CN107993250A - A kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus - Google Patents
A kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus Download PDFInfo
- Publication number
- CN107993250A CN107993250A CN201710822109.7A CN201710822109A CN107993250A CN 107993250 A CN107993250 A CN 107993250A CN 201710822109 A CN201710822109 A CN 201710822109A CN 107993250 A CN107993250 A CN 107993250A
- Authority
- CN
- China
- Prior art keywords
- picture
- region
- frame
- feature
- target data
- 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
- 238000004458 analytical method Methods 0.000 title claims abstract description 10
- 238000001914 filtration Methods 0.000 claims abstract description 53
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 35
- 238000000034 method Methods 0.000 claims abstract description 24
- 230000004044 response Effects 0.000 claims abstract description 13
- 238000000605 extraction Methods 0.000 claims description 25
- 230000008859 change Effects 0.000 claims description 6
- 230000021615 conjugation Effects 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 5
- 230000007935 neutral effect Effects 0.000 claims description 2
- 230000008569 process Effects 0.000 abstract description 10
- 238000012549 training Methods 0.000 description 11
- 238000001514 detection method Methods 0.000 description 7
- 241000208340 Araliaceae Species 0.000 description 6
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 6
- 235000003140 Panax quinquefolius Nutrition 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 235000008434 ginseng Nutrition 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000013441 quality evaluation Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 230000036632 reaction speed Effects 0.000 description 1
- 239000004575 stone Substances 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4084—Scaling of whole images or parts thereof, e.g. expanding or contracting in the transform domain, e.g. fast Fourier transform [FFT] domain scaling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/262—Analysis of motion using transform domain methods, e.g. Fourier domain methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus, by the first frame, carrying out sectional drawing to object frame region and obtaining target data picture；Target data picture is extracted to the feature of target data picture by convolutional neural networks；Target data filtering parameter is calculated by wave filter according to the feature of the target data picture and obtains prediction block；In the next frame, sectional drawing is carried out to picture by the prediction block and obtains region of search picture, the feature of region of search picture is extracted by convolutional neural networks；Region of search filtering parameter is calculated by wave filter according to the feature of described search region picture；According to described search region filtering parameter obtain response maximum region as final position frame and etc. be trained, and then obtain an efficiently quick CNN tracing model.Make the computing architecture process that repetition is eliminated in practical application, solve the problems, such as and fast track accurate for multiple target pedestrian and analysis.
Description
Technical field
The present invention relates to a kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus.
Background technology
At present, more there are network redundancy in present deep learning tracing algorithm, tracking speed is slow, and model is big, it is difficult to
Practicality, can not real-time tracing, the problems such as it is even more impossible to carry out multi-target tracking.To recalculate and carry by model algorithm every time
Feature is taken, and then can be formed and cause computing repeatedly, makes reaction speed significantly limited.So it is badly in need of a kind of regarding based on correlation filtering
Frequency object tracking method and its intelligent apparatus, are asked to solve the real accurate and fast track now for multiple target pedestrian with what is analyzed
Topic.
The content of the invention
The technical problem to be solved in the present invention is solving, traditional tracing algorithm network redundancy is more, and tracking speed is slow, model
Greatly, it is difficult to it is practical, can not real-time tracing, it is even more impossible to carry out the technical problem of multi-target tracking.
Above-mentioned technical problem is solved, the present invention provides a kind of object video method for tracing based on correlation filtering, it is wrapped
Include：
Sequence of frames of video is zoomed into same scale；
In the first frame, sectional drawing is carried out to object frame region and obtains target data picture；
Target data picture is extracted to the feature of target data picture by convolutional neural networks；
Target data filtering parameter is calculated by wave filter according to the feature of the target data picture, according to the mesh
Mark data filtering parameter obtains prediction block；
In the next frame, sectional drawing is carried out to picture by the prediction block and obtains region of search picture；
Region of search picture is extracted to the feature of region of search picture by convolutional neural networks；
Region of search filtering parameter is calculated by wave filter according to the feature of described search region picture；
Response maximum region is obtained as final position frame according to described search region filtering parameter.
Further, obtained according to the feature of described search region pictureAnd in the target at the heart journey is high
The responsive tags y ∈ R of this shapeM×N, wherein, M represents the width of Spatial Dimension, and N represents the height of Spatial Dimension, D representative feature passages；
According to formulaFiltering parameter is obtained, wherein, wlIt is the wave filter ginseng of passage l
Number, ★ represent circulation relevant operation, and parameter lambda >=0 is regularization parameter,Represent the discrete Fourier transform of y, y*Represent plural number y
Conjugation；
According to formulaResponse maximum region is obtained as final position frame.
Further, the convolutional neural networks are the twin convolutional neural networks of three layers of shared parameter without zero padding.
Further, the frame picture is amplified 2.5 times before the object frame region progress sectional drawing, the prediction block is to picture
The frame picture is amplified 2.5 times before progress sectional drawing.
Present invention also provides a kind of intelligent apparatus, it includes：
Unit for scaling, for sequence of frames of video to be zoomed to same scale；
Sectional drawing unit, in the first frame, carrying out sectional drawing to object frame region and obtaining target data picture；
Extraction unit, for target data picture to be extracted to the feature of target data picture by convolutional neural networks；
Computing unit, target data filtering ginseng is calculated for the feature according to the target data picture by wave filter
Number, prediction block is obtained according to the target data filtering parameter；
The sectional drawing unit, is additionally operable in the next frame, and carrying out sectional drawing to picture by the prediction block obtains the field of search
Domain picture；
The extraction unit, is additionally operable to extract region of search picture by convolutional neural networks the spy of region of search picture
Sign；
The computing unit, is additionally operable to calculate region of search by wave filter according to the feature of described search region picture
Filtering parameter；
Acquiring unit, for obtaining corresponding maximum region as final position frame according to described search region filtering parameter.
Further, the extraction unit, the feature for being additionally operable to described search region picture obtainAnd
In the target at the heart Cheng Gaosi shapes responsive tags y ∈ RM×N, wherein, M represents the width of Spatial Dimension, and N represents Spatial Dimension
Height, D representative feature passages；
The computing unit is additionally operable to according to formulaFiltering parameter is obtained, wherein, wl
It is the filter parameter of passage l, ★ represents circulation relevant operation, and parameter lambda >=0 is regularization parameter,Represent the direct computation of DFT of y
Leaf transformation, y*The conjugation of plural number y is represented, ⊙ represents the multiplication of matrix correspondence position；
The acquiring unit, is additionally operable to according to formulaObtain response maximum region.
Further, the extraction unit, the convolutional neural networks for extracting feature are three layers of being total to without zero padding
Enjoy the twin convolutional neural networks of parameter.
Further, the unit for scaling, is additionally operable to amplify the first frame picture before the object frame region carries out sectional drawing
2.5 times, the prediction block by the second frame picture before picture progress sectional drawing to amplifying 2.5 times.
Beneficial effects of the present invention：
1. speed is fast：Reach the tracking speed to more than single goal 100fps on i5cpu, and the video frame of current main-stream
The 25fps of rate, therefore algorithm can carry out real-time tracing to the target object in video.
2. model is small：Model size is 76k, and as the system based on convolutional neural networks, the model of lightweight causes this
System can be readily used in embedded device.
3. accuracy rate is high：It is fast in speed, on the basis of model is small, change standard data set OTB of the tracing model in tracking
Reach very high accuracy rate in (object tracking benchmark) and VOT (visual object tracking),
The pedestrian being entirely capable of suitable for reality follows the trail of scene.
4. multi-target tracking：The tracing model of mainstream is difficult to be tracked multiple target at present, the tracking that the present invention designs
System can be tracked multiple targets in real time.
5. training method end to end：The training frame of the tracing system can combine convolutional neural networks and related filter
Ripple, reduces and adjusts the cumbersome of ginseng manually, and can reach more preferably performance.
Brief description of the drawings
Fig. 1 is the flow chart of the object video method for tracing to correlation filtering of one embodiment of the application；
Fig. 2 is the Organization Chart of the object video tracking intelligent apparatus to correlation filtering of another embodiment of the application；
Fig. 3 is the application training process overall flow schematic diagram；
Fig. 4 is the flow diagram of the application first embodiment concrete application；
Fig. 5 is the flow diagram of the application second embodiment concrete application；
Fig. 6 is the first schematic diagram of use state of the application image concrete application；
Fig. 7 is the second schematic diagram of use state of the application image concrete application；
Fig. 8 is the 3rd schematic diagram of use state of the application image concrete application；
Embodiment：
Following embodiments are only clearly to invent this example, and not to the limit of embodiments of the present invention
It is fixed.For those of ordinary skill in the field, it is various forms of that other can also be made on the basis of the following description
Change changes, and these belong to protection of the spiritual obvious changes or variations drawn of the invention still in the present invention
Among scope.
Step shown in the flowchart of the accompanying drawings can be in the computer system of such as a group of computer-executable instructions
Perform.Also, although logical order is shown in flow charts, in some cases, can be with suitable different from herein
Sequence performs shown or described step.
More there are network redundancy in present deep learning tracing algorithm, tracking speed is slow, and model is big, it is difficult to and it is practical,
Can not real-time tracing, the problems such as multi-target tracking can not be carried out.In the present invention, by the network structure of lightweight, base is passed through
In the end-to-end training method of the correlation filtering of deep learning, the shortcomings that traditional filtering method needs to adjust ginseng manually is avoided.Pass through
The twin deep learning network of the lightweight of one 3 layers of design carries out feature extraction, while ensure that speed, is substantially reduced
Model capacity.Multiscale analysis is carried out to feature using correlation filtering, can accurately export the position of target.
As shown in figure 3, process is divided into data processing, feature extraction, three parts of correlation filtering are discussed：
1. data processing
Sequence of frames of video is zoomed into same scale, for training picture, centered on the position of object frame, frame is amplified
2.5 times of region carries out sectional drawing, as training picture.Test when, the first frame according to object frame amplify 2.5 times region into
Row sectional drawing, by the sectional drawing of generation by convolutional neural networks, calculates filter parameter.Since the second frame, previous frame is utilized
Prediction block position amplify 2.5 times to the frame carry out sectional drawing, sectional drawing is equally passed through into the filter of previous frame by convolutional neural networks
Ripple device parameter calculates the position of final frame.In tracing process, the renewal of device parameter is filtered every 5 frames so that wave filter
The change of the form light and ambient background of target can constantly be adapted to.
2. feature extraction
In the training process, using the twin convolutional neural networks of three layers of shared parameter without zero padding respectively to target data
Feature extraction is carried out with region of search.In this way, carry out correlation filtering meter with regard to two eigenmatrixes can be obtained, then by eigenmatrix
Calculate, training label is the Gauss shape characteristic pattern of one and input data formed objects, and the response in target area is maximum.Testing
During, amplify 2.5 times of progress sectional drawings according only to the prediction block position of previous frame, region of search is obtained, by region of search picture
Data are input in three-layer coil product neutral net, are obtained final eigenmatrix, by feature convolution by correlation filter, are obtained
The final prediction block position of the frame.
3. correlation filtering
By the feature extraction in 2, obtainIn the target at the heart Cheng Gaosi shapes responsive tags y ∈
RM×N, wherein M represents the width of Spatial Dimension, and N represents the height of Spatial Dimension, and D representative feature passages, the purpose of correlation filtering is to look for
To a filtering parameter so that the loss function ε of ridge regression is minimum：
Wherein, wlIt is the filter parameter of passage l.★ represents circulation relevant operation, i.e., cyclic shift behaviour is to w and first
Make, carrying out related operation, parameter lambda >=0 is regularization parameter.The solution of the equation can be expressed as：
Here,Represent the discrete Fourier transform of y.y*Represent the conjugation of plural number y.⊙ represents the multiplication of matrix correspondence position.
In detection process, we carry out feature extraction for region of search, obtainPass through the following formula：
The region for the maximum that meets with a response.
In its another embodiment, according to Fig. 1, the present invention provides a kind of object video based on correlation filtering
Method for tracing, it includes：
S101, same scale is zoomed to by sequence of frames of video；
S102, in the first frame, carries out sectional drawing to object frame region and obtains target data picture；
Region of search, is extracted the feature of target data picture by S103 by convolutional neural networks；
S104, target data filtering parameter is calculated according to the feature of the target data picture by wave filter, according to
The target data filtering parameter obtains prediction block；
S105, in the next frame, carries out sectional drawing to picture by the prediction block and obtains region of search picture；
Region of search picture, is extracted the feature of region of search picture by S106 by convolutional neural networks；
S107, region of search filtering parameter is calculated according to the feature of described search region picture by wave filter；
S108, response maximum region is obtained as final position frame according to described search region filtering parameter.
First, video is all zoomed into same scale per frame, sectional drawing is carried out centered on the position of object frame, as training
Picture.Feature is extracted by convolutional neural networks to the training picture sectional drawing (i.e. above-mentioned target data picture) of generation, according to
The feature of extraction is filtered device calculating, target data filtering parameter is calculated, according to the feelings of the target data filtering parameter
Condition speculates to obtain a prediction block, which is to think that the region of maximum probability occurs in target in a hypothesis, is passed through
This region further carries out sectional drawing in the next frame.Region of search picture is obtained, convolutional Neural is passed through to region of search picture
Network carries out extraction feature, and region of search filtering parameter is calculated again by wave filter, is filtered and joined according to described search region
Number obtains response maximum region as final position frame.In the above process, the renewal of device parameter is filtered every 5 frames so that filter
Ripple device can constantly adapt to the change of the form light and ambient background of target.
In another alternative embodiment, how will specifically calculate and be illustrated, this method further includes：
Obtained according to the feature of described search region pictureAnd in the target at the heart Cheng Gaosi shapes response
Label y ∈ RM×N, wherein, M represents the width of Spatial Dimension, and N represents the height of Spatial Dimension, D representative feature passages；
According to formulaFiltering parameter is obtained, wherein, wlIt is the wave filter ginseng of passage l
Number, ★ represent circulation relevant operation, and parameter lambda >=0 is regularization parameter,Represent the discrete Fourier transform of y, y*Represent plural number y
Conjugation；⊙ represents the multiplication of matrix correspondence position.Wherein, ★ represents circulation relevant operation and circulative shift operation is done to w and first
And calculate.
According to formulaObtain response maximum region.
Secondly, in fact twice extraction with calculate filtering parameter process be as, difference is the second extraction feature
When, the maximum region of response can be also obtained as final frame position.
In another alternative embodiment, as shown in figure 3, this method further includes：
Again, the convolutional neural networks are the twin convolutional neural networks of three layers of shared parameter without zero padding, three layers of this kind
Structure is the convolutional neural networks of lightweight.Twin is two convolutional neural networks, passes through filter after carrying out extraction feature respectively
Ripple device, which calculate, obtains filtering parameter, and then obtains the specific region of frame.
In another alternative embodiment, this method further includes：
The frame picture is amplified 2.5 times before the object frame region progress sectional drawing, the prediction block carries out sectional drawing to picture
It is preceding that the frame picture is amplified 2.5 times.The feature extraction after being captured as again is amplified to it to lay the first stone.It has passed through
After above-mentioned training process, an efficiently quick CNN tracing model (framework of feature) is trained.Pass through above-mentioned tracing model
In service stage afterwards, directly by carrying out feature extraction framework to processing picture, process is recalculated among eliminating, greatly
Width improves work efficiency.
Above-mentioned model is illustrated the embodiment after it as follows：
Example 1：Tracing system as shown in Figure 4, first, the target location of the first frame is detected using target detection technique, will
Multiple targets to be followed the trail of are added in tracking queue.Next frame picture is inputted, then traversal tracking queue, for each tracking
Object invocation tracing algorithm obtains the position of the target in the next frame.The target is obtained after the position of next frame, is passed through
Whether the threshold decision target frames out.If the target have left screen, which is removed into tracking object queue.
Every 24 frames, a target detection is called, the result of target detection is calculated into IOU with the result of tracking, if mesh
Mark the IOU ＜ 0.1 of some result of detection and all targets of tracking, then it is assumed that new target adds screen, by the target
Add in tracking queue.If IOU ＞ 0.5, the frame of tracking is substituted using the frame of target detection, carries out position correction.Wherein,
IOU(intersection over Union)：Hand over and compare, two intersection of sets collection divided by two union of sets collection.
It is (meeting either condition) to judge the condition whether target frames out：
Perdict_score ＜ threshold
H/w ＞ threshold1
W/h ＞ threshold2
| x1 |/W ＜ threshold3
| W-x2 |/W ＜ threshold3
| y1 |/H ＜ threshold4
| H-y1 |/H ＜ threshold4
Wherein, h and w is respectively the height and width of object, and H and W are respectively the height and width of frame.(x1, y1) is the target upper left corner
Point coordinates, (x2, y2) be the target lower right corner point coordinates.
Example 2：Quality evaluation system as shown in Figure 5.Our tracking can follow the trail of a target for a long time, and
There are numerous redundancies between video frame and frame, therefore, quality is most within a period of time for the simply target that we are put in storage
High, a most representative figure.
As shown in figs 6-8, it is video that one section of pedestrian goes across the road, three figures are the first frame, the 20th frame, and the 40th frame respectively
Design sketch, we first to the first frame carry out pedestrian detection, the pedestrian target is tracked inside ensuing frame.One
In the section time, call our quality evaluation algorithm to be assessed, select pedestrian's result storage.Here since all frames should
Target is all angle of leaning to one side, therefore we choose the highest small picture storage of a quality score.
Present invention also provides a kind of intelligent apparatus, as shown in Fig. 2, it includes：
Unit for scaling, for sequence of frames of video to be zoomed to same scale；
Sectional drawing unit, in the first frame, carrying out sectional drawing to object frame region and obtaining target data picture；
Extraction unit, for target data picture to be extracted to the feature of target data picture by convolutional neural networks；
Computing unit, target data filtering ginseng is calculated for the feature according to the target data picture by wave filter
Number, prediction block is obtained according to the target data filtering parameter；
The sectional drawing unit, is additionally operable in the next frame, and carrying out sectional drawing to picture by the prediction block obtains the field of search
Domain picture；
The extraction unit, is additionally operable to extract region of search picture by convolutional neural networks the spy of region of search picture
Sign；
The computing unit, is additionally operable to calculate region of search by wave filter according to the feature of described search region picture
Filtering parameter；
Acquiring unit, for obtaining response maximum region as final position frame according to described search region filtering parameter.
Further, the extraction unit, the feature for being additionally operable to described search region picture obtainAnd
In the target at the heart Cheng Gaosi shapes responsive tags y ∈ RM×N, wherein, M represents the width of Spatial Dimension, and N represents Spatial Dimension
Height, D representative feature passages；
The computing unit, is additionally operable to according to formulaFiltering parameter is obtained, wherein, wl
It is the filter parameter of passage l, ★ represents circulation relevant operation, and parameter lambda >=0 is regularization parameter,Represent the direct computation of DFT of y
Leaf transformation, y*Represent the conjugation of plural number y；
The acquiring unit, is additionally operable to according to formulaObtain response maximum region.
Further, the extraction unit, the convolutional neural networks for extracting feature are three layers of being total to without zero padding
Enjoy the twin convolutional neural networks of parameter.
Further, the unit for scaling, is additionally operable to amplify the first frame picture before the object frame region carries out sectional drawing
2.5 times, the prediction block by the second frame picture before picture progress sectional drawing to amplifying 2.5 times.
Although disclosed herein embodiment as above, the content be only readily appreciate the present invention and use
Embodiment, is not limited to the present invention.Technical staff in any fields of the present invention, is taken off not departing from the present invention
On the premise of the spirit and scope of dew, any modification and change, but the present invention can be carried out in the form and details of implementation
Scope of patent protection, still should be subject to the scope of the claims as defined in the appended claims.
Claims (8)
1. a kind of fast multi-target pedestrian tracking and analysis method, it is characterised in that this method includes：
Sequence of frames of video is zoomed into same scale；
In the first frame, sectional drawing is carried out to object frame region and obtains target data picture；
Target data picture is extracted to the feature of target data picture by convolutional neural networks；
Target data filtering parameter is calculated by wave filter according to the feature of the target data picture, according to the number of targets
Prediction block is obtained according to filtering parameter；
In the next frame, sectional drawing is carried out to picture by the prediction block and obtains region of search picture；
Region of search picture is extracted to the feature of region of search picture by convolutional neural networks；
Region of search filtering parameter is calculated by wave filter according to the feature of described search region picture；
Corresponding maximum region is obtained as final position frame according to described search region filtering parameter.
2. fast multi-target pedestrian tracking according to claim 1 and analysis method, according to described search region picture
Feature calculates region of search filtering parameter by wave filter, and corresponding maximum region is obtained according to described search region filtering parameter
As final position frame, it is characterised in that this method further includes：
Obtained according to the feature of described search region pictureAnd in the target at the heart Cheng Gaosi shapes responsive tags
y∈RM×N, wherein, M represents the width of Spatial Dimension, and N represents the height of Spatial Dimension, D representative feature passages；
According to formulaFiltering parameter is obtained, wherein, wlIt is the filter parameter of passage l, ★
Circulation relevant operation is represented, parameter lambda >=0 is regularization parameter,Represent the discrete Fourier transform of y, y*Represent being total to for plural number y
Yoke, ⊙ represent the multiplication of matrix correspondence position；
According to formulaResponse maximum region is obtained as final position frame.
3. fast multi-target pedestrian tracking according to claim 1 and analysis method, it is characterised in that this method is also wrapped
Include：
The convolutional neural networks are the twin convolutional neural networks of three layers of shared parameter without zero padding.
4. fast multi-target pedestrian tracking according to claim 1 and analysis method, it is characterised in that this method is also wrapped
Include：
The frame picture is amplified 2.5 times before the object frame region progress sectional drawing；
The prediction block by the frame picture before picture progress sectional drawing to amplifying 2.5 times.
5. a kind of intelligent apparatus, it is characterised in that it includes：
Unit for scaling, for sequence of frames of video to be zoomed to same scale；
Sectional drawing unit, in the first frame, carrying out sectional drawing to object frame region and obtaining target data picture；
Extraction unit, for target data picture to be extracted to the feature of target data picture by convolutional neural networks；
Computing unit, for calculating target data filtering parameter by wave filter according to the feature of the target data picture,
Prediction block is obtained according to the target data filtering parameter；
The sectional drawing unit, is additionally operable in the next frame, and carrying out sectional drawing to picture by the prediction block obtains region of search figure
Piece；
The extraction unit, is additionally operable to extract region of search picture by convolutional neural networks the feature of region of search picture；
The computing unit, is additionally operable to calculate region of search filtering by wave filter according to the feature of described search region picture
Parameter；
Acquiring unit, for obtaining response maximum region as final position frame according to described search region filtering parameter.
6. intelligent apparatus according to claim 5, it is characterised in that it is further included：
The extraction unit, the feature for being additionally operable to described search region picture obtainAnd journey at the heart in the target
The responsive tags y ∈ R of Gauss shapeM×N, wherein, M represents the width of Spatial Dimension, and N represents the height of Spatial Dimension, and D representative features lead to
Road；
The computing unit, is additionally operable to according to formulaFiltering parameter is obtained, wherein, wlIt is logical
The filter parameter of road l, ★ represent circulation relevant operation, and parameter lambda >=0 is regularization parameter,The discrete fourier for representing y becomes
Change, y*The conjugation of plural number y is represented, ⊙ represents the multiplication of matrix correspondence position；
The acquiring unit, is additionally operable to according to formulaResponse maximum region is obtained as final position
Frame.
7. intelligent apparatus according to claim 5, it is characterised in that it is further included：
The extraction unit, the convolutional neural networks for extracting feature are the twin convolution of three layers of shared parameter without zero padding
Neutral net.
8. intelligent apparatus according to claim 5, it is characterised in that it includes：
The unit for scaling, is additionally operable to that the first frame picture is amplified 2.5 times before the object frame region progress sectional drawing, the prediction
Frame by the second frame picture before picture progress sectional drawing to amplifying 2.5 times.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710822109.7A CN107993250A (en) | 2017-09-12 | 2017-09-12 | A kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710822109.7A CN107993250A (en) | 2017-09-12 | 2017-09-12 | A kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107993250A true CN107993250A (en) | 2018-05-04 |
Family
ID=62029754
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710822109.7A Pending CN107993250A (en) | 2017-09-12 | 2017-09-12 | A kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107993250A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108830170A (en) * | 2018-05-24 | 2018-11-16 | 杭州电子科技大学 | A kind of end-to-end method for tracking target indicated based on layered characteristic |
| CN108846344A (en) * | 2018-06-05 | 2018-11-20 | 中南大学 | A kind of pedestrian's posture multiple features INTELLIGENT IDENTIFICATION method merging deep learning |
| CN109063759A (en) * | 2018-07-20 | 2018-12-21 | 浙江大学 | A kind of neural network structure searching method applied to the more attribute forecasts of picture |
| CN109064493A (en) * | 2018-08-01 | 2018-12-21 | 北京飞搜科技有限公司 | A kind of method for tracking target and device based on meta learning |
| CN109101932A (en) * | 2018-08-17 | 2018-12-28 | 佛山市顺德区中山大学研究院 | The deep learning algorithm of multitask and proximity information fusion based on target detection |
| CN109446889A (en) * | 2018-09-10 | 2019-03-08 | 北京飞搜科技有限公司 | Object tracking method and device based on twin matching network |
| CN109727272A (en) * | 2018-11-20 | 2019-05-07 | 南京邮电大学 | A kind of method for tracking target based on double branch's space-time regularization correlation filters |
| CN109727246A (en) * | 2019-01-26 | 2019-05-07 | 福州大学 | Comparative learning image quality evaluation method based on twin network |
| CN109886996A (en) * | 2019-01-15 | 2019-06-14 | 东华大学 | A kind of visual pursuit optimization method |
| CN110633594A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Target detection method and device |
| CN111951298A (en) * | 2020-06-25 | 2020-11-17 | 湖南大学 | Target tracking method fusing time series information |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1924894A (en) * | 2006-09-27 | 2007-03-07 | 北京中星微电子有限公司 | Multiple attitude human face detection and track system and method |
| CN103679687A (en) * | 2012-09-18 | 2014-03-26 | 杭州海康威视数字技术股份有限公司 | Target tracking method of intelligent tracking high-speed dome camera |
| CN103985136A (en) * | 2014-03-21 | 2014-08-13 | 南京大学 | Target tracking method based on local feature point feature flow pattern |
| CN106875425A (en) * | 2017-01-22 | 2017-06-20 | 北京飞搜科技有限公司 | A kind of multi-target tracking system and implementation method based on deep learning |
| CN106909625A (en) * | 2017-01-20 | 2017-06-30 | 清华大学 | A kind of image search method and system based on Siamese networks |
-
2017
- 2017-09-12 CN CN201710822109.7A patent/CN107993250A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1924894A (en) * | 2006-09-27 | 2007-03-07 | 北京中星微电子有限公司 | Multiple attitude human face detection and track system and method |
| CN103679687A (en) * | 2012-09-18 | 2014-03-26 | 杭州海康威视数字技术股份有限公司 | Target tracking method of intelligent tracking high-speed dome camera |
| CN103985136A (en) * | 2014-03-21 | 2014-08-13 | 南京大学 | Target tracking method based on local feature point feature flow pattern |
| CN106909625A (en) * | 2017-01-20 | 2017-06-30 | 清华大学 | A kind of image search method and system based on Siamese networks |
| CN106875425A (en) * | 2017-01-22 | 2017-06-20 | 北京飞搜科技有限公司 | A kind of multi-target tracking system and implementation method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| QIANG WANG, JIN GAO, JUNLIANG XING, MENGDAN ZHANG, WEIMING HU: ""DCFNET: DISCRIMINANT CORRELATION FILTERS NETWORK FOR VISUAL TRACKING"", 《ARXIV》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108830170A (en) * | 2018-05-24 | 2018-11-16 | 杭州电子科技大学 | A kind of end-to-end method for tracking target indicated based on layered characteristic |
| CN108830170B (en) * | 2018-05-24 | 2022-03-18 | 杭州电子科技大学 | End-to-end target tracking method based on layered feature representation |
| CN108846344A (en) * | 2018-06-05 | 2018-11-20 | 中南大学 | A kind of pedestrian's posture multiple features INTELLIGENT IDENTIFICATION method merging deep learning |
| CN108846344B (en) * | 2018-06-05 | 2022-05-17 | 中南大学 | Pedestrian posture multi-feature intelligent identification method integrating deep learning |
| CN110633594A (en) * | 2018-06-21 | 2019-12-31 | 北京京东尚科信息技术有限公司 | Target detection method and device |
| CN109063759A (en) * | 2018-07-20 | 2018-12-21 | 浙江大学 | A kind of neural network structure searching method applied to the more attribute forecasts of picture |
| CN109064493A (en) * | 2018-08-01 | 2018-12-21 | 北京飞搜科技有限公司 | A kind of method for tracking target and device based on meta learning |
| CN109064493B (en) * | 2018-08-01 | 2021-03-09 | 苏州飞搜科技有限公司 | Target tracking method and device based on meta-learning |
| CN109101932B (en) * | 2018-08-17 | 2020-07-24 | 佛山市顺德区中山大学研究院 | Multi-task and proximity information fusion deep learning method based on target detection |
| CN109101932A (en) * | 2018-08-17 | 2018-12-28 | 佛山市顺德区中山大学研究院 | The deep learning algorithm of multitask and proximity information fusion based on target detection |
| CN109446889A (en) * | 2018-09-10 | 2019-03-08 | 北京飞搜科技有限公司 | Object tracking method and device based on twin matching network |
| CN109446889B (en) * | 2018-09-10 | 2021-03-09 | 苏州飞搜科技有限公司 | Object tracking method and device based on twin matching network |
| CN109727272A (en) * | 2018-11-20 | 2019-05-07 | 南京邮电大学 | A kind of method for tracking target based on double branch's space-time regularization correlation filters |
| CN109727272B (en) * | 2018-11-20 | 2022-08-12 | 南京邮电大学 | Target tracking method based on double-branch space-time regularization correlation filter |
| CN109886996A (en) * | 2019-01-15 | 2019-06-14 | 东华大学 | A kind of visual pursuit optimization method |
| CN109886996B (en) * | 2019-01-15 | 2023-06-06 | 东华大学 | Visual tracking optimization method |
| CN109727246A (en) * | 2019-01-26 | 2019-05-07 | 福州大学 | Comparative learning image quality evaluation method based on twin network |
| CN109727246B (en) * | 2019-01-26 | 2022-05-13 | 福州大学 | Comparative learning image quality evaluation method based on twin network |
| CN111951298A (en) * | 2020-06-25 | 2020-11-17 | 湖南大学 | Target tracking method fusing time series information |
| CN111951298B (en) * | 2020-06-25 | 2024-03-08 | 湖南大学 | Target tracking method integrating time sequence information |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107993250A (en) | A kind of fast multi-target pedestrian tracking and analysis method and its intelligent apparatus | |
| CN111259850B (en) | Pedestrian re-identification method integrating random batch mask and multi-scale representation learning | |
| CN106096577A (en) | Target tracking system in a kind of photographic head distribution map and method for tracing | |
| CN104573731B (en) | Fast target detection method based on convolutional neural networks | |
| CN110427905A (en) | Pedestrian tracting method, device and terminal | |
| CN109816689A (en) | A kind of motion target tracking method that multilayer convolution feature adaptively merges | |
| CN107644430A (en) | Target following based on self-adaptive features fusion | |
| CN109829443A (en) | Video behavior recognition methods based on image enhancement Yu 3D convolutional neural networks | |
| CN109271888A (en) | Personal identification method, device, electronic equipment based on gait | |
| CN108665481A (en) | Multilayer depth characteristic fusion it is adaptive resist block infrared object tracking method | |
| CN107341805A (en) | Background segment and network model training, image processing method and device before image | |
| CN107169994A (en) | Correlation filtering tracking based on multi-feature fusion | |
| CN108334847A (en) | A kind of face identification method based on deep learning under real scene | |
| CN108229282A (en) | Critical point detection method, apparatus, storage medium and electronic equipment | |
| CN109241871A (en) | A kind of public domain stream of people's tracking based on video data | |
| CN109558815A (en) | A kind of detection of real time multi-human face and tracking | |
| CN108121945A (en) | A kind of multi-target detection tracking, electronic equipment and storage medium | |
| CN108446694A (en) | A kind of object detection method and device | |
| CN108121931A (en) | two-dimensional code data processing method, device and mobile terminal | |
| CN107463881A (en) | A kind of character image searching method based on depth enhancing study | |
| CN110674886B (en) | Video target detection method fusing multi-level features | |
| Liu et al. | ACDnet: An action detection network for real-time edge computing based on flow-guided feature approximation and memory aggregation | |
| CN112766123B (en) | Crowd counting method and system based on criss-cross attention network | |
| CN108830170A (en) | A kind of end-to-end method for tracking target indicated based on layered characteristic | |
| Dong et al. | Lightweight and efficient neural network with SPSA attention for wheat ear detection |
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: 20180504 |