CN110349181A - One kind being based on improved figure partition model single camera multi-object tracking method - Google Patents
One kind being based on improved figure partition model single camera multi-object tracking method Download PDFInfo
- Publication number
- CN110349181A CN110349181A CN201910503562.0A CN201910503562A CN110349181A CN 110349181 A CN110349181 A CN 110349181A CN 201910503562 A CN201910503562 A CN 201910503562A CN 110349181 A CN110349181 A CN 110349181A
- Authority
- CN
- China
- Prior art keywords
- detection block
- sliding window
- track
- small fragment
- appearance
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/231—Hierarchical techniques, i.e. dividing or merging pattern sets so as to obtain a dendrogram
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/223—Analysis of motion using block-matching
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30241—Trajectory
Abstract
The invention discloses one kind to be based on improved figure partition model single camera multi-object tracking method, belongs to target tracking domain.The present invention uses two layers of reasoning structure, and single camera multiple target tracking is seen mapping partitioning problem, is solved with two-value integer programming (BIP).In the stage of hierarchical reasoning, first stage uses shorter sliding window, will belong to same person's detection block and is divided into the same figure subregion and forms short track small fragment；The sliding window of any is grown in second stage use, and the track small fragment for belonging to the same person is divided into the same figure subregion and forms long track.For second stage sliding window near the intersection of non-overlapping segment it is possible that the case where missing inspection, the method that the present invention proposes overlapping sliding window improves algorithm keeps track precision to reduce omission factor.On the other hand, it is easy to happen between target identity caused by mutually blocking when multiple target tracking to convert, therefore the method that the present invention proposes profile constraints, reduces the generation of identity conversion.
Description
Technical field
The invention belongs to target tracking domains, are based on improved figure partition model single camera more particularly, to one kind
Multi-object tracking method.
Background technique
Single camera multiple target tracking (Multi-target Single camera Tracking, MTSCT) is computer
Important component in vision, what it was studied is all location informations of multiple interesting targets in one section of successive video frames,
The motion profile of all interesting targets is obtained, and then analyzes the behavior of tracking target.
Single camera multi-object tracking method has obtained some good as a result, including that Kalman filtering, particle are filtered now
Wave, correlation filtering, optimal Hungary Algorithm, two-value integer programming etc..In actual application, although Kalman filtering and grain
The methods of son filtering usually intuitively is readily appreciated that by probabilistic inference, but their usual more difficult deductions, therefore the present invention is normal
The method for often selecting optimal Hungary, two-value integer programming etc. to minimize based on energy function.But these methods still exist
The problem of how optimizing calculating, Ergys Ristani et al. can effectively solve this using the figure partition model of hierarchical reasoning and ask
Topic, the algorithm using one two layers hierarchical reasoning structure reduce data scale, respectively in short time and prolonged sliding window
Every group observations are divided into corresponding pedestrian's body according to space-time and appearance similitude with two-value integer programming (BIP) by interior operation
Part, first layer forms track small fragment, and the second layer forms long track.But the former algorithm second layer is carried out on non-overlap window
, in this way in the intersection of window it is possible that missing inspection, on the other hand, when carrying out the data correlation of track small fragment, mesh
The partial occlusion because caused by being closer may occur between mark and target, may result in track algorithm in this way and mistake occurs
Matching.
Summary of the invention
In view of the drawbacks of the prior art, it is an object of the invention to solve the more mesh of prior art diagram partition model single camera
Mark tracking there is technical issues that missing inspection,.
To achieve the above object, in a first aspect, the embodiment of the invention provides one kind to be based on improved figure partition model list
Camera multi-object tracking method, method includes the following steps:
S1. the video flowing got to single camera extracts the detection block of all targets in every frame picture, each detection block
A corresponding target, and select first not to be overlapped sliding window length and the second overlapping sliding window length, wherein the second overlapping
Sliding window length is greater than first and is not overlapped sliding window length, and the length of the lap of the second overlapping sliding window length
The integral multiple for not being overlapped sliding window length for first；
It S2. is first not to be overlapped in the sliding window of sliding window length in length, to all detection blocks in the sliding window,
Hierarchical cluster is carried out with the Euclidean distance of detection block central point, obtains k1A space-time group；
S3. the foundation of the motion relevance and appearance similitude of detection block as data correlation is used, to each space-time group
In detection block carry out data correlation, by solve BIP repartition into a new subset, obtain the inspection for belonging to same target
Frame is surveyed, these detection blocks are connected into track small fragment according to frame sequence；
S4. in the overlapping sliding window that length is the second overlapping sliding window length, according to the similitude of external appearance characteristic,
All track small fragments in the sliding window are divided into k2Group appearance group；
S5. the foundation of the motion relevance and appearance similitude of track small fragment as data correlation is used, to each outer
Track small fragment in sight group carries out data correlation, repartitions into a new subset by solving BIP, obtains belonging to identical
These track small fragments are connected into the corresponding pursuit path of each target according to frame sequence by the track small fragment of target.
Specifically, for each detection block, its HSV histogram feature is calculated as external appearance characteristic, then detection block D1And D2It
Between appearance Similarity measures formula it is as follows:
sima(D1,D2)=max [1- α * d (φ1,φ2),0]
Wherein, φ1And φ2It is D respectively1And D2External appearance characteristic, d (φ1,φ2) indicate the distance between external appearance characteristic, α
For the weight coefficient of distance.
Specifically, it is foundation with the distance between detection block, finds out apart from current detection frame D1=(φ1,p1,t1,v1) most
Close detection block D2=(φ2,p2,t2,v2), motion relevance calculation formula is as follows:
simst(D1,D2)=max [1- β (e (D1,D2)+e(D2,D1)),0]
Wherein, e (D1,D2)=‖ q1-p2‖2What is calculated is forward error, i.e. detection block D2Physical location p2With detection block D1
In time t2Estimated location q1=p1+v1(t2-t1) between Euclidean distance；e(D2, D1) what is calculated is backward error, β is indicated
Receptible error degree, p1、t1、v1Respectively indicate detection block D1Physical location, speed and corresponding frame time, p2、t2、v2Point
It Biao Shi not detection block D2Physical location, speed and corresponding frame time.
Specifically, in step S3, the weight calculation formula on each pair of node side is as follows:
Wherein, λ indicates the width of the intermediate zone between positive negative correlation, sima(D1,D2) indicate detection block D1And D2Between
Appearance similitude, simst(D1,D2) indicate detection block D1And D2Between motion relevance.
Specifically, for each track small fragment, the HSV histogram for all detection blocks that the track small fragment includes is calculated
The average value of feature is as external appearance characteristic.
Specifically, when calculating the motion relevance of track small fragment, track actual speed is replaced with the average speed of track,
Then the forward direction range error and backward range error of two sections of tracks are calculated separately.
Specifically, in step S5, increase the space constraint of track and punished come the path segment more to overlapping region,
Specifically, in the appearance weight on original graph model side multiplied by a new variable αij, multiplied by 1- α in movement weightij；
Wherein, Ti∩TjThe overlap proportion of path segment i and path segment j are represented, μ indicates overlapping region threshold value.
Second aspect, the embodiment of the invention provides a kind of computer readable storage medium, the computer-readable storage mediums
Computer program is stored in matter, which realizes described in above-mentioned first aspect when being executed by processor based on improvement
Figure partition model single camera multi-object tracking method.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, have below beneficial to effect
Fruit:
1. the present invention is directed to using the missing inspection problem that sliding window may cause in window intersection is not overlapped, by being layered
The second layer of reasoning structure is using overlapping sliding window, so that the shorter track small fragment of sliding window intersection participates in reasoning
In the process, achieve the effect that the omission factor for reducing single camera multiple target tracking, improve algorithm keeps track precision.
2. the present invention suffers too close for target and identity that is may cause converts (erroneous matching) problem, by carrying out
The space constraint of track is introduced when path matching to inhibit this phenomenon, the path segment more to overlapping region is punished
It penalizes, to reduce the error hiding number of single camera multiple target tracking.
Detailed description of the invention
Fig. 1 is provided in an embodiment of the present invention a kind of based on improved figure partition model single camera multi-object tracking method
Flow chart；
Fig. 2 (a) is video flowing provided in an embodiment of the present invention；
Fig. 2 (b) is space-time group provided in an embodiment of the present invention；
Fig. 2 (c) is track small fragment provided in an embodiment of the present invention；
Fig. 2 (d) is appearance group provided in an embodiment of the present invention；
Fig. 2 (e) is overlapping sliding window provided in an embodiment of the present invention；
Fig. 2 (f) is long track provided in an embodiment of the present invention；
Fig. 3 is provided in an embodiment of the present invention using the effect contrast figure for not being overlapped sliding window and sliding window；
Fig. 4 (a) is the multiple target tracking result schematic diagram provided in an embodiment of the present invention that trajectory range constraint is not added；
Fig. 4 (b) is the multiple target tracking result schematic diagram of addition trajectory range constraint provided in an embodiment of the present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
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.
Single camera multiple target tracking is carried out using the mode of batch processing, is in simple terms exactly to input a batch to have several frames
Continuous picture extracts external appearance characteristic to these pictures and motion feature calculates similitude, most like several frame pictures series connection
Get up be exactly a people pursuit path.But when multiple target tracking, the picture of the target of unknown number is handled simultaneously,
Therefore the data volume of processing is very more.
In order to reduce solved the complex nature of the problem, it is complicated that the present invention using two layers reasoning structure reduces algorithm
Degree, regards single camera multiple target tracking as a figure partitioning problem.Figure subregion is the batch detection frame root that will be inputted in fact
It is divided into the set for belonging to the detection block of different people according to their appearance and kinematic similarity, this is a np hard problem, uses two-value
Integer programming solves.It is at all levels to be carried out in sliding window to reach real-time purpose, in the stage of hierarchical reasoning, first stage
Using shorter sliding window, same person's detection block will be belonged to it is divided into the same figure subregion and form short track small fragment；
The sliding window of any is grown in second stage use, and the track small fragment for belonging to the same person is divided into the same figure subregion and is formed
Long track.For second stage sliding window near the intersection of non-overlapping segment it is possible that the case where missing inspection, this
The method that invention proposes overlapping sliding window improves algorithm keeps track precision to reduce omission factor.On the other hand, multiple target with
It is easy to happen between target identity caused by mutually blocking when track to convert, therefore the method that the present invention proposes profile constraints, reduce
The generation of identity conversion.
Be based on improved figure partition model single camera multi-object tracking method as shown in Figure 1, a kind of, this method include with
Lower step:
The video flowing that step S1. gets single camera extracts the detection block of all targets in every frame picture, Mei Gejian
The corresponding target of frame is surveyed, and selects first not to be overlapped sliding window length and the second overlapping sliding window length, wherein second
Overlapping sliding window length is greater than first and is not overlapped sliding window length, and the lap of the second overlapping sliding window length
Length is the first integral multiple for not being overlapped sliding window length.
The detection block of every frame picture, these detection block packets are extracted using DPM detector to the video flowing as shown in Fig. 2 (a)
Video frame information is contained.The detection block of all frames is screened, screens out the low detection block of confidence level, except the visual field interested
Detection block, excessive detection block, obtain effective detection block.It will test coordinate of the frame on picture and be converted into world coordinates, with inspection
The centre coordinate of frame is surveyed as location information.
Step S2. is not overlapped in the sliding window of sliding window length in length for first, to all inspections in the sliding window
Frame is surveyed, hierarchical cluster is carried out with the Euclidean distance of detection block central point, obtains k1A space-time group.
Hierarchical cluster is carried out with the Euclidean distance of detection block central point, obtains k1A space-time group, as shown in Fig. 2 (b), distance
Close detection block is divided into one group, kiIt can be arranged according to the actual situation.
Step S3. uses the foundation of the motion relevance and appearance similitude of detection block as data correlation, to it is each when
Detection block in empty group carries out data correlation, repartitions into a new subset by solving BIP, obtains belonging to same target
Detection block, these detection blocks are connected into track small fragment according to frame sequence.
Multiple target tracking is seen into mapping partitioning problem, is inputted as the set V of n node, node can be individual detection
Frame is also possible to fused track small fragment, and final set V is divided into multiple set, and each set includes the same identity
Node, the same set is interior to accumulate highest positive correlation weight, accumulates highest negative correlation weight between different sets, i.e., complete
At the solution of figure subregion.Scheme G=(V, E, W), for a pair of of node (u, v) in set V, the side between node pair belongs to E, side
On weight wuvBelong to W, wuvIndicate that node to the correlation on the side (u, v), judges that node (u, v) belongs to the same identity
Confidence level.
Above-mentioned figure partition model can be derived as two-value integer programming problem (BIP), be defined as follows:
Set X is to binary variable xuvThe set of all possible combinations of assignment, if node u and v are identical identity,
So illustrate xuvIt is 1.xuvIt is a binary variable, takes 0 or 1.wuvIt is weight.The constraint condition of inequality indicates transitivity, such as
Fruit node u and v are the same identity, then node u and t are also the same identity.
For all detection blocks, its HSV histogram feature is calculated as external appearance characteristic, if φ1And φ2It is detection block respectively
D1And D2External appearance characteristic, then D1And D2Between appearance similitude such as following formula:
sima(D1,D2)=max [1- α * d (φ1,φ2),0]
Wherein, d (φ1, φ2) indicate the distance between external appearance characteristic, herein be EMD (Earth Mover ' s
Distance) distance seeks the distance of histogram.α takes 1 herein.
For the motion feature of each node, the speed of uniform motion model reasoning detection block is used.Between detection block
Distance be foundation, the detection block nearest apart from current detection frame is considered the inspection that the next frame of current detection frame is likely to occur
Frame is surveyed, then calculates the speed of each pair of detection block that may be nearest, minimum speed is the estimating speed of current detection frame, the speed
Degree is no more than threshold speed.
To detection block D1=(φ1, p1, t1,v1) and D2=(φ2, p2, t2, v2), the calculation formula of motion relevance such as 2-3
It is shown.
simst(D1,D2)=max [1- β (e (D1,D2)+e(D2,D1)),0]
e(D1,D2)=‖ q1-p2‖2What is calculated is forward error, i.e. detection block D2Physical location p2With detection block D1When
Between t2Estimated location q1=p1+v1(t2-t1) between Euclidean distance.Similarly, e (D2,D1) what is calculated is backward error.β is indicated
Receptible error degree, takes 1 herein.
After the appearance and motion relevance for calculating each pair of node, each pair of section is defined using the nonlinear combination of the two
Weight on point side.
Wherein, λ indicates the width of the intermediate zone between positive negative correlation, value 6；Work as sima(D1, D2)=simst(D1,D2)
Appearance similitude is indicated when=0.5, and to target is judged, whether similar effect is consistent with operation correlation.
First layer, which generates track small fragment to the testing result of input, can retain the institute of detection block as shown in Fig. 2 (c)
There is information, and can be reduced the input of next level.
Step S4. is in the overlapping sliding window that length is the second overlapping sliding window length, according to the similar of external appearance characteristic
Property, all track small fragments in the sliding window are divided into k2Group appearance group.
Compare appearance similitude if it is to two sections of path segments, then external appearance characteristic is the flat of total HSV histogram feature
Mean value.It is clustered according to appearance similitude using k-means and track small fragment is divided into the k as shown in Fig. 2 (d)2Group appearance group, k2
Value be arranged on demand in an experiment.
As shown in figure 3, the different track small fragment of circle, pentagon, star-like expression in figure, connects expression and merges
The long path segment come, if the shorter track small fragment of the circle of intersection can be due in sliding window 1 using sliding window is not overlapped
Confidence level is low to be judged as empty inspection to be rejected, but actually the long path segment of circle of it and sliding window 2 belong to it is same
A identity causes track lack part correctly to track if giving up.Therefore the present invention is slided using the overlapping as shown in Fig. 2 (e)
Dynamic window eliminates the generation of such phenomenon, and overlap length is set as 4 times of first layer length of window, in this way in fusion path segment
When, the circle small fragment of 1 intersection of sliding window can also participate in the reasoning process of sliding window 2.
Step S5. uses the foundation of the motion relevance and appearance similitude of track small fragment as data correlation, to every
Track small fragment in a appearance group carries out data correlation, repartitions into a new subset by solving BIP, is belonged to
These track small fragments are connected into the corresponding pursuit path of each target according to frame sequence by the track small fragment of same target.
When calculating motion relevance to path segment, principle is same as above.The practical speed in track is replaced with the average speed of track
Degree, then calculates separately the forward direction range error and backward range error of two sections of tracks.
For generating track small fragment and when track is likely to occur blocks or missing inspection problem, filled out with linear interpolation method
It is filled with smooth track.
In actual scene, moved when there is the people that the situation blocked between target and target, especially two walk side by side
State is similar, when appearance distinguishes unconspicuous due to blocking, it may occur that identity transfer problem.Two rails as shown in Fig. 4 (a)
Mark segment suffers close and has similar motion state, and when target is excessively close apart, influence of the HSV feature to target correlation is small
When motion relevance feature bring influences, the identity conversion of target will occur, as shown in box, close to the track of road
(the corresponding target close to road) is connected in the target in market, and identity conversion occurs for target.Increase trajectory range about
Shown in Shu Yihou such as Fig. 4 (b), trajectory range constrain so that suffer between excessively close target the correlation of motion feature weaken, by
Appearance effects are bigger, as shown in box, it can be seen that can preferably distinguish two sections of tracks, there is no identity conversion is existing
As.The space constraint that the present invention increases track is punished come the path segment more to overlapping region.To original graph model side
On appearance weight on multiplied by a new variable αij, multiplied by 1- α in movement weightij。
Wherein, Ti∩TjRepresent the overlap proportion of path segment i and path segment j.Trajectory range constraint in this algorithm is only
Corresponding punishment, this kind punishment exponentially multiple increasing with the increase of overlapping region are made to the track that overlapping region is more than 60%
It is long.Second level makes inferences input trajectory small fragment, generates such as the longer complete trajectory of Fig. 2 (f).
The present invention is by it compared with other single camera multiple target tracking algorithms are on Town Center public data collection
MOTA (multiple target tracking accuracy rate) and IDswithes (identity conversion number) index, as shown in table 1.
Multiple target tracking accuracy rate index is primarily used to reflection algorithm erroneous detection number during tracking, missing inspection number and identity
Convert the summation of number specific gravity shared in tracking target sum.The calculation formula of MOTA is as follows:
Wherein, gtFor the target sum of t moment really marked.idtThe respectively erroneous detection target of t moment
Number, missing inspection number of targets and identity convert number.The quantity of erroneous matching when what identity conversion number indicated is tracking.
Table 1
4500 frame of Town Center video sequence overall length is under the more complex streetscape scene for having more pedestrian to block
Data set, the present invention using 12 frames short time window and 12 seconds long-time windows and 5 appearance groups because this
Scape is more crowded.It is proposed by the present invention based on the single camera multiple target tracking algorithm of figure partition model in Town Center data
MOTA on collection is 79.25%, improves 11.95% than Leal-Taxixe et al., the identity of the mentioned algorithm of the present invention converts number
Reduce 34 than Leal-Taxixe et al..Analysis reason may be because the algorithm of Leal-Taxixe et al. consider it is more
The time-space relationship and motion feature of track, discriminating power when not accounting for external appearance characteristic, therefore being blocked to target compared with
It is weak, it is easy to appear the phenomenon that identity is converted.Algorithm of the invention considers external appearance characteristic and space-time characteristic simultaneously, therefore to blocking
There is certain robustness, identity conversion quantity is reduced after the constraint of application space, overcomes space-time characteristic phase to a certain extent
Pedestrian's occlusion issue when close.Zamir et al. is also that multiple target tracking problem is carried out data correlation with graph model, it is also considered that
Hierarchical reasoning, but data correlation is separately carried out one by one to target when the algorithm is every time to video frame batch processing,
And algorithm of the invention carries out batch processing to video frame using sliding window, while carrying out data pass to multiple objects in batch processing
Connection, the experimental results showed that, the raising 3.66% of MOTA ratio Zamir of the mentioned algorithm of the present invention et al., it was confirmed that inventive algorithm
Validity.BASELINE algorithm is the single camera multiple target tracking algorithm that Ristani et al. is proposed, the present invention proposes calculation
Method improves 0.82% than the MOTA of the algorithm, and identity, which converts number, reduces by 26, and analysis reason may be that the algorithm is sliding in non-overlap
Batch processing is carried out to video frame on dynamic window, MOTA may be caused to reduce due to missing inspection, in addition the present invention increases the sky between track
Between constrain so that erroneous matching caused by identity is converted is reduced to improving tracking accuracy.
More than, the only preferable specific embodiment of the application, but the protection scope of the application is not limited thereto, and it is any
Within the technical scope of the present application, any changes or substitutions that can be easily thought of by those familiar with the art, all answers
Cover within the scope of protection of this application.Therefore, the protection scope of the application should be subject to the protection scope in claims.
Claims (8)
1. one kind be based on improved figure partition model single camera multi-object tracking method, which is characterized in that this method include with
Lower step:
S1. the video flowing got to single camera, extracts the detection block of all targets in every frame picture, and each detection block is corresponding
One target, and select first not to be overlapped sliding window length and the second overlapping sliding window length, wherein the second overlapping sliding
Length of window is greater than first and is not overlapped sliding window length, and the length of the lap of the second overlapping sliding window length is the
One is not overlapped the integral multiple of sliding window length；
S2. it is not overlapped in the sliding window of sliding window length in length for first, to all detection blocks in the sliding window, with inspection
The Euclidean distance for surveying frame central point carries out hierarchical cluster, obtains k1A space-time group；
S3. the foundation of the motion relevance and appearance similitude of detection block as data correlation is used, in each space-time group
Detection block carries out data correlation, repartitions into a new subset by solving BIP, obtains the detection for belonging to same target
These detection blocks are connected into track small fragment according to frame sequence by frame；
S4. in the overlapping sliding window that length is the second overlapping sliding window length, according to the similitude of external appearance characteristic, the cunning
All track small fragments in dynamic window are divided into k2Group appearance group；
S5. the foundation of the motion relevance and appearance similitude of track small fragment as data correlation is used, to each appearance group
In track small fragment carry out data correlation, by solve BIP repartition into a new subset, obtain belonging to same target
Track small fragment, these track small fragments are connected into the corresponding pursuit path of each target according to frame sequence.
2. the method as described in claim 1, which is characterized in that for each detection block, calculate its HSV histogram feature conduct
External appearance characteristic, then detection block D1And D2Between appearance Similarity measures formula it is as follows:
sima(D1, D2)=max [1- α * d (φ1, φ2), 0]
Wherein, φ1And φ2It is D respectively1And D2External appearance characteristic, d (φ1, φ2) indicate the distance between external appearance characteristic, α be away from
From weight coefficient.
3. method according to claim 1 or 2, which is characterized in that with the distance between detection block be foundation, find out distance and work as
Preceding detection block D1=(φ1, p1, t1, v1) nearest detection block D2=(φ2, p2, t2, v2), motion relevance calculation formula is as follows:
simst(D1, D2)=max [1- β (e (D1, D2)+e(D2, D1)), 0]
Wherein, e (D1, D2)=| | q1-p2||2What is calculated is forward error, i.e. detection block D2Physical location p2With detection block D1?
Time t2Estimated location q1=p1+v1(t2-t1) between Euclidean distance；e(D2, D1) what is calculated is backward error, β indicates energy
The error degree of receiving, p1、t1、v1Respectively indicate detection block D1Physical location, speed and corresponding frame time, p2、t2、v2Respectively
Indicate detection block D2Physical location, speed and corresponding frame time.
4. method as described in any one of claims 1 to 3, which is characterized in that the weight meter in step S3, on each pair of node side
It is as follows to calculate formula:
Wherein, λ indicates the width of the intermediate zone between positive negative correlation, sima(D1, D2) indicate detection block D1And D2Between appearance
Similitude, simst(D1, D2) indicate detection block D1And D2Between motion relevance.
5. method according to claim 2, which is characterized in that for each track small fragment, calculate the track small fragment packet
The average value of the HSV histogram feature of all detection blocks contained is as external appearance characteristic.
6. method as claimed in claim 3, which is characterized in that when calculating the motion relevance of track small fragment, with track
Average speed replaces track actual speed, then calculates separately the forward direction range error and backward range error of two sections of tracks.
7. method as claimed in claim 4, which is characterized in that in step S5, the space constraint for increasing track comes to overlay region
The more path segment in domain punished, specifically, in the appearance weight on original graph model side multiplied by a new variable
αij, multiplied by 1- α in movement weightij；
Wherein, Ti∩TjThe overlap proportion of path segment i and path segment j are represented, μ indicates overlapping region threshold value.
8. a kind of computer readable storage medium, which is characterized in that be stored with computer on the computer readable storage medium
Program, the computer program are realized as described in any one of claim 1 to 7 based on improved figure point when being executed by processor
Section model single camera multi-object tracking method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910503562.0A CN110349181B (en) | 2019-06-12 | 2019-06-12 | Single-camera multi-target tracking method based on improved graph partitioning model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910503562.0A CN110349181B (en) | 2019-06-12 | 2019-06-12 | Single-camera multi-target tracking method based on improved graph partitioning model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110349181A true CN110349181A (en) | 2019-10-18 |
| CN110349181B CN110349181B (en) | 2021-04-06 |
Family
ID=68181889
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910503562.0A Expired - Fee Related CN110349181B (en) | 2019-06-12 | 2019-06-12 | Single-camera multi-target tracking method based on improved graph partitioning model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110349181B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111768427A (en) * | 2020-05-07 | 2020-10-13 | 普联国际有限公司 | Multi-moving-target tracking method and device and storage medium |
| CN111814885A (en) * | 2020-07-10 | 2020-10-23 | 云从科技集团股份有限公司 | Method, system, device and medium for managing image frames |
| CN112784738A (en) * | 2021-01-21 | 2021-05-11 | 上海云从汇临人工智能科技有限公司 | Moving object detection alarm method, device and computer readable storage medium |
| CN113536862A (en) * | 2020-04-21 | 2021-10-22 | 北京爱笔科技有限公司 | Clustering method, device, equipment and storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8391548B1 (en) * | 2008-05-21 | 2013-03-05 | University Of Southern California | Tracking multiple moving targets in digital video |
| CN104200488A (en) * | 2014-08-04 | 2014-12-10 | 合肥工业大学 | Multi-target tracking method based on graph representation and matching |
| WO2016077026A1 (en) * | 2014-11-12 | 2016-05-19 | Nec Laboratories America, Inc. | Near-online multi-target tracking with aggregated local flow descriptor (alfd) |
| CN106600631A (en) * | 2016-11-30 | 2017-04-26 | 郑州金惠计算机系统工程有限公司 | Multiple target tracking-based passenger flow statistics method |
| CN107545582A (en) * | 2017-07-04 | 2018-01-05 | 深圳大学 | Video multi-target tracking and device based on fuzzy logic |
| CN108765459A (en) * | 2018-04-18 | 2018-11-06 | 中国人民解放军国防科技大学 | Semi-online visual multi-target tracking method based on small trajectory graph association model |
| CN109859245A (en) * | 2019-01-22 | 2019-06-07 | 深圳大学 | Multi-object tracking method, device and the storage medium of video object |
-
2019
- 2019-06-12 CN CN201910503562.0A patent/CN110349181B/en not_active Expired - Fee Related
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8391548B1 (en) * | 2008-05-21 | 2013-03-05 | University Of Southern California | Tracking multiple moving targets in digital video |
| CN104200488A (en) * | 2014-08-04 | 2014-12-10 | 合肥工业大学 | Multi-target tracking method based on graph representation and matching |
| WO2016077026A1 (en) * | 2014-11-12 | 2016-05-19 | Nec Laboratories America, Inc. | Near-online multi-target tracking with aggregated local flow descriptor (alfd) |
| CN106600631A (en) * | 2016-11-30 | 2017-04-26 | 郑州金惠计算机系统工程有限公司 | Multiple target tracking-based passenger flow statistics method |
| CN107545582A (en) * | 2017-07-04 | 2018-01-05 | 深圳大学 | Video multi-target tracking and device based on fuzzy logic |
| CN108765459A (en) * | 2018-04-18 | 2018-11-06 | 中国人民解放军国防科技大学 | Semi-online visual multi-target tracking method based on small trajectory graph association model |
| CN109859245A (en) * | 2019-01-22 | 2019-06-07 | 深圳大学 | Multi-object tracking method, device and the storage medium of video object |
Non-Patent Citations (1)
| Title |
|---|
| ERGYS RISTANI ET AL.: "tracking multiple people online and in real time", 《COMPUTER VISION -- ACCV 2014》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113536862A (en) * | 2020-04-21 | 2021-10-22 | 北京爱笔科技有限公司 | Clustering method, device, equipment and storage medium |
| CN111768427A (en) * | 2020-05-07 | 2020-10-13 | 普联国际有限公司 | Multi-moving-target tracking method and device and storage medium |
| CN111768427B (en) * | 2020-05-07 | 2023-12-26 | 普联国际有限公司 | Multi-moving-object tracking method, device and storage medium |
| CN111814885A (en) * | 2020-07-10 | 2020-10-23 | 云从科技集团股份有限公司 | Method, system, device and medium for managing image frames |
| CN111814885B (en) * | 2020-07-10 | 2021-06-22 | 云从科技集团股份有限公司 | Method, system, device and medium for managing image frames |
| CN112784738A (en) * | 2021-01-21 | 2021-05-11 | 上海云从汇临人工智能科技有限公司 | Moving object detection alarm method, device and computer readable storage medium |
| CN112784738B (en) * | 2021-01-21 | 2023-09-19 | 上海云从汇临人工智能科技有限公司 | Moving object detection alarm method, moving object detection alarm device and computer readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110349181B (en) | 2021-04-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110349181A (en) | One kind being based on improved figure partition model single camera multi-object tracking method | |
| CN107967451B (en) | Method for counting crowd of still image | |
| Hu et al. | A novel object tracking algorithm by fusing color and depth information based on single valued neutrosophic cross-entropy | |
| US20130272573A1 (en) | Multi-view object detection using appearance model transfer from similar scenes | |
| Nguyen et al. | Anomaly detection in traffic surveillance videos with gan-based future frame prediction | |
| Lamas et al. | Human pose estimation for mitigating false negatives in weapon detection in video-surveillance | |
| CN108986453A (en) | A kind of traffic movement prediction method based on contextual information, system and device | |
| CN105654516B (en) | Satellite image based on target conspicuousness is to ground weak moving target detection method | |
| CN110264495A (en) | A kind of method for tracking target and device | |
| Gan et al. | Online CNN-based multiple object tracking with enhanced model updates and identity association | |
| CN103150552B (en) | A kind of driving training management method based on number of people counting | |
| CN116128932A (en) | Multi-target tracking method | |
| CN110531618A (en) | Closed loop based on effective key frame detects robot self-localization error cancelling method | |
| Gündüz et al. | Efficient multi-object tracking by strong associations on temporal window | |
| Valappil et al. | CNN-SVM based vehicle detection for UAV platform | |
| Cheng et al. | YOLOv3 Object Detection Algorithm with Feature Pyramid Attention for Remote Sensing Images. | |
| Li et al. | Object tracking in satellite videos: Correlation particle filter tracking method with motion estimation by Kalman filter | |
| Fürst et al. | Lrpd: Long range 3d pedestrian detection leveraging specific strengths of lidar and rgb | |
| CN116363694A (en) | Multi-target tracking method of unmanned system crossing cameras matched with multiple pieces of information | |
| Pang et al. | Federated Learning for Crowd Counting in Smart Surveillance Systems | |
| CN112287906B (en) | Template matching tracking method and system based on depth feature fusion | |
| CN111598158B (en) | Warehouse inventory state sensing method based on AI analysis technology | |
| CN112418149A (en) | Abnormal behavior detection method based on deep convolutional neural network | |
| CN110889347A (en) | Density traffic flow counting method and system based on space-time counting characteristics | |
| Fan et al. | Video anomaly detection using CycleGan based on skeleton features |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210406 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |