CN106920253A - It is a kind of based on the multi-object tracking method for blocking layering - Google Patents
It is a kind of based on the multi-object tracking method for blocking layering Download PDFInfo
- Publication number
- CN106920253A CN106920253A CN201710073835.3A CN201710073835A CN106920253A CN 106920253 A CN106920253 A CN 106920253A CN 201710073835 A CN201710073835 A CN 201710073835A CN 106920253 A CN106920253 A CN 106920253A
- Authority
- CN
- China
- Prior art keywords
- field pictures
- target
- state
- foreground area
- multiple targets
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30232—Surveillance
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on the multi-object tracking method for blocking layering, including:Build background scene, the hierarchical model that dividing processing obtains background scene is carried out in background scene, the morpheme parameter that background difference processing extracts foreground area is carried out to current frame image and background scene, determine in current frame image target association relation in foreground area and previous frame image, if a foreground area and a target association, then state of the target in current frame image is obtained with monotrack method, if a foreground area and multiple target associations, then state of multiple targets in current frame image is obtained with Markov chain Monte Carlo method, if foreground area not with any target association, the foreground area is then thought newly to enter target.The present invention is solved in multiple target tracking due to being blocked between target and tracking failure problem caused by target is blocked with background, it is adaptable to the intelligent video monitoring system of target following technology is depended under open environment.
Description
Technical field
The invention belongs to area of pattern recognition, more particularly, to a kind of based on the multi-object tracking method for blocking layering.
Background technology
Video frequency object tracking is always the focus and difficulties in computer vision research field, in order to retrieved
The movement locus of all targets in video sequence, and then for follow-up computer vision system higher level recognition performance provides effectively guarantor
Barrier, has practical value very high to accelerating the automatic traffic management based on target following technology, intellectuality.But multiple target
The development of tracking technique but encounters many restrictions, and wherein occlusion issue is that generally existing shows in multiple target tracking application
As, occlusion issue includes being blocked between mutually blocking between multiple targets and background and target, meanwhile, Multitarget Tracking
Also need to solve the uncertain problem of moving target number, these problems tracking object tape accurate to target tracking algorism carrys out very big shadow
Ring, or even the validity for having a strong impact on video monitoring system.
The content of the invention
For the disadvantages described above of prior art, the invention provides a kind of based on the multi-object tracking method for blocking layering,
Its object is to solve to cause the technical problem of multiple target tracking failure due to occlusion issue in the prior art.
To achieve the above object, it is the invention provides a kind of based on the multi-object tracking method for blocking layering including following
Step:
(1) background modeling treatment is carried out to preceding N two field pictures, background scene is obtained;
(2) dividing processing is carried out to background scene, the hierarchical model of background scene is obtained;
(3) background difference processing is carried out to N+1 two field pictures and background scene, each foreground zone in N+1 two field pictures is extracted
The morpheme parameter in domain, and according to each target in the morpheme parameter extraction N+1 two field pictures of each foreground area in N+1 two field pictures
Morpheme parameter and template characteristic, and morpheme parameter and the hierarchical model of background scene according to each target in N+1 two field pictures carry
Take in N+1 two field pictures the layer for blocking each target and background scene in hierarchical relationship and N+1 two field pictures between each target
Hierarchical relationship is blocked between secondary model;
(4) background difference processing is carried out to t two field pictures and background scene, each foreground area in acquisition t two field pictures
Morpheme parameter;
Morpheme parameter according to each target in the morpheme parameter and t-1 two field pictures of each foreground area in t two field pictures is obtained
The incidence relation of each foreground area and each target in t-1 two field pictures in t two field pictures;
(5) incidence relation according to each target in each foreground area in the t two field pictures and t-1 two field pictures judges the
The quantity of target associated by each foreground area in t two field pictures;
If a foreground area and a target association in t-1 two field pictures in t two field pictures, according to t-1 two field pictures
In the target morpheme parameter, t two field pictures in the morpheme parameter of the foreground area, the hierarchical model of background scene and the mesh
The feature templates being marked in t-1 two field pictures carry out monotrack treatment and obtain the shape of the target in t frame video images
State;
If a foreground area and multiple target associations in t-1 two field pictures in t two field pictures, according to t-1 two field pictures
The morpheme parameter of the foreground area in morpheme parameter, the t two field pictures of middle multiple targets, the hierarchical model of background scene and many
Feature templates of the individual target in t-1 two field pictures are carried out in Markov chain Monte Carlo treatment acquisition t two field pictures
The state of multiple target;
If a foreground area is not associated with any one target in t-1 two field pictures in t two field pictures, position is extracted
The state of each target of the foreground area in t two field pictures;
(6) using each target t two field pictures state as each target t two field pictures tracking result, and utilize each mesh
It is marked on feature templates in t-1 two field pictures of the state and each target of t two field pictures and updates each target in t two field pictures
Feature templates;
(7) whether judgment frame order t is equal to the totalframes amount M of video, if so, then terminating, otherwise, makes t=t+1, enters
Step (4);
Wherein, N+2≤t≤M, M are the totalframes amount of video.
The present invention provide based on the multi-object tracking method for blocking layering, for foreground area in current frame image with it is upper
The relation of each target in one two field picture, whether there is hiding relation and judges whether target is the new mesh for entering between judging target
Mark, the situation that there is hiding relation between target, mutually block many is obtained using the Markov Chain Monte Carlo method of sampling
Individual target, the situation between target in the absence of hiding relation, is obtained in the tracking result of present frame using monotrack method
The target, for the new target for entering present frame, obtains the state of target in the tracking result of present frame, as the target with
Track result.The present invention can solve the problem that and blocked between target and target, blocked between target and background scene and newly enter target
Multiple target tracking problem in the case of three kinds.
Preferably, step (4) includes following sub-step:
(41) background difference processing is carried out to t two field pictures and background scene, each foreground area in acquisition t two field pictures
Morpheme parameter;
(42) according to the morpheme ginseng of each target in the morpheme parameter and t-1 two field pictures of each foreground area in t two field pictures
Number judges the overlapping cases of each target in foreground area and t-1 two field pictures in t two field pictures;
If the overlapping region area and t-1 two field pictures of target should in foreground area and t-1 two field pictures in t two field pictures
The region area ratio of target more than 10%, then target association in the foreground area and t-1 frequency images in t two field pictures,
Otherwise the foreground area is not associated with the target in t-1 frequency images in t two field pictures;
(43) all targets in all foreground areas and t-1 two field pictures in traversal t two field pictures, obtain t two field pictures
In each foreground area and t-1 two field pictures each target incidence relation.
Preferably, step (5) if in a foreground area and a target association in t-1 two field pictures in t two field pictures
When, obtaining the state of the target in t two field pictures includes following sub-step:
(511) the morpheme parameter to the target in t-1 two field pictures carries out multiple stochastical sampling acquisition positioned at the t frame figures
Multiple particles as in foreground area, and morpheme parameter and the hierarchical model of background scene to each particle carries out random combine
Obtain each particle and be possible to barrier bed time relation with the hierarchical model of background scene;
(512) hierarchical model according to each particle and background scene blocks hierarchical relationship, extracts each particle at this
The de-occlusion region under hierarchical relationship is blocked, each is obtained in the de-occlusion region under blocking hierarchical relationship used in each particle
Particle is in the template characteristic under blocking hierarchical relationship;
(513) template characteristic according to each particle in the case where difference blocks hierarchical relationship and the target are in t-1 two field pictures
Template characteristic obtain the observation likelihood probability of each particle in the case where difference blocks hierarchical relationship, it is and maximum to observe likelihood probability
The morpheme parameter of particle as the target in t two field pictures morpheme parameter, to observe the maximum particle of likelihood probability and background
The hiding relation of the hierarchical model of scene is closed in t two field pictures as the target with the level that blocks of the hierarchical model of background scene
System, and by the target in t two field pictures morpheme parameter and the target in t two field pictures with the screening of the hierarchical model of background scene
Gear hierarchical relationship is used as the state of the target in t two field pictures.
The monotrack method provided in the present invention obtains single target tracking result in current frame image, by right
Background scene carries out dividing processing, before obtaining the hierarchical model of background scene, and the every two field picture of utilization background subtraction acquisition
Scene area and target, and morpheme parameter to target carries out stochastical sampling, and multiple grains are produced in current frame image foreground area
Son, by every two field picture between each particle and the hierarchical model of background scene hiding relation random combine, reasoning is each
The circumstance of occlusion of the hierarchical model of particle and background scene, and according to each particle and the circumstance of occlusion of the hierarchical model of background scene
The de-occlusion region of particle is extracted, for determining observation likelihood probability of the particle relative to the target, to observe likelihood probability
Maximum particle state is used as target tracking mode in the current frame.
Preferably, according to formulaUpdate i-th template of target in t two field pictures special
Levy, wherein,To utilize the template characteristic of i-th target of target de-occlusion region extraction, γ in t two field picturesiIt is i-th
The de-occlusion region of target accounts for i-th percentage of target,It is i-th template characteristic of target in t-1 two field pictures, 1≤
I≤L, L are the total quantity of target in t two field pictures.
Preferably, step (5) if in a foreground area and multiple target associations in t-1 two field pictures in t two field pictures
When, obtaining state of multiple targets in t two field pictures includes following sub-step:
(521) screening in the state acquisition t-1 two field pictures according to multiple targets in t-1 two field pictures between multiple targets
In gear hierarchical relationship and t-1 two field pictures hierarchical relationship is blocked between multiple targets and the hierarchical model of background scene;
According between multiple target in t-1 two field pictures block hierarchical relationship, t-1 two field pictures in multiple targets and the back of the body
The morpheme gain of parameter t for blocking multiple targets in hierarchical relationship and t-1 two field pictures between the hierarchical model of scape scene
The original state of multiple targets in two field picture;
(522) state parameter in state of the multiple targets after r-th renewal in t two field pictures is randomly updated,
Obtain state of the multiple targets after the r+1 renewal in t two field pictures;And the multiple targets after being updated according to the r+1
State in t two field pictures, update for r-th after state of multiple targets in t two field pictures and multiple targets the
The reception that feature templates in t-1 two field pictures obtain state of the multiple targets after the r+1 renewal in t two field pictures is general
Rate;
(523) whether state of the multiple targets after the r+1 renewal in t two field pictures is received according to probabilistic determination,
If so, state of the multiple targets after the r+1 renewal of record in t two field pictures, and enter step (524), otherwise, directly
Into step (524);
(524) judge to update order r and whether be equal to update total degree R, if so, then making r=r+1, and enter step
(522), otherwise into step (525);
(525) by the renewal of all receiving after state of multiple targets in t two field pictures in each parameter average,
Obtain mean state of multiple targets in t two field pictures;Using mean state of multiple targets in t two field pictures as multiple
State of the target in t two field pictures.
In the multi-object tracking method that the present invention is provided, foreground area and multiple mesh in previous frame image in current frame image
When mark has incidence relation, there is hiding relation in multiple targets, in present frame by the Markov Chain Monte Carlo method of sampling
Blocking between more fresh target and target block hierarchical relationship and each target morpheme parameter between hierarchical relationship, target and background
The state of multiple targets in current frame image after being updated, receives with multiple targets in the current frame image after new according to probability
State, the final optimum state for obtaining multiple targets in current frame image.
Preferably, according to formula in step (522)Obtain r+
1 update after state of multiple targets in t two field pictures the probability of acceptance;
Wherein,It is the observation likelihood probability of state of the multiple targets after r-th renewal in t two field pictures; Be the multiple targets after being updated at r-th in t two field pictures in the state of the
The j observation likelihood probability of target,It is state of the multiple targets after r-th renewal in t two field pictures
Motion model probability,Φ is to be desired forVariance
It is the probability function of the Gaussian Profile of Σ,It is that j-th pre- measured center of the minimum rectangle of target is surrounded in t two field pictures
Point coordinates,Be r-th update after multiple target t two field pictures in the state of surround j-th minimum rectangle of target
Center point coordinate;Be state update before and after suggestion distribution ratio, 1≤j≤Q, Q be with t two field pictures before
There is the target number of incidence relation in scene area.
In general, by the contemplated above technical scheme of the present invention compared with prior art, can obtain down and show
Beneficial effect:
1st, the present invention obtains background scene by background modeling, and extracts target according to background scene and each two field picture, no
Needs know target classification in advance, are not influenceed by weather illumination, and this is consistent with actual intelligent video monitoring application demand.
2nd, the present invention uses layering thought, and each mesh in the foreground area and each frame of each frame is extracted by background subtraction
Mark, and the hierarchical model that segmentation obtains background scene is carried out to background scene so that the present invention can effectively set up target and mesh
Hierarchical relationship is blocked between the hierarchical model for blocking hierarchical relationship and target and background scene between mark, is conducive to reasoning target
Details is blocked between target and between target and background, so as to thoroughly solve be asked because blocking the bottleneck for causing tracking to fail
Topic, substantially increases the robustness of track algorithm.
3rd, effective guarantee is provided to follow-up computer vision system higher level recognition performance, to accelerating to be based on multiple target tracking
The automatic traffic management of technology, intellectuality have practical value very high, and concrete application can be related to friendship and regulate reason, machine
The application scenarios such as people's navigation.
Brief description of the drawings
Fig. 1 is the flow chart based on the multi-object tracking method for blocking layering that the present invention is provided.
Specific embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as additionally, technical characteristic involved in invention described below each implementation method
Not constituting conflict each other can just be mutually combined.
The flow chart based on the multi-object tracking method for blocking layering that Fig. 1 is provided for the present invention, multi-object tracking method
Comprise the following steps:
(1) giving needs the sequence of video images of tracking, and background modeling treatment is carried out to preceding N two field pictures, obtains ambient field
Scape, background scene is built in the embodiment that the present invention is provided using mixed Gaussian background modeling method, and the value of N is complicated according to background
Degree determines that the complexity of background is determined from illumination variation and the degree of crowding, and background is more complicated, N values are bigger.
(2) dividing processing is carried out to background scene, the hierarchical model of background scene is obtained, in the embodiment that the present invention is provided
The simple linear Iterative Clustering (Simple Linear Iterative Clustering, SLIC) of use, obtains background
Each visual consistency region in scene, hierarchical relationship is blocked so as to subsequent builds target and background scene.
(3) background difference processing is carried out to background scene and N+1 two field pictures, extracts each foreground zone in N+1 two field pictures
The morpheme parameter in domain, using each target of morpheme parameter initialization of each foreground area in N+1 two field pictures in N+1 two field pictures
Morpheme parameter and template characteristic, and morpheme parameter and the hierarchical model of background scene according to each target in N+1 two field pictures carry
Take in N+1 two field pictures the layer for blocking each target and background scene in hierarchical relationship and N+1 two field pictures between each target
Hierarchical relationship is blocked between secondary model, is implemented as:
(31) background scene extracted using preceding N frames, foreground zone in N+1 two field pictures is detected using background subtraction
Domain, due to illumination, water wave, the interference such as leaf so that there is noise spot in the foreground area of detection, wave filter technology is utilized for this
Filter noise spot;Foreground area cavitation is directed to, with the further filling cavity of morphological dilations corrosion principle, side is smoothed
Edge, purifies foreground area, and the close foreground area of combined distance, extracts foreground area, finally with encirclement foreground area
Minimum rectangle center point coordinate and length and width as foreground area in N+1 two field pictures morpheme parameter.
(32) the morpheme ginseng with each target of morpheme gain of parameter of foreground area in N+1 two field pictures in N+1 two field pictures
Number, i.e., by the use of the center point coordinate and length and width of the minimum matrix for surrounding each target as each target N+1 two field pictures shape
Position parameter.
(33) the triple channel color histogram and HOG textural characteristics of each target are extracted as the target in N+1 two field pictures
In external appearance characteristic, extract motion feature of the Optical-flow Feature of each target as the target in N+1 two field pictures, use the target
Template characteristic as the target in N+1 two field pictures of external appearance characteristic and motion feature, and according to each mesh in N+1 two field pictures
The hierarchical model of target morpheme parameter and background scene extract in N+1 two field pictures between each target block hierarchical relationship and
In N+1 two field pictures hierarchical relationship is blocked between each target and the hierarchical model of background scene.
(4) tracked since N+2 two field pictures, the background scene extracted using preceding N frames is detected using background subtraction
The morpheme parameter of the foreground area in t two field pictures, wherein, t >=N+2.
Morpheme parameter according to each target in the morpheme parameter and t-1 two field pictures of each foreground area in t two field pictures is obtained
The incidence relation of each foreground area and each target in t-1 two field pictures in t two field pictures.It is implemented as:
(41) shape of the foreground area in t two field pictures is measured using background subtraction go-on-go to t two field pictures and background scene
Position parameter.
(42) according to the shape of certain target in the morpheme parameter and t-1 two field pictures of a certain foreground area in t two field pictures
In the parameter determination t two field pictures of position in the foreground area and t-1 two field pictures the target overlapping region.
If the overlapping region area and t-1 frame figures of the target in the foreground area of t two field pictures and t-1 two field pictures
The region area ratio of the target more than 10%, then with the target in t-1 frequency images deposit in t two field pictures by foreground area as in
In incidence relation, the foreground area does not exist incidence relation with the target in t-1 frequency images otherwise in t two field pictures.
(43) all targets in all foreground areas and t-1 two field pictures in traversal t two field pictures, obtain t two field pictures
In each foreground area and t-1 two field pictures each target incidence relation.
The incidence relation of each foreground area in t two field pictures and each target of t-1 two field pictures is divided into three groups, respectively
It is:
In t-1 two field pictures only have a target it is relevant with a foreground area in t two field pictures, now we
Think that the target is not blocked with other targets;
Only multiple targets are relevant with a foreground area in t two field pictures in t-1 two field pictures, now we
There is situation about blocking mutually between thinking these targets;
No one of t-1 two field pictures target is relevant with a foreground area in t two field pictures, now we
The foreground area is the target for newly entering scene in thinking t two field pictures.
(5) incidence relation according to each target in each foreground area in the t two field pictures and t-1 two field pictures judges the
The quantity of target associated by each foreground area in t two field pictures;
(51) if an only target is relevant with a foreground area in t two field pictures in t-1 two field pictures, this
When it is considered that the target is not blocked with other targets, the target is obtained in t two field pictures using monotrack method
In state, be implemented as:
(511) the morpheme parameter with the target in t-1 two field pictures is as initial value, to the target in t-1 two field pictures
Morpheme parameter carry out stochastical sampling, randomly generate multiple grains with the related foreground area of the target in the t two field pictures
Son, and morpheme parameter and the hierarchical model of background scene according to each particle carry out random combine and obtain each particle and background scene
Hierarchical model be possible to barrier bed time relation.
Particularly, multiple particles are produced as follows:
Centered on the center point coordinate of minimum rectangle of the target in the morpheme parameter of t-1 two field pictures, with Gauss
Perturbation scheme, produces the center point coordinate of new minimum rectangle around centre, and with the target t-1 two field pictures shape
The length of the minimum rectangle in the parameter of position and a width of center, produce the length and width of new minimum rectangle in the way of Gauss disturbance, from
And obtain particle, and when particle is randomly generated require particle be located at t two field pictures in the related foreground area of the target
In, 1000 particles are randomly generated in the present embodiment.
(512) hierarchical model according to each particle and background scene blocks hierarchical relationship, extracts each particle at this
The de-occlusion region under hierarchical relationship is blocked, threeway is extracted in the de-occlusion region under blocking hierarchical relationship with each particle
Road color histogram and HOG textural characteristics, in the external appearance characteristic under blocking hierarchical relationship, are existed as the particle with each particle
It is special in the motion under blocking hierarchical relationship as the particle that Optical-flow Feature is extracted in the de-occlusion region blocked under hierarchical relationship
Levy, with the particle in the external appearance characteristic under blocking hierarchical relationship and in the motion feature under blocking hierarchical relationship as the grain
Son is in the template characteristic under blocking hierarchical relationship.
(513) according to each particle the difference with the hierarchical model of background scene block template characteristic under hierarchical relationship with
Template characteristic of the target in t-1 two field pictures obtains each particle and blocks level in the difference of the hierarchical model with background scene
Observation likelihood probability under relation, and to observe the morpheme parameter of the maximum particle of likelihood probability as the mesh in t two field pictures
Target morpheme parameter, to observe the particle of likelihood probability maximum with the hiding relation of the hierarchical model of background scene as the target
With the hierarchical relationship that blocks of the hierarchical model of background scene in t two field pictures, and by the target in t two field pictures morpheme parameter
With the target hierarchical relationship is blocked in t two field pictures as the target in t two field pictures with the hierarchical model of background scene
State.
Obtain the observation likelihood probability tool that each particle is blocked under hierarchical relationship in the difference with the hierarchical model of background scene
Body is embodied as:Target osTemplate characteristic in t-1 two field picturesWherein, msIt is the target in t-1
Template characteristic in two field picture,It is the external appearance characteristic of the target in t-1 two field pictures,It is the target in t-1 frames
Motion feature in image, hierarchical relationship is blocked according to k-th particle and the hierarchical model of background scene, is extracted in this and is blocked
The de-occlusion region of k-th particle under hierarchical relationship, used in this block hierarchical relationship under k-th particleDe-occlusion regionExtract k-th particleIn the external appearance characteristic under blocking hierarchical relationship and in the motion feature under blocking hierarchical relationship
As k-th particleIn the template characteristic under blocking hierarchical relationship,Wherein, 1≤k≤W, W are grain
Sub- number,Hierarchical relationship lower template feature is blocked at this for k-th particle,It is to utilize de-occlusion regionExtract
The external appearance characteristic of the particle,It is to utilize de-occlusion regionThe motion feature of the particle for extracting, then k-th particleCompared with
The observation likelihood probability of the target is defined asHere ρ is that can be chosen for Pasteur's vector distance formula.
(52) if there is multiple targets relevant with certain foreground area in t two field pictures in t-1 two field pictures, it is believed that this
There is mutual hiding relation between a little targets, multiobjective optimization state, tool are obtained using Markov Chain Monte Carlo sampling algorithm
Body is embodied as:
(521) object set that note is associated with a certain foreground area simultaneously is O={ oq, q=1,2,3 ... Q, Q be with together
The quantity of the associated target of one foreground area, first with object set O={ oqIn screening of each target in t-1 two field pictures
Gear hierarchical relationship, object set O={ oqIn each target block level with the hierarchical model of background scene in t-1 two field pictures
Relation and object set O={ oqIn each target morpheme parameter in t-1 two field pictures define each target in t two field pictures
Original state, i.e. Markov Chain Monte Carlo sampling original state.
Each target united state in t two field pictures is Ωt={ Xt,Gt, wherein, XtTo surround object set O={ oqIn
The center point coordinate of the minimum rectangle of each target and the length and width of minimum rectangle, i.e., the morpheme parameter of each target, GtIt is object set O
={ oqIn block hierarchical relationship and object set O={ o between each targetqIn each target and background scene hierarchical model
Hiding relation.
(522) state parameter in the state of t two field pictures of the multiple targets after r-th renewal is randomly updated, is obtained
Obtain state of the multiple targets after the r+1 renewal in t two field pictures;And the multiple targets after being updated according to the r+1 exist
State in t two field pictures, update for r-th after state and multiple target of multiple targets in t two field pictures in t-1
Template characteristic obtains the probability of acceptance of state of the multiple targets after the r+1 renewal in t two field pictures in two field picture.R's
Initial value is 1, and the multiple targets after the 1st renewal are that multiple targets are initial in t two field pictures in the state of t two field pictures
State.
Randomly update a state parameter in state of the multiple targets after r-th renewal in t two field pictures, including with
Lower three kinds:
Hierarchical relationship renewal is blocked in state of multiple targets in t two field pictures after updating for r-th between target:
Two targets are selected in object set O at random, and change the hiding relation between them, according to probabilityRandomly select pth target op
With q-th target oq, wherein, op,oq∈ O, q=1,2,3 ... Q, p=1,2,3 ... Q update pth target opWith q-th target oq
Occlusion state.If pth target o in the united state of the multiple target t two field pictures after being updated at r-thpBy q-th target oq
Block, after renewal, pth target o in state of multiple targets in t two field pictures after being updated at the r+1pBlock q-th
Target oqIf, pth target o in state of multiple targets in t two field pictures after being updated at r-thpBlock q-th target oq,
After renewal, pth target o in state of multiple targets in t two field pictures after being updated at the r+1pBy q-th target oqHide
Gear.
Target and background scene blocks level pass in state of multiple targets in t two field pictures after updating for r-th
System updates:Certain target is selected at random in object set O, if the target is handed over the hierarchical model domain of the existence of background scene
It is folded, then change the hiding relation of the target and the hierarchical model of background scene, according to probabilityTarget is randomly selected, the target is updated
Hierarchical states are blocked with the hierarchical model of background scene, if shape of the multiple targets after being updated at r-th in t two field pictures
The target is blocked by the hierarchical model of background scene in state, and after renewal, multiple targets are in t two field pictures after being updated at the r+1
In state in the target do not blocked by the hierarchical model of background scene, if r-th update after multiple targets in t frame figures
The target is not blocked by the hierarchical model of background scene in state as in, and after renewal, multiple targets exist after being updated at the r+1
The target is blocked by the hierarchical model of background scene in state in t two field pictures.
The morpheme parameter of each target updates in state of multiple targets in t two field pictures after updating for r-th:At random more
After the morpheme gain of parameter the r+1 of certain target updates in state of multiple targets in t two field pictures after updating for new r-th
State of multiple targets in t two field pictures in the target morpheme parameter, also cry according to suggestion distributed update target morpheme
Parameter.Certain target is selected according to probability in object set O, according to probabilityTarget is randomly selected, the morpheme parameter of the target is updated,
Update the size of the minimum rectangle for including the minimum rectangle center point coordinate of the target and surrounding the target.We divide suggestion
Cloth model is set to Gauss disturbance model:If certain target in state of the multiple targets after updating for r-th in t two field pictures
Morpheme parameter isAfter renewal, the r+1 update after state of multiple targets in t two field pictures in the target morpheme
Parameter isWherein, ε is four-dimensional zero-mean gaussian distribution, i.e. ε~N (0, Σ), covariance Σ are diagonal matrix,
Element is respectively (20,20,10,10) on diagonal, and the center two-dimensional coordinate of the minimum rectangle for surrounding the target is corresponded to respectively
Variance, the variance for surrounding the length of the minimum rectangle of the target, the variance wide of the minimum rectangle for surrounding the target.
State of multiple targets in t two field pictures after being updated according to the r+1 and the multiple targets after r-th renewal
State in t two field pictures obtains the probability of acceptance of state of the multiple targets after the r+1 renewal in t two field pictures.
In particular, according to formulaAfter determining the r+1 renewal
State of multiple targets in t two field pictures the probability of acceptance.
Wherein,It is the observation likelihood of state of the multiple targets after r-th renewal in t two field pictures; Be r-th update after multiple targets in t two field pictures in the state of jth
The observation likelihood probability of individual target, 1≤j≤Q, Q are the target that there is incidence relation with same foreground area in t two field pictures
Number, the multiple targets after r-th renewal are obtained in t according to state of the multiple targets after r-th renewal in t two field pictures
J-th de-occlusion region of target in the state of in two field picture, the multiple targets after being updated according to r-th are in t two field pictures
In the state of j-th target de-occlusion region obtain the multiple targets after updating for r-th in the t two field pictures in the state of the
The j template characteristic of targetWherein,Multiple targets after being updated for r-th are in t two field pictures
J-th external appearance characteristic of target under state,Multiple targets after being updated for r-th in the t two field pictures in the state of jth
The motion feature of individual target, the template characteristic of j-th target in t-1 two field pictures isWherein,For
The external appearance characteristic of j-th target in t-1 two field pictures,The motion feature for being j-th target in t-1 two field pictures,
Multiple targets after then updating for r-th in the t two field pictures in the state of j-th target with respect to j-th mesh in t-1 two field pictures
Target observation likelihood is defined asHere ρ is that can be chosen for Pasteur's vector distance formula.
It is the motion model probability of state of the multiple targets after r-th renewal in t two field pictures,If each target motion follows constant velocity linear's motion model, j-th target exists in object set O
Predicted position in t two field pictures is It is that j-th minimum rectangle of target is surrounded in t two field pictures
Prediction center point coordinate,It is that j-th center point coordinate of the minimum rectangle of target is surrounded in t-1 two field pictures,For
J-th speed of target in t-1 two field pictures, position according to j-th target in current frame image with previous frame image
Position obtain.System modeling be withTo expect, variance is the Gaussian Profile of Σ, then
Φ is to be desired forVariance is the probability function of the Gaussian Profile of Σ,Multiple target t frame figures after being updated for r-th
J-th center point coordinate of the minimum rectangle of target is surrounded in the state of as in, variance Σ is set to (20,20) in the present embodiment
Diagonal matrix.
Suggestion before and after being updated for state is distributed ratio.Because suggestion is distributed as Gauss model, due to Gauss
The symmetry of model,It is constantly equal to 1.
(523) three kinds of state parameter more new strategies are performed at random, i.e., occlusion state renewal, target and background scene between target
Between occlusion state update or each target morpheme parameter update, so as to produce new state, whether receive r-th according to probabilistic determination
The state of multiple targets after renewal in t two field pictures, specifically refers to, and is produced at random with non-uniform probability on [0,1] interval
A raw random number β, if the r+1 update after state of multiple targets in t two field pictures acceptance probability α more than with
Machine number β, then record state of the multiple targets after the r+1 renewal in t two field pictures, is not connect otherwise into step (524)
Multiple targets after being updated by the r+1 are directly entered step (524) in the state of t two field pictures;
(524) judge to update order r and whether be equal to update total degree R, if so, then making r=r+1, and enter step
(522), otherwise into step (525);Wherein, update total degree R and typically take 103And the above order of magnitude.
(525) by the renewal of all receiving after state of multiple targets in t two field pictures in each state parameter ask equal
Value, obtains mean state of multiple targets in t two field pictures;Using mean state of multiple targets in t two field pictures as
The state of multiple targets in t two field pictures.
In the embodiment that the present invention is provided, the renewal of each lower 5000 receiving of Markov Chain Monte Carlo sample record
State of the multiple targets afterwards in t two field pictures, is designated asThe t after the renewal of preceding 1000 receiving is abandoned first
The state of multiple targets in two field picture, then multiple targets after remaining 4000 renewals of receiving are in t two field pictures
State, for the unit sampling interval, retains state of the multiple targets after 400 renewals of receiving in t two field pictures with 100,
Each parameter in state of the multiple targets after 400 renewals of receiving in t two field pictures is averaged, as multiple targets
State.
(53) if no one of t-1 two field pictures target is relevant with certain foreground area in t two field pictures, should
Target is newly, into the target of scene, the initial morpheme parameter of fresh target directly to be initialized with foreground area in foreground area, and
Initial morpheme parameter and fresh target according to fresh target obtain the layer of fresh target and background scene with the hierarchical model of background scene
Block hierarchical relationship between secondary model, with the hierarchical model of the initial morpheme parameter and fresh target of fresh target and background scene it
Between block state of the hierarchical relationship as fresh target.
(6) using each target t two field pictures state as each target t two field pictures tracking result, and utilize each mesh
It is marked on the state and each target of t the two field pictures feature templates in t-1 two field pictures and updates each target in t two field pictures
Feature templates.
Specifically, according to formulaMore template characteristic of the fresh target in t two field pictures, its
In, 1≤i≤L, L are the total quantity of target in t two field pictures, It is i-th target in t frame figures
Using i-th template characteristic of target de-occlusion region extraction as in,For i-th target utilizes i-th in t two field pictures
The external appearance characteristic that individual target de-occlusion region is extracted,For i-th target utilizes i-th non-screening of target in t two field pictures
Keep off the motion feature of extracted region, γiFor the de-occlusion region of i-th target accounts for i-th percentage of target, The template characteristic for being i-th target in t-1 two field pictures,For i-th target exists
External appearance characteristic in t-1 two field pictures,The motion feature for being i-th target in t-1 two field pictures.
(7) whether judgment frame order t is equal to the totalframes amount M of video, if so, then terminating, otherwise, makes t=t+1, enters
Step (4).
Sum it up, the present invention provide based on the multi-object tracking method for blocking layering, it is adaptable to all intelligent videos
Pedestrian movement's track following in object detecting and tracking in monitoring system, such as user and streetscape scene, track of vehicle demarcation etc.
Deng.The present invention can effectively solve the problems, such as multiple target tracking in the complex scene for exist occlusion issue.First with background modeling skill
Art builds background scene;Then the method for carrying out background difference processing using present frame and background scene detects foreground area;
Dividing processing is carried out to background scene, the hierarchical model of background scene is obtained, the layer of foreground area and background scene is being isolated
On secondary model basis, between description target and target, the hiding relation between target and background scene.With each prospect of present frame
Whether incidence relation between region and each target of previous frame determines between target with the presence or absence of hiding relation and determines target
It is the new target for entering scene.For not existing hiding relation between target and target, the mesh is obtained using monotrack method
It is marked on the state of present frame.For there is hiding relation between target and target, with the centre bit of the minimum rectangle for surrounding target
The length for putting, surrounding the minimum rectangle of target describes the morpheme parameter of target with wide, and level is blocked between target and target
The morpheme parameter for blocking hierarchical relationship and multiple target between relation, target and background scene as multiple targets shape
State, the unshielding part of shelter target is obtained using the hiding relation between target and target, target and background scene, using mesh
The combination of the external appearance characteristic and velocity characteristic of mark unshielding part describes shelter target, finally using Markov chain Meng Teka
Sieve carrys out in reasoning scene the optimum state of the multiple targets mutually blocked.Involved background modeling in the present invention, super-pixel point
Cut, various features extract expression, template renewal strategy is applied to computer vision neighborhood algorithms most in use.
As it will be easily appreciated by one skilled in the art that the foregoing is only presently preferred embodiments of the present invention, it is not used to
The limitation present invention, all any modification, equivalent and improvement made within the spirit and principles in the present invention etc., all should include
Within protection scope of the present invention.
Claims (6)
1. it is a kind of based on the multi-object tracking method for blocking layering, it is characterised in that to comprise the following steps:
(1) background modeling treatment is carried out to preceding N two field pictures, background scene is obtained;
(2) dividing processing is carried out to the background scene, the hierarchical model of background scene is obtained;
(3) background difference processing is carried out to N+1 two field pictures and the background scene, each foreground zone in N+1 two field pictures is extracted
The morpheme parameter in domain, and according to each mesh in the morpheme parameter extraction N+1 two field pictures of each foreground area in the N+1 two field pictures
Target morpheme parameter and template characteristic, and according to the morpheme parameter and the background scene of each target in the N+1 two field pictures
Hierarchical model obtain in N+1 two field pictures between each target block in hierarchical relationship and N+1 two field pictures each target with
Hierarchical relationship is blocked between the hierarchical model of background scene;
(4) background difference processing is carried out to t two field pictures and the background scene, each foreground area in acquisition t two field pictures
Morpheme parameter;
The morpheme parameter of each target is obtained in morpheme parameter and t-1 two field pictures according to each foreground area in the t two field pictures
The incidence relation of each foreground area and each target in t-1 two field pictures in t two field pictures;
(5) incidence relation according to each target in each foreground area in the t two field pictures and t-1 two field pictures judges t frames
The quantity of target associated by each foreground area in image;
If foreground area and a target association in t-1 two field pictures in t two field pictures, according to should in t-1 two field pictures
The morpheme parameter of the foreground area, the hierarchical model of the background scene and the mesh in morpheme parameter, the t two field pictures of target
The feature templates being marked in t-1 two field pictures carry out monotrack treatment and obtain the shape of the target in t frame video images
State;
If a foreground area and multiple target associations in t-1 two field pictures in t two field pictures, according to many in t-1 two field pictures
The morpheme parameter of the foreground area, the hierarchical model of the background scene and institute in morpheme parameter, the t two field pictures of individual target
Stating feature templates of multiple targets in t-1 two field pictures carries out Markov chain Monte Carlo treatment acquisition t frame figures
The state of multiple target as in;
If a foreground area is not associated with any one target in t-1 two field pictures in t two field pictures, extract and be located at the
The state of each target of the foreground area in t two field pictures;
(6) using each target t two field pictures state as each target t two field pictures tracking result, and using described
Each target updates each target in t frame figures in the feature templates in t-1 two field pictures of state and each target of t two field pictures
Feature templates as in;
(7) whether judgment frame order t is equal to the totalframes amount M of video, if so, then terminating, otherwise, t=t+1 is made, into step
(4);
Wherein, N+2≤t≤M, M are the totalframes amount of video.
2. multi-object tracking method according to claim 1, it is characterised in that the step (4) includes following sub-step:
(41) background difference processing is carried out to t two field pictures and background scene, the morpheme of each foreground area in t two field pictures is obtained
Parameter;
(42) in the morpheme parameter and the t-1 two field pictures according to each foreground area in the t two field pictures each target shape
Position parameter judges the overlapping cases of each target in foreground area and the t-1 two field pictures in the t two field pictures;
If in the t two field pictures in foreground area and the t-1 two field pictures target overlapping region area and t-1 frame figures
As the region area ratio of the target is more than 10%, then the foreground area is closed with the target in t-1 frequency images in t two field pictures
Connection, otherwise the foreground area is not associated with the target in t-1 frequency images in t two field pictures;
(43) all targets in all foreground areas and t-1 two field pictures in traversal t two field pictures, obtain each in t two field pictures
The incidence relation of each target of foreground area and t-1 two field pictures.
3. multi-object tracking method according to claim 1, it is characterised in that the step (5) if in t two field pictures
One foreground area with during a target association, obtained in t-1 two field pictures the state of the target in t two field pictures include with
Lower sub-step:
(511) carry out multiple stochastical sampling to the morpheme parameter of the target in t-1 two field pictures and obtain to be located in the t two field pictures
Multiple particles in foreground area, and morpheme parameter and the hierarchical model of background scene to each particle carries out random combine acquisition
Each particle is possible to barrier bed time relation with the hierarchical model of background scene;
(512) hierarchical model according to each particle and background scene blocks hierarchical relationship, extracts each particle and is blocked at this
De-occlusion region under hierarchical relationship, each particle is obtained used in each particle in the de-occlusion region under blocking hierarchical relationship
In the template characteristic under blocking hierarchical relationship;
(513) template characteristic according to each particle in the case where difference blocks hierarchical relationship and the mould of the target in t-1 two field pictures
Plate features obtain observation likelihood probability of each particle in the case where difference blocks hierarchical relationship, and the particle maximum to observe likelihood probability
Morpheme parameter as the target in t two field pictures morpheme parameter, to observe the maximum particle of likelihood probability and background scene
The hiding relation of hierarchical model block hierarchical relationship with the hierarchical model of background scene in t two field pictures as the target,
And morpheme parameter and the target are blocked in t two field pictures with the hierarchical model of background scene in t two field pictures by the target
Hierarchical relationship is used as the state of the target in t two field pictures.
4. multi-object tracking method according to claim 1, it is characterised in that according to formula in the step (6)I-th template characteristic of target in t two field pictures is updated, wherein,For i-th target exists
The template characteristic extracted using target de-occlusion region in t two field pictures, γiFor the de-occlusion region of i-th target accounts for i-th
The percentage of target,The template characteristic for being i-th target in t-1 two field pictures, 1≤i≤L, L are mesh in t two field pictures
Target total quantity.
5. multi-object tracking method according to claim 1, it is characterised in that step (5) if in one in t two field pictures
Foreground area is with during multiple target associations, state of the multiple targets of acquisition in t two field pictures includes following in t-1 two field pictures
Sub-step:
(521) barrier bed in the state acquisition t-1 two field pictures according to multiple targets in t-1 two field pictures between multiple targets
In secondary relation and t-1 two field pictures hierarchical relationship is blocked between multiple targets and the hierarchical model of background scene;
According between multiple target in t-1 two field pictures block hierarchical relationship, t-1 two field pictures in multiple targets and ambient field
The morpheme gain of parameter t frame figures for blocking multiple targets in hierarchical relationship and t-1 two field pictures between the hierarchical model of scape
The original state of multiple targets as in;
(522) state parameter in state of the multiple targets after r-th renewal in t two field pictures is randomly updated, is obtained
The r+1 update after state of multiple targets in t two field pictures;And the multiple targets after being updated according to the r+1 are in t
State in two field picture, update for r-th after state and multiple target of multiple targets in t two field pictures in t-1 frames
Feature templates in image obtain the probability of acceptance of state of the multiple targets after the r+1 renewal in t two field pictures;
(523) whether state of the multiple targets after the r+1 renewal in t two field pictures is received according to probabilistic determination, if so,
State of the multiple targets after the r+1 renewal in t two field pictures is recorded, and enters step (524), otherwise, be directly entered
Step (524);
(524) judge to update order r and whether be equal to update total degree R, if so, then making r=r+1, and enter step (522), it is no
Then enter step (525);
(525) by the renewal of all receiving after state of multiple targets in t two field pictures in each parameter average, obtain
Mean state of multiple targets in t two field pictures;Using mean state of multiple targets in t two field pictures as multiple targets
State in t two field pictures.
6. multi-object tracking method according to claim 5, it is characterised in that according to formula in the step (522)Obtain the r+1 update after multiple targets in t two field pictures
The probability of acceptance of state;
Wherein,It is the observation likelihood probability of state of the multiple targets after r-th renewal in t two field pictures; Be r-th update after multiple targets in t two field pictures in the state of jth
The observation likelihood probability of individual target,It is the fortune of state of the multiple targets after r-th renewal in t two field pictures
Movable model probability,Φ is to be desired forVariance is
The probability function of the Gaussian Profile of Σ,It is that j-th prediction central point of the minimum rectangle of target is surrounded in t two field pictures
Coordinate,Be r-th update after multiple target t two field pictures in the state of surround j-th minimum rectangle of target
Center point coordinate;Suggestion before and after being updated for state is distributed ratio, and 1≤j≤Q, Q are same with t two field pictures
There is the target number of incidence relation in foreground area.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710073835.3A CN106920253A (en) | 2017-02-10 | 2017-02-10 | It is a kind of based on the multi-object tracking method for blocking layering |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710073835.3A CN106920253A (en) | 2017-02-10 | 2017-02-10 | It is a kind of based on the multi-object tracking method for blocking layering |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106920253A true CN106920253A (en) | 2017-07-04 |
Family
ID=59453675
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710073835.3A Pending CN106920253A (en) | 2017-02-10 | 2017-02-10 | It is a kind of based on the multi-object tracking method for blocking layering |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106920253A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109446942A (en) * | 2018-10-12 | 2019-03-08 | 北京旷视科技有限公司 | Method for tracking target, device and system |
CN110136174A (en) * | 2019-05-22 | 2019-08-16 | 北京华捷艾米科技有限公司 | A kind of target object tracking and device |
CN110390279A (en) * | 2019-07-08 | 2019-10-29 | 丰图科技(深圳)有限公司 | Coordinate recognition method, device, equipment and computer readable storage medium |
CN110517483A (en) * | 2019-08-06 | 2019-11-29 | 杭州博信智联科技有限公司 | A kind of traffic information processing method and digital rail roadside unit |
CN111260686A (en) * | 2020-01-09 | 2020-06-09 | 滨州学院 | Target tracking method and system for anti-shielding multi-feature fusion of self-adaptive cosine window |
CN111708021A (en) * | 2020-07-15 | 2020-09-25 | 四川长虹电器股份有限公司 | Personnel tracking and identifying algorithm based on millimeter wave radar |
CN112802054A (en) * | 2021-02-04 | 2021-05-14 | 重庆大学 | Mixed Gaussian model foreground detection method fusing image segmentation |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2000082145A (en) * | 1998-01-07 | 2000-03-21 | Toshiba Corp | Object extraction device |
CN101141633A (en) * | 2007-08-28 | 2008-03-12 | 湖南大学 | Moving object detecting and tracing method in complex scene |
CN101408982A (en) * | 2008-10-29 | 2009-04-15 | 南京邮电大学 | Object-tracking method base on particle filtering and movable contour model |
CN102509306A (en) * | 2011-10-08 | 2012-06-20 | 西安理工大学 | Specific target tracking method based on video |
CN103077539A (en) * | 2013-01-23 | 2013-05-01 | 上海交通大学 | Moving object tracking method under complicated background and sheltering condition |
JP2014093023A (en) * | 2012-11-06 | 2014-05-19 | Canon Inc | Object detection device, object detection method and program |
WO2014135910A1 (en) * | 2013-03-08 | 2014-09-12 | JACQUEMET, Jean-Philippe | Method of replacing objects in a video stream and computer program |
US20160004929A1 (en) * | 2014-07-07 | 2016-01-07 | Geo Semiconductor Inc. | System and method for robust motion detection |
CN105335986A (en) * | 2015-09-10 | 2016-02-17 | 西安电子科技大学 | Characteristic matching and MeanShift algorithm-based target tracking method |
-
2017
- 2017-02-10 CN CN201710073835.3A patent/CN106920253A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2000082145A (en) * | 1998-01-07 | 2000-03-21 | Toshiba Corp | Object extraction device |
CN101141633A (en) * | 2007-08-28 | 2008-03-12 | 湖南大学 | Moving object detecting and tracing method in complex scene |
CN101408982A (en) * | 2008-10-29 | 2009-04-15 | 南京邮电大学 | Object-tracking method base on particle filtering and movable contour model |
CN102509306A (en) * | 2011-10-08 | 2012-06-20 | 西安理工大学 | Specific target tracking method based on video |
JP2014093023A (en) * | 2012-11-06 | 2014-05-19 | Canon Inc | Object detection device, object detection method and program |
CN103077539A (en) * | 2013-01-23 | 2013-05-01 | 上海交通大学 | Moving object tracking method under complicated background and sheltering condition |
WO2014135910A1 (en) * | 2013-03-08 | 2014-09-12 | JACQUEMET, Jean-Philippe | Method of replacing objects in a video stream and computer program |
US20160004929A1 (en) * | 2014-07-07 | 2016-01-07 | Geo Semiconductor Inc. | System and method for robust motion detection |
CN105335986A (en) * | 2015-09-10 | 2016-02-17 | 西安电子科技大学 | Characteristic matching and MeanShift algorithm-based target tracking method |
Non-Patent Citations (2)
Title |
---|
李峥: "智能视频监控中的遮挡目标跟踪技术研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
***: "固定背景下单_多目标行人跟踪算法研究", 《中国优秀硕士论文全文数据库 信息科技辑》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109446942A (en) * | 2018-10-12 | 2019-03-08 | 北京旷视科技有限公司 | Method for tracking target, device and system |
CN110136174A (en) * | 2019-05-22 | 2019-08-16 | 北京华捷艾米科技有限公司 | A kind of target object tracking and device |
CN110136174B (en) * | 2019-05-22 | 2021-06-22 | 北京华捷艾米科技有限公司 | Target object tracking method and device |
CN110390279A (en) * | 2019-07-08 | 2019-10-29 | 丰图科技(深圳)有限公司 | Coordinate recognition method, device, equipment and computer readable storage medium |
CN110517483A (en) * | 2019-08-06 | 2019-11-29 | 杭州博信智联科技有限公司 | A kind of traffic information processing method and digital rail roadside unit |
CN111260686A (en) * | 2020-01-09 | 2020-06-09 | 滨州学院 | Target tracking method and system for anti-shielding multi-feature fusion of self-adaptive cosine window |
CN111260686B (en) * | 2020-01-09 | 2023-11-10 | 滨州学院 | Target tracking method and system for anti-shielding multi-feature fusion of self-adaptive cosine window |
CN111708021A (en) * | 2020-07-15 | 2020-09-25 | 四川长虹电器股份有限公司 | Personnel tracking and identifying algorithm based on millimeter wave radar |
CN111708021B (en) * | 2020-07-15 | 2022-04-15 | 四川长虹电器股份有限公司 | Personnel tracking and identifying algorithm based on millimeter wave radar |
CN112802054A (en) * | 2021-02-04 | 2021-05-14 | 重庆大学 | Mixed Gaussian model foreground detection method fusing image segmentation |
CN112802054B (en) * | 2021-02-04 | 2023-09-01 | 重庆大学 | Mixed Gaussian model foreground detection method based on fusion image segmentation |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106920253A (en) | It is a kind of based on the multi-object tracking method for blocking layering | |
CN110517288B (en) | Real-time target detection tracking method based on panoramic multi-path 4k video images | |
CN110688987B (en) | Pedestrian position detection and tracking method and system | |
CN102542289B (en) | Pedestrian volume statistical method based on plurality of Gaussian counting models | |
Dai et al. | Multi-task faster R-CNN for nighttime pedestrian detection and distance estimation | |
CN102136059B (en) | Video- analysis-base smoke detecting method | |
CN104183142B (en) | A kind of statistical method of traffic flow based on image vision treatment technology | |
Bouvie et al. | Tracking and counting vehicles in traffic video sequences using particle filtering | |
CN104091348A (en) | Multi-target tracking method integrating obvious characteristics and block division templates | |
CN103927520A (en) | Method for detecting human face under backlighting environment | |
CN102147861A (en) | Moving target detection method for carrying out Bayes judgment based on color-texture dual characteristic vectors | |
CN101957997A (en) | Regional average value kernel density estimation-based moving target detecting method in dynamic scene | |
Shaikh et al. | Moving object detection approaches, challenges and object tracking | |
CN109359563A (en) | A kind of road occupying phenomenon real-time detection method based on Digital Image Processing | |
CN103136537A (en) | Vehicle type identification method based on support vector machine | |
CN105654508A (en) | Monitoring video moving target tracking method based on self-adaptive background segmentation and system thereof | |
Jiang et al. | An Approach for Crowd Density and Crowd Size Estimation. | |
CN103049788A (en) | Computer-vision-based system and method for detecting number of pedestrians waiting to cross crosswalk | |
CN113763427A (en) | Multi-target tracking method based on coarse-fine shielding processing | |
CN104091332A (en) | Method for optimizing multilayer image segmentation of multiclass color texture images based on variation model | |
CN106643725B (en) | A kind of robot localization air navigation aid based on floor tile contour line | |
Gao et al. | Region-based moving shadow detection using watershed algorithm | |
Shiru et al. | Research on multi-feature front vehicle detection algorithm based on video image | |
CN112215109A (en) | Vehicle detection method and system based on scene analysis | |
Hu et al. | An improved vehicle detection and tracking model |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170704 |