognition method in video image |
| CN106503728A (en) * | 2016-09-30 | 2017-03-15 | 深圳云天励飞技术有限公司 | A kind of image-recognizing method and device |
| CN106557760A (en) * | 2016-11-28 | 2017-04-05 | 江苏鸿信系统集成有限公司 | Monitoring system is filtered in a kind of image frame retrieval based on video identification technology |
| CN106846361A (en) * | 2016-12-16 | 2017-06-13 | 深圳大学 | Method for tracking target and device based on intuitionistic fuzzy random forest |
| CN106981043A (en) * | 2016-11-23 | 2017-07-25 | 四川大学 | A kind of high-precision three-dimensional information fast acquiring method based on random forest |
| CN107360552A (en) * | 2017-08-01 | 2017-11-17 | 电子科技大学 | A kind of indoor orientation method of multi-categorizer overall situation dynamic fusion |
| CN107423702A (en) * | 2017-07-20 | 2017-12-01 | 西安电子科技大学 | Video target tracking method based on TLD tracking systems |
| CN107786942A (en) * | 2017-09-30 | 2018-03-09 | 平安科技（深圳）有限公司 | Positioner, method and computer-readable recording medium based on wireless device |
| CN107944382A (en) * | 2017-11-20 | 2018-04-20 | 北京旷视科技有限公司 | Method for tracking target, device and electronic equipment |
| CN108399748A (en) * | 2018-03-08 | 2018-08-14 | 重庆邮电大学 | A kind of road travel time forecasting methods based on random forest and clustering algorithm |
| CN108563982A (en) * | 2018-01-05 | 2018-09-21 | 百度在线网络技术（北京）有限公司 | Method and apparatus for detection image |
| CN109697447A (en) * | 2017-10-20 | 2019-04-30 | 富士通株式会社 | Disaggregated model construction device, method and electronic equipment based on random forest |
| CN110765952A (en) * | 2019-10-24 | 2020-02-07 | 上海眼控科技股份有限公司 | Vehicle illegal video processing method and device and computer equipment |
| CN111385655A (en) * | 2018-12-29 | 2020-07-07 | 武汉斗鱼网络科技有限公司 | Advertisement bullet screen detection method and device, server and storage medium |
| WO2020183316A1 (en) * | 2019-03-12 | 2020-09-17 | International Business Machines Corporation | Deep forest model development and training |
| CN113392777A (en) * | 2021-06-17 | 2021-09-14 | 西安应用光学研究所 | Real-time target detection method based on online learning strategy |
| CN114170664A (en) * | 2021-12-11 | 2022-03-11 | 南京行者易智能交通科技有限公司 | Face image clustering method and device for performing link prediction based on self-attention mechanism |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102799900A (en) * | 2012-07-04 | 2012-11-28 | 西南交通大学 | Target tracking method based on supporting online clustering in detection |
| JP2012243196A (en) * | 2011-05-23 | 2012-12-10 | Morpho Inc | Image identification device, image identification method, image identification program and recording medium |
-
2013
- 2013-08-09 CN CN201310353511.7A patent/CN103400391B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2012243196A (en) * | 2011-05-23 | 2012-12-10 | Morpho Inc | Image identification device, image identification method, image identification program and recording medium |
| CN102799900A (en) * | 2012-07-04 | 2012-11-28 | 西南交通大学 | Target tracking method based on supporting online clustering in detection |
Non-Patent Citations (2)
| Title |
|---|
| 刘足华 等: "基于随机森林的目标检测和定位", 《计算机工程》 * |
| 王守超 等: "基于在线学习和结构约束的目标跟踪算法", 《计算机工程》 * |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104299243B (en) * | 2014-09-28 | 2017-02-08 | 南京邮电大学 | Target tracking method based on Hough forests |
| CN104299243A (en) * | 2014-09-28 | 2015-01-21 | 南京邮电大学 | Target tracking method based on Hough forests |
| CN105809707A (en) * | 2014-12-30 | 2016-07-27 | 江苏慧眼数据科技股份有限公司 | Pedestrian tracking method based on random forest algorithm |
| CN105809707B (en) * | 2014-12-30 | 2018-11-27 | 江苏慧眼数据科技股份有限公司 | A kind of pedestrian tracting method based on random forests algorithm |
| CN105389589B (en) * | 2015-11-06 | 2018-09-18 | 北京航空航天大学 | A kind of chest X ray piece rib cage detection method returned based on random forest |
| CN105389589A (en) * | 2015-11-06 | 2016-03-09 | 北京航空航天大学 | Random-forest-regression-based rib detection method of chest X-ray film |
| CN106295532A (en) * | 2016-08-01 | 2017-01-04 | 河海大学 | A kind of human motion recognition method in video image |
| CN106295532B (en) * | 2016-08-01 | 2019-09-24 | 河海大学 | A kind of human motion recognition method in video image |
| CN106503728A (en) * | 2016-09-30 | 2017-03-15 | 深圳云天励飞技术有限公司 | A kind of image-recognizing method and device |
| CN106981043A (en) * | 2016-11-23 | 2017-07-25 | 四川大学 | A kind of high-precision three-dimensional information fast acquiring method based on random forest |
| CN106557760A (en) * | 2016-11-28 | 2017-04-05 | 江苏鸿信系统集成有限公司 | Monitoring system is filtered in a kind of image frame retrieval based on video identification technology |
| CN106846361A (en) * | 2016-12-16 | 2017-06-13 | 深圳大学 | Method for tracking target and device based on intuitionistic fuzzy random forest |
| CN107423702B (en) * | 2017-07-20 | 2020-06-23 | 西安电子科技大学 | Video target tracking method based on TLD tracking system |
| CN107423702A (en) * | 2017-07-20 | 2017-12-01 | 西安电子科技大学 | Video target tracking method based on TLD tracking systems |
| CN107360552B (en) * | 2017-08-01 | 2020-02-07 | 电子科技大学 | Indoor positioning method for global dynamic fusion of multiple classifiers |
| CN107360552A (en) * | 2017-08-01 | 2017-11-17 | 电子科技大学 | A kind of indoor orientation method of multi-categorizer overall situation dynamic fusion |
| CN107786942A (en) * | 2017-09-30 | 2018-03-09 | 平安科技（深圳）有限公司 | Positioner, method and computer-readable recording medium based on wireless device |
| CN109697447A (en) * | 2017-10-20 | 2019-04-30 | 富士通株式会社 | Disaggregated model construction device, method and electronic equipment based on random forest |
| CN107944382A (en) * | 2017-11-20 | 2018-04-20 | 北京旷视科技有限公司 | Method for tracking target, device and electronic equipment |
| CN107944382B (en) * | 2017-11-20 | 2019-07-12 | 北京旷视科技有限公司 | Method for tracking target, device and electronic equipment |
| CN108563982B (en) * | 2018-01-05 | 2020-01-17 | 百度在线网络技术（北京）有限公司 | Method and apparatus for detecting image |
| CN108563982A (en) * | 2018-01-05 | 2018-09-21 | 百度在线网络技术（北京）有限公司 | Method and apparatus for detection image |
| CN108399748A (en) * | 2018-03-08 | 2018-08-14 | 重庆邮电大学 | A kind of road travel time forecasting methods based on random forest and clustering algorithm |
| CN111385655A (en) * | 2018-12-29 | 2020-07-07 | 武汉斗鱼网络科技有限公司 | Advertisement bullet screen detection method and device, server and storage medium |
| WO2020183316A1 (en) * | 2019-03-12 | 2020-09-17 | International Business Machines Corporation | Deep forest model development and training |
| US11893499B2 (en) | 2019-03-12 | 2024-02-06 | International Business Machines Corporation | Deep forest model development and training |
| CN110765952A (en) * | 2019-10-24 | 2020-02-07 | 上海眼控科技股份有限公司 | Vehicle illegal video processing method and device and computer equipment |
| CN113392777A (en) * | 2021-06-17 | 2021-09-14 | 西安应用光学研究所 | Real-time target detection method based on online learning strategy |
| CN114170664A (en) * | 2021-12-11 | 2022-03-11 | 南京行者易智能交通科技有限公司 | Face image clustering method and device for performing link prediction based on self-attention mechanism |
| CN114170664B (en) * | 2021-12-11 | 2024-08-27 | 南京行者易智能交通科技有限公司 | Face image clustering method and device for link prediction based on self-attention mechanism |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103400391B (en) | 2016-08-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103400391A (en) | Multiple-target tracking method and device based on improved random forest | |
| CN102842045B (en) | A kind of pedestrian detection method based on assemblage characteristic | |
| CN105512640B (en) | A kind of people flow rate statistical method based on video sequence | |
| CN104978570B (en) | The detection and recognition methods of traffic sign in driving video based on incremental learning | |
| CN110084151B (en) | Video abnormal behavior discrimination method based on non-local network deep learning | |
| CN107239762A (en) | Patronage statistical method in a kind of bus of view-based access control model | |
| CN109767427A (en) | The detection method of train rail fastener defect | |
| CN108921875A (en) | A kind of real-time traffic flow detection and method for tracing based on data of taking photo by plane | |
| CN107133974A (en) | The vehicle type classification method that Gaussian Background modeling is combined with Recognition with Recurrent Neural Network | |
| CN106778472A (en) | The common invader object detection and recognition method in transmission of electricity corridor based on deep learning | |
| CN111462488A (en) | Intersection safety risk assessment method based on deep convolutional neural network and intersection behavior characteristic model | |
| CN104112282A (en) | A method for tracking a plurality of moving objects in a monitor video based on on-line study | |
| CN110348437A (en) | It is a kind of based on Weakly supervised study with block the object detection method of perception | |
| CN102360434B (en) | Target classification method of vehicle and pedestrian in intelligent traffic monitoring | |
| CN106778570A (en) | A kind of pedestrian's real-time detection and tracking | |
| CN102902974A (en) | Image based method for identifying railway overhead-contact system bolt support identifying information | |
| CN104299243A (en) | Target tracking method based on Hough forests | |
| CN109993061A (en) | A kind of human face detection and tracing method, system and terminal device | |
| CN105930808A (en) | Moving object tracking method based on vector boosting template updating | |
| CN106127161A (en) | Fast target detection method based on cascade multilayer detector | |
| CN102708384B (en) | Bootstrapping weak learning method based on random fern and classifier thereof | |
| CN113450573A (en) | Traffic monitoring method and traffic monitoring system based on unmanned aerial vehicle image recognition | |
| CN104732248A (en) | Human body target detection method based on Omega shape features | |
| CN102024149B (en) | Method of object detection and training method of classifier in hierarchical object detector | |
| CN106710228A (en) | Implementation method of passenger and freight separated traffic parameter monitoring system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20160810 |