CN110298868A - A kind of multiscale target tracking of high real-time - Google Patents
A kind of multiscale target tracking of high real-time Download PDFInfo
- Publication number
- CN110298868A CN110298868A CN201910559301.0A CN201910559301A CN110298868A CN 110298868 A CN110298868 A CN 110298868A CN 201910559301 A CN201910559301 A CN 201910559301A CN 110298868 A CN110298868 A CN 110298868A
- Authority
- CN
- China
- Prior art keywords
- scale
- target
- frame
- node
- response
- 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
- 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
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
-
- 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
- G06T2207/20048—Transform domain processing
- G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]
-
- 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
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of multiscale target trackings with height real-time.This method includes extracting Fast Field direction histogram (fhog) feature to target and its surrounding background area, generates positive negative sample by cyclic shift simulation, two-dimensional Gaussian function is as sample label, Ridge Regression Modeling Method training correlation filter；Subsequent frame obtains target position using filter response characteristics, calculates target scale using the method combined based on scale prediction with scale pond；Repetition training process carries out interpolation update to correlation filter.The present invention utilizes target histories motion information, realize the optimization to frequency domain operation and size estimation mode, while ensureing tracking accuracy, core correlation filter (KCF) method operational efficiency can be optimized about 43%, the real-time of height is that correlation filtering class method for tracking target is provided safeguard to the transplanting of the development board of the low operational capability such as embedded system, can be applied to the fields such as intelligent monitoring, space flight and aviation, unmanned.
Description
Technical field
The invention belongs to the monotrack field of computer vision, a kind of multiscale target of high real-time especially set out
Tracking.
Background technique
Target following is to have incorporated the technologies such as automatic control and information science on the basis of image procossing, it can
The information such as the size for positioning target in each frame image, and obtaining target, thus by target area from background area
In split, and then complete being tracked in entire video sequence image to target of the task.
In recent years, the hot spot of discriminate tracking area research mainly concentrated the tracking based on correlation filter to improve
On.Correlation filtering is originally a kind of signal processing method, defeated to judge by output response of the input signal after wave filter
Enter the correlation of signal.It, can be by finding peak response, to determine the center of tracking target in target following.For
Realize it is effective update object module, quickly and accurately Scale Estimation Method is very necessary.Scale meter is solved at present
The method of calculation carries out target using translation filter mainly by way of scale pond on the image block of multiple dimensioned scaling
Detection takes the maximum translation position of response and place scale.This method is simple, it can be readily appreciated that but it is excessively coarse, lead to method
Overall complexity can increase with the increase of institute's dipping degree number and in times several levels, destroy nuclear phase and close filter tracking method itself
The superiority of high speed performance, so that reality cannot be completed when running on the development board of some low operational capabilities such as embeded processor
The requirement of when property is unable to satisfy the real landing demand of tracking technique.
In order to optimize the above-mentioned target scale estimation method based on scale pond, the redundancy present in it is reduced, tracking is improved
Method operational efficiency.Because the increase for calculating scale has powerful destructive power, therefore consider how to reduce the number of dimension calculation,
By the way that the mechanism of scale prediction to be incorporated therein, it is excellent that speed can be promoted while both ensureing tracking pinpoint accuracy
Gesture.
Summary of the invention
For the present invention in order to solve in object tracking process, the inefficient problem of size estimation proposes a kind of high real-time
Multiscale target tracking, guarantee tracking it is high-precision and meanwhile improve correlation filtering class tracking goal approach tracking speed
Degree.
To realize the above-mentioned technical purpose, the technical solution adopted by the present invention mainly comprises the steps that
Step 1, area-of-interest is determined.Target to be tracked, processing target week are chosen with rectangle frame in initial image frame
The 2.5 times of pixel regions enclosed adjust long edge size to 96, and shorter edge carries out scaling in the ratio that long side changes；
Step 2, feature is extracted.Fast Field direction histogram (fhog) feature is tieed up to region of interesting extraction 31；
Step 3, training correlation filtering tracker.Feature vector and dimensional Gaussian peak function are transformed into frequency-domain calculations, obtained
To correlation filter；
Step 4, target position is detected.It is same as step 1 extracts area-of-interest in next frame image, using trained
The correlation filtering tracker arrived, response maximum are target position；
Step 5, size estimation.Scale prediction is combined with scale pond, determines target scale.
In size estimation, priority is divided for the dimension calculation in scale pond first.It can be by every frame target scale
Change and is divided into three kinds of situations: 0.95,1,1.05, i.e., small scale, normally, large scale.Target is respectively indicated with this to become smaller, it is constant with
And the case where becoming larger.
Traditional Scale Estimation Method is that the image block of different scale is taken to extract feature respectively, is sought with correlation filter
Response, compares response, and the maximum is best scale.This method overall complexity and scale number increase in times several levels,
Thus it is affected to speed.Division response characteristic of the priority approach based on correlation filter proposed by the present invention, with target
The case where actually becoming larger citing can be reasoned out if obtained response is variable R as R (0.95) < R (1) < R (1.05).Therefore
If first calculating and comparing R (1) and R (1.05), practical maximum value that you can get it, without calculating the response to small scale again.Together
Reason, conversely, becoming smaller if target is practical, without the response for calculating large scale again.It does not change for target physical size
The case where, can assign at random first it is a kind of in above situation calculated, then calculate the response of the third scale and be compared.
The present invention uses Statistics-Based Method, and using target actual motion information, confirmation calculates certain scale first
Response.The dimensional variation situation of 10 frames of statistics past, which kind of dimensional variation is in the majority, i.e., is preferentially calculated in present frame with the scale.
No matter what kind of situation it is, it is both needed to response when calculating R (1), i.e. Scale invariant, then can be just compared.
The statistical method is completed in constant time complexity to past frame by the way of creating special circular linked list
Dimensional variation situation statistics, tracking velocity is impacted to avoid high time complexity.
In order to further increase tracking velocity, the present invention is after counting multiframe information, for scale almost without change
The target of change is tracked using in such a way that frame is multiple dimensioned.I.e. a frame is multiple dimensioned, and multiframe single scale is alternately.It both can be with
Timely update target scale information, and every frame can effectively be avoided all to carry out multiscale tracing and waste time.
The present invention has also done related optimization to redundant computation, stores and utilizes the frequency domain operation of feature vector, reduction nuclear phase
The operation times of filter Fast Fourier Transform (FFT) (FFT) are closed, optimization improves 20% operational efficiency.
Frequency domain operation proposed by the present invention and Scale Estimation Method can be applied in all correlation filtering class trackings,
And it is not limited to nuclear phase and closes filter (KCF) tracking.
Detailed description of the invention
Fig. 1: method overall flow figure
Fig. 2: circular linked list schematic diagram
Specific embodiment
The embodiment of the invention will now be described in detail with reference to the accompanying drawings.
As shown in Fig. 1 overall flow figure, the invention discloses a kind of multiscale target trackings of high real-time, specifically
The following steps are included:
Step 1, area-of-interest is determined.Target to be tracked, processing target week are chosen with rectangle frame in initial image frame
The 2.5 times of pixel regions enclosed adjust long edge size to 96, and shorter edge carries out scaling in the ratio that long side changes；
Step 2, feature is extracted.Fast Field direction histogram (fhog) feature is tieed up to region of interesting extraction 31；
Step 3, training correlation filtering tracker and redundancy optimization.Feature vector and dimensional Gaussian peak function are transformed into frequency
Domain calculates, and obtains correlation filter, training formula such as formula (1):
Wherein, α is the implicit coefficient of correlation filter, and k is gaussian kernel function variable k (x, x),
Such as (2), x is training image block eigenvector to calculation formula, and y is the label vector that two-dimensional Gaussian function is constituted, and λ is
Regularization factors.
In formula (2), F-1For the inverse transformation of FFT,For x be transformed into frequency domain as a result, ⊙ is dot-product operation, * expression takes
Conjugation, σ is bandwidth.Two equal x of input of the kernel function in training, therefore practical need to carry out a FFT to x, this can
Reduce by training 1/2 FFT calculation times.
Step 4, target position is detected.It is same as step 1 extracts area-of-interest in next frame image, using trained
The correlation filtering tracker arrived, detection formula such as formula (3):
Wherein,It being exported for the response of all positions, α is to acquire in step 3,For gaussian kernel function variable k (x, z),
Calculation formula such as (4).Wherein x is training image block eigenvector, z test image block eigenvector.Response maximum is mesh
Mark position.
Two inputs for detecting Kernel Function are respectively x and z,WithFor x and z be transformed into frequency domain as a result, multiple dimensioned
Link is calculated, detection process is repeated to the image block of different scale, i.e., only change test image block eigenvector z, therefore store
FFT (x) when scaling op for the first time, calls directly in subsequent arithmetic.
Step 5, size estimation.Scale prediction is combined with scale pond, determines target scale.
In this step size estimation, priority is divided for the dimension calculation in scale pond first.By every frame target scale
Variation be divided into three kinds of situations: 0.95,1,1.05, i.e., small scale, normally, large scale.Target is respectively indicated with this to become smaller, it is constant
And the case where becoming larger.Division response characteristic of the priority approach based on correlation filter proposed by the present invention, with target reality
The case where becoming larger citing can be reasoned out if obtained response is variable R as R (0.95) < R (1) < R (1.05).If therefore first
Calculate and compare R (1) and R (1.05), practical maximum value that you can get it, without calculating the response to small scale again.Similarly, if
Target is practical to become smaller, without the response for calculating large scale again.The case where not changing for target physical size, can be with
Assign at random first it is a kind of in above situation calculated, then calculate the response of the third scale and be compared.
The present invention uses Statistics-Based Method, and using target actual motion information, confirmation calculates certain scale first
Response.The dimensional variation situation of 10 frames of statistics past, which kind of dimensional variation is in the majority, i.e., is preferentially calculated in present frame with the scale.
No matter what kind of situation it is, it is both needed to response when calculating R (1), i.e. Scale invariant, then can be just compared.The statistical method is adopted
With the mode for creating special circular linked list, the system to the dimensional variation situation of past frame is completed in constant time complexity
Meter, impacts tracking velocity to avoid high time complexity.Circular linked list schematic diagram is as shown in Fig. 2, specifically include as follows:
Definition node first, each node have bidirectional pointer, are respectively directed to front and back node, intra-node defines mScale
Variable, for recording current scale state, state, which is divided into, to become larger (value 1), constant (value 0), becomes smaller (value is -1) and default
(value is 2).
After creating circular linked list, it is directed toward arbitrary node using pointer, every to pass through a frame, pointer moves down a node, works as frame
When number is greater than queue length, pointer is directed toward node initial in chained list, so that the nodal information is capped, realizes nearest 10 frame letter
The record of breath.
In order to reduce the time complexity of statistical yardstick information, a mSum variable is safeguarded, for recording all frame scales
The sum of state mScale, formula such as (5):
MSum=∑ mScale (mScale ≠ 2) (5)
The default situations that mScale is 2 indicate that node is also uncovered, cannot include.
Since mScale design is the statistical variable of 0 mean value, and if more than 0, then illustrate the frame that preceding 10 frame scale becomes larger
Number is more, if being equal to 0, the frame number of Scale invariant is more, if the frame number that scale becomes smaller is more less than 0, mSum variable is repaired
Change and is located at the time of this frame information is set every time, formula such as (6):
MSum=mSum-mScale (i)+mScale (j) (6)
It subtracts the dimensional information mScale (i) for currently pointing to node, in addition the dimensional information mScale (j) of present frame,
Therefore the occupied time complexity of statistical yardstick change information is constant complexity, will not reduce tracking velocity.
Use head node information realization full to team simultaneously and the judgement of team's dummy status, if head node nanoscale regime is default,
Then head node is not used by, and can determine whether team's sky；If the previous node nanoscale regime of head node be it is non-default, can determine whether that team is full.
The time complexity of two states inquiry is constant complexity, will not reduce tracking velocity.
For scale, almost unconverted target, the present invention are tracked using in such a way that frame is multiple dimensioned.I.e. a frame is more
Scale, multiframe single scale is alternately.Every 5 frame progress is primary multiple dimensioned, i.e., a frame is multiple dimensioned, 4 frame single scales.It both can be timely
Target scale information is updated, and every frame can effectively be avoided all to carry out multiscale tracing and waste time.
It is to sum up that a kind of multiscale target tracking of high real-time provided by the invention is explained in detail.This theory
Bright book content should not be construed as limitation of the invention, and protection scope of the present invention should be limited to the appended claims.
Claims (5)
1. a kind of multiscale target tracking of high real-time, which comprises the following steps:
Step 1, area-of-interest is determined；Target to be tracked is chosen with rectangle frame in initial image frame, extracts target and its week
Enclose background area；
Step 2, feature is extracted；To region of interesting extraction Fast Field direction histogram feature；
Step 3, training correlation filtering tracker；Feature vector and dimensional Gaussian peak function are transformed into frequency-domain calculations, obtain phase
Close filter；
Step 4, target position is detected；Same such as step 1 extracts area-of-interest in next frame image, is obtained using training
Correlation filtering tracker, response maximum are target position；
Step 5, size estimation；Scale prediction is combined with scale pond, determines target scale.
2. a kind of multiscale target tracking of high real-time according to claim 1, which is characterized in that based on correlation
The response characteristic of filter is that the dimension calculation in scale pond divides priority；The step of dividing scale priority specifically includes
Below:
Firstly, the variation of every frame target scale is divided into three kinds of situations: 0.95,1,1.05, i.e., small scale, normally, large scale；With
This respectively indicates target and becomes smaller, constant and the case where become larger；
Target is practical when becoming larger, if obtained response is variable R, reasons out as R (0.95) < R (1) < R (1.05)；If therefore first
R (1) and R (1.05) are calculated and compared, practical maximum value is obtained, without calculating the response to small scale again；
Become smaller if target is practical, reasons out as R (0.95) > R (1) > R (1.05)；If therefore first calculating and comparing R (1) and R
(0.95), practical maximum value is obtained, without calculating the response of large scale again；
If target physical size does not change, assign at random first it is a kind of in above situation calculated, then calculate the third
The response of scale is compared；
Statistics-Based Method, using target actual motion information, confirmation calculates the response of certain scale first；The statistics past 10
The dimensional variation situation of frame, which kind of dimensional variation is in the majority, i.e., is preferentially calculated in present frame with the scale；Either which kind of feelings paid attention to
Condition is both needed to response when calculating R (1), i.e. Scale invariant, can just be compared in this way；Statistics-Based Method is specific as follows:
Definition node first, each node have bidirectional pointer, are respectively directed to front and back node, and intra-node defines mScale change
Amount, for recording current scale state, state, which is divided into, to become larger (value 1), constant (value 0), becomes smaller (value is -1) and default (value
For 2)；
After creating circular linked list, it is directed toward arbitrary node using pointer, every to pass through a frame, pointer moves down a node, when frame number is big
When queue length, pointer is directed toward node initial in chained list, so that the nodal information is capped, realizes nearest 10 frame information
Record；
In order to reduce the time complexity of statistical yardstick information, a mSum variable is safeguarded, for recording all frame nanoscale regimes
The sum of mScale then illustrates what preceding 10 frame scale became larger since mScale design is the statistical variable of 0 mean value, and if more than 0
Frame number is more, if being equal to 0, the frame number of Scale invariant is more, if the frame number that scale becomes smaller is more less than 0, mSum variable
Modification subtracts the dimensional information for currently pointing to node, in addition the scale of present frame positioned at the time of this frame information is arranged every time
Information；
Use head node information realization full to team simultaneously and the judgement of team's dummy status, if head node nanoscale regime is default, head
Node is not used by, and judges team's sky；If the previous node nanoscale regime of head node be it is non-default, judge that team is full.
3. a kind of multiscale target tracking of high real-time according to claim 2, which is characterized in that for statistics
The unconverted target of scale, tracking is using in such a way that frame is multiple dimensioned；I.e. a frame is multiple dimensioned, multiframe single scale alternately into
Row.
4. a kind of multiscale target tracking of high real-time according to claim 3, which is characterized in that every 5 frame into
Row is primary multiple dimensioned, i.e., a frame is multiple dimensioned, 4 frame single scales.
5. a kind of multiscale target tracking of high real-time according to claim 1, which is characterized in that store and sharp
With the frequency domain operation of feature vector；
It is closed in filter tracks method in nuclear phase, either linear kernel or Gaussian kernel, it will be right in trained and detection process
Two feature vectors of input are transformed into frequency domain and are handled；Gaussian kernel calculation formula such as (1):
Linear kernel calculation formula such as (2):
Wherein, k (x, x') is about feature vector x, the kernel function of x', F-1For the inverse transformation of FFT,WithIt is transformed into for x, x'
Frequency domain as a result, ⊙ be dot-product operation, * expression take conjugation, σ is bandwidth；
Training formula such as formula (3), wherein α is the implicit coefficient of correlation filter, and k is linear kernel function variable k (x, x), and x is
Training image block eigenvector, y are the label vector that two-dimensional Gaussian function is constituted, and λ is regularization factors；Kernel function in training
Two equal x of input, therefore practical need to only carry out a FFT to x；
Detection formula such as formula (4), whereinIt is exported for the response of all positions,For kernel function variable k (x, z), wherein x is
Training image block eigenvector, z test image block eigenvector；Response maximum is target position；In multiple dimensioned meter
Link is calculated, detection process is repeated to the image block of different scale, i.e., only change test image block eigenvector z, therefore stored first
FFT (x) when subdimension operation, calls directly in subsequent arithmetic.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910559301.0A CN110298868B (en) | 2019-06-26 | 2019-06-26 | High-instantaneity multi-scale target tracking method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910559301.0A CN110298868B (en) | 2019-06-26 | 2019-06-26 | High-instantaneity multi-scale target tracking method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110298868A true CN110298868A (en) | 2019-10-01 |
| CN110298868B CN110298868B (en) | 2021-06-25 |
Family
ID=68028836
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910559301.0A Active CN110298868B (en) | 2019-06-26 | 2019-06-26 | High-instantaneity multi-scale target tracking method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110298868B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112686609A (en) * | 2020-12-31 | 2021-04-20 | 江苏佳利达国际物流股份有限公司 | Intelligent unmanned logistics transportation method and system based on optimization efficiency evaluation algorithm |
| CN112991394A (en) * | 2021-04-16 | 2021-06-18 | 北京京航计算通讯研究所 | KCF target tracking method based on cubic spline interpolation and Markov chain |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106204638A (en) * | 2016-06-29 | 2016-12-07 | 西安电子科技大学 | A kind of based on dimension self-adaption with the method for tracking target of taking photo by plane blocking process |
| CN107103616A (en) * | 2017-04-27 | 2017-08-29 | 中国科学院长春光学精密机械与物理研究所 | A kind of method for tracking target and system |
| CN108010067A (en) * | 2017-12-25 | 2018-05-08 | 北京航空航天大学 | A kind of visual target tracking method based on combination determination strategy |
| CN108346159A (en) * | 2018-01-28 | 2018-07-31 | 北京工业大学 | A kind of visual target tracking method based on tracking-study-detection |
| CN108550161A (en) * | 2018-03-20 | 2018-09-18 | 南京邮电大学 | A kind of dimension self-adaption core correlation filtering fast-moving target tracking method |
| CN109410247A (en) * | 2018-10-16 | 2019-03-01 | 中国石油大学(华东) | A kind of video tracking algorithm of multi-template and adaptive features select |
| CN109685073A (en) * | 2018-12-28 | 2019-04-26 | 南京工程学院 | A kind of dimension self-adaption target tracking algorism based on core correlation filtering |
-
2019
- 2019-06-26 CN CN201910559301.0A patent/CN110298868B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106204638A (en) * | 2016-06-29 | 2016-12-07 | 西安电子科技大学 | A kind of based on dimension self-adaption with the method for tracking target of taking photo by plane blocking process |
| CN107103616A (en) * | 2017-04-27 | 2017-08-29 | 中国科学院长春光学精密机械与物理研究所 | A kind of method for tracking target and system |
| CN108010067A (en) * | 2017-12-25 | 2018-05-08 | 北京航空航天大学 | A kind of visual target tracking method based on combination determination strategy |
| CN108346159A (en) * | 2018-01-28 | 2018-07-31 | 北京工业大学 | A kind of visual target tracking method based on tracking-study-detection |
| CN108550161A (en) * | 2018-03-20 | 2018-09-18 | 南京邮电大学 | A kind of dimension self-adaption core correlation filtering fast-moving target tracking method |
| CN109410247A (en) * | 2018-10-16 | 2019-03-01 | 中国石油大学(华东) | A kind of video tracking algorithm of multi-template and adaptive features select |
| CN109685073A (en) * | 2018-12-28 | 2019-04-26 | 南京工程学院 | A kind of dimension self-adaption target tracking algorism based on core correlation filtering |
Non-Patent Citations (3)
| Title |
|---|
| MENGDAN ZHANG ET AL: "ROBUST VISUAL TRACKING USING JOINT SCALE-SPATIAL CORRELATION FILTERS", 《2015 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING》 * |
| 丁建伟 等: "快速尺度估计的核相关滤波目标跟踪", 《科学技术与工程》 * |
| 孙营 等: "基于可变形部件改进模型的夜间车辆检测方法", 《计算机工程》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112686609A (en) * | 2020-12-31 | 2021-04-20 | 江苏佳利达国际物流股份有限公司 | Intelligent unmanned logistics transportation method and system based on optimization efficiency evaluation algorithm |
| CN112686609B (en) * | 2020-12-31 | 2021-08-13 | 江苏佳利达国际物流股份有限公司 | Intelligent unmanned logistics transportation method and system based on optimization efficiency evaluation algorithm |
| CN112991394A (en) * | 2021-04-16 | 2021-06-18 | 北京京航计算通讯研究所 | KCF target tracking method based on cubic spline interpolation and Markov chain |
| CN112991394B (en) * | 2021-04-16 | 2024-01-19 | 北京京航计算通讯研究所 | KCF target tracking method based on cubic spline interpolation and Markov chain |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110298868B (en) | 2021-06-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | YOLO-ACN: Focusing on small target and occluded object detection | |
| Xu et al. | Multimodal cross-layer bilinear pooling for RGBT tracking | |
| Zhang et al. | A background-aware correlation filter with adaptive saliency-aware regularization for visual tracking | |
| CN111915644A (en) | Real-time target tracking method of twin guiding anchor frame RPN network | |
| Gu et al. | Class-incremental instance segmentation via multi-teacher networks | |
| CN110298868A (en) | A kind of multiscale target tracking of high real-time | |
| CN114898403A (en) | Pedestrian multi-target tracking method based on Attention-JDE network | |
| An et al. | Transitive transfer learning-based anchor free rotatable detector for SAR target detection with few samples | |
| Zeng et al. | A novel tensor decomposition-based efficient detector for low-altitude aerial objects with knowledge distillation scheme | |
| Sowmya et al. | Heavy-vehicle detection based on YOLOv4 featuring data augmentation and transfer-learning techniques | |
| Xu et al. | Hierarchical convolution fusion-based adaptive Siamese network for infrared target tracking | |
| Wu et al. | M-FasterSeg: An efficient semantic segmentation network based on neural architecture search | |
| CN109978923A (en) | One kind being based on double-template dimension self-adaption correlation filtering method for tracking target and system | |
| Moridvaisi et al. | An extended KCF tracking algorithm based on TLD structure in low frame rate videos | |
| Ma et al. | Robust visual tracking via adaptive feature channel selection | |
| CN112991394A (en) | KCF target tracking method based on cubic spline interpolation and Markov chain | |
| CN108154522A (en) | Target tracking system | |
| Fan et al. | High-speed tracking based on multi-CF filters and attention mechanism | |
| Lotfi et al. | Object localization through a single multiple-model switching CNN and a superpixel training approach | |
| CN116523957A (en) | Multi-target tracking method, system, electronic equipment and storage medium | |
| Wang et al. | VTAAN: Visual tracking with attentive adversarial network | |
| Liu et al. | Combined with the residual and multi-scale method for Chinese thermal power system record text recognition | |
| Zhou et al. | A strong feature representation for siamese network tracker | |
| Yin et al. | Pig target detection from image based on improved YOLO V3 | |
| Du et al. | Target tracking algorithm for pedestrians movement based on kernel-correlation filtering |
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 |