CN106372650B - A kind of compression tracking based on motion prediction - Google Patents
A kind of compression tracking based on motion prediction Download PDFInfo
- Publication number
- CN106372650B CN106372650B CN201610701199.XA CN201610701199A CN106372650B CN 106372650 B CN106372650 B CN 106372650B CN 201610701199 A CN201610701199 A CN 201610701199A CN 106372650 B CN106372650 B CN 106372650B
- Authority
- CN
- China
- Prior art keywords
- target
- tracking
- frame
- classifier
- motion
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
- G06F18/24155—Bayesian classification
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
Compression tracking based on motion prediction of the invention carries out motion prediction to video object according to present frame, obtains the direction of motion of target;According to motion vector, the distance for calculating the movement of front cross frame target reduces the acquisition of candidate samples according to this apart from adjust automatically search range;Using adaptive tracing window optimization, positive and negative sample set is acquired again, extracts the feature of two sample sets, carries out the update of Naive Bayes Classifier, and records present frame target position, the target position traced into, undated parameter.It has the beneficial effect that and greatly reduces search time, reduce the acquisition of candidate samples, to improve real-time and robustness of the compression tracking under some complex scenes.
Description
Technical field
The invention belongs to technical field of computer vision more particularly to a kind of compression trackings based on motion prediction.
Background technique
With the rapid development of electronic computer technology, computer vision becomes popular research topic.Intelligent video prison
Control gradually infiltrates into daily life, is automatically analyzed using video sequence image to detect, track and identify monitoring field
Target in scape, and then analyze and determine target and make countermeasure.And video frequency object tracking is the key that portion in intelligent monitor system
Point, merge that image procossing, pattern-recognition, signal processing and control etc. is multi-field, multidisciplinary project.Due to tracking by
Therefore the influence of several factors, type change, the variation of illumination, the problems such as blocking of object for being particularly due to target establish one
Robust, adaptive method for tracking target are still a challenging problem.
In recent years, one kind is found efficiently to pay close attention to the tracking of robust by researcher.2012, Zhang et al.
[document 1] (Zhang K H, Zhang L, Yang M H.Real-time compressive tracking [C] .European
Conference on Computer Vision, 2012:864-877) propose compression track algorithm (Compressive
Tracking, CT), algorithm uses optimal experimental configuration by experiment, and the method to next frame image procossing is according to before
The rectangle frame of 20 Euclidean distance radiuses is all used as candidate region around one frame target upper left angle point, to each extracted region 50
Haar-like characteristic value.It is finally screened, is selected most using characteristic value of the Naive Bayes Classifier to these candidate regions
The region of macrotaxonomy device number of responses, the as target area of present frame.When determining target area, using this extensive
Search strategy is the more wasteful calculating time.Therefore, motion prediction is introduced into track algorithm, with prediction direction, to this
It is chosen on a large scale on a direction, candidate region number is reduced to other directions.
Common motion forecast method has very much, such as Luo et al. [document 2] (Luo H L, Zhong B K, Kong F
S.Object tracking algorithm by combining the predicted target position with
Compressive tracking [J] .Journal of Image and Graphics, 2014,19 (6): 875-885.) it mentions
The position for using Mean shift to carry out target motion prediction, and passing through prediction out influences tracing positional, raising tracking accuracy.
(Yang Dongdong, Chang Danhua, Han Xia wait the improvement of moving object detection and tracking algorithm to Yang Dongdong et al. [document 3]
With realization [J] laser and infrared, 2010,40 (2): 205-209) prediction that is moved using motion history image, is improved
Tracking accuracy.
2016, Zhang et al. [document 4] (Zhang K H, Liu Q S, Wu Y, et al.Robust visual
tracking via convolutional networks without training[J].IEEE Transactions on
Image Processing, 2016,25 (4): 1779-1792) propose convolutional neural networks algorithm (Convolutional
Networks without Training, CNT), tracking performance is obviously improved.These algorithms are used during tracking, it can
To improve the performance of target following.But these algorithm comparisons are complicated, computational complexity is relatively high, cannot better meet reality
Border demand.
Summary of the invention
The present invention in order to overcome the deficiencies of the prior art, provides that a kind of complexity is low, robustness is high based on motion prediction
Compression tracking, be specifically realized by the following technical scheme:
The compression tracking based on motion prediction, comprising:
Initialization is chosen in first frame image and tracks target area, the sampled images near first frame image target area
Block carries out feature extraction to described image block and dimensionality reduction obtains the feature vector of each image block, establishes classifier;
Track target and carry out motion prediction to target: for t+1 frame image, t is the integer more than or equal to 2, is used
Target position in front cross frame image obtains target motion vectors and prediction direction, and according to the target motion vectors and prediction
Direction obtains the predicted position of target, the sampled images block near the predicted position that t frame traces into, and then obtains each figure
As the feature vector of block, target that the classifier traces into the corresponding image block of maximum classifier as present frame;
It determines the final tracking result of present frame: adaptive optimization being carried out to tracking window, in the target proximity, again
It is sampled, and is selected by classifier, using the corresponding image block of maximum classifier as final tracking target;And it records
Present frame target position traces into target position, updates classifier parameters.
The further design of the compression tracking based on motion prediction is that the classifier is using Bayes point
Class device H (v), such as formula (1),
Wherein, p (y=1)=p (y=0), y ∈ { 0,1 } are positive negative sample marks, and y=0 represents negative sample, and y=1 generation
Table positive sample, low dimensional space v=(v1,....,vn)T,viFor i-th of element in v.
The further design of the compression tracking based on motion prediction is, sets the condition in classifier H (v)
Distribution p (vi| y=1) and p (vi| y=0) it is Gaussian Profile, and meet WhereinWithThe respectively expectation and standard deviation of positive sample probability, andWithRespectively
The expectation and standard deviation of negative sample probability,To be desired forIt is with standard deviationGaussian Profile,For
It is desired forIt is with standard deviationGaussian Profile, N represents the symbol of Gaussian Profile, and Bayes classifier parameter is updated such as formula
(2), formula (3):
Wherein, μ1、σ1Indicate the expectation and standard deviation of positive sample after updating, λ > 0 is Studying factors, n representative sample number, vi(k) represent kth sample indicates in lower dimensional space, k
Represent k-th of sample.
The further design of the compression tracking based on motion prediction is that the sampling of described image block passes through excellent
The positive and negative samples selection changed, which is realized, to be acquired, positive sample selection gist success rate formula, success rate formula such as formula (4):
Wherein, CI is the calculated target frame region of algorithm, and RI is realistic objective region;When the value of Rt in a frame is greater than
0.5, then it is assumed that tracking result is correct.
The compression tracking based on motion prediction it is further design be, the optimization of positive and negative samples selection, according to
According to the success rate formula, the optimization formula such as formula (5) of positive and negative samples selection:
Wherein, w, h are the width and height of target window, and l is offset, and n and m are horizontal and vertical offset components,
Positive sample selection region is success rate in 0.9 or more region, and negative sample is success rate in 0.5 region below.
The compression tracking based on motion prediction it is further design be, the motion prediction specifically include as
Lower step:
Step 1) sets former frame target position as (x1,y1), present frame target position is (x2,y2), then motion vector β
Are as follows:
β=(x2,y2)-(x1,y1);
Step 2) defines a matrix A, the matrix A size and search according to target direction of motion, definition of search region
Range size is identical, it is divided into four quadrants according to rectangular coordinate system, the quadrant where β is set as candidate region, mesh
Candidate region is marked to carry out with matrix A and after operation, be just only left to rely on the candidate target in the direction of motion of former frame prediction.
Step 3) defines another matrix B, copes with the tracking in present frame in the case of the unexpected deflecting of target:
Use matrix A and matrix B progress or operation as sampling matrix;
Step 4) calculates the distance d of front cross frame target movement, is adaptively adjusted according to the distance according to motion vector
Weight is arranged by the distance and target window size of movement, then weight ω in search range are as follows:
Wherein, w, h are the width and height of target window;Initial search frequency range is adaptively adjusted using weight ω.
The compression tracking based on motion prediction it is further design be, the adaptive optimization of tracking window,
Specifically: window size is adjusted with the distance of two pixels every time, feature extraction is carried out in these windows, and use pattra leaves
This classifier is classified, and is found best matching result, that is, is obtained the adaptive optimization of tracking window.
Advantages of the present invention:
First frame the present invention is based on the compression tracking of motion prediction in target video sequence gives the initial shape that sets the goal
In the case where state, motion prediction is carried out to video object according to present frame, obtains the direction of motion of target according to motion vector, meter
The distance for calculating the movement of front cross frame target greatly reduces search time according to this apart from adjust automatically search range, reduces
The acquisition of candidate samples;Using adaptive tracing window optimization, to improve compression tracking under some complex scenes
Real-time and robustness.
Detailed description of the invention
The flow diagram of Fig. 1 the method for the present invention.
Fig. 2 the method for the present invention positive sample selection optimization schematic diagram.
Fig. 3 compressed sensing algorithm keeps track schematic diagram and the method for the present invention adaptive tracing window optimization track schematic diagram.
Fig. 4 the method for the present invention, CT and CNT tracking result schematic diagram.
Fig. 5 the method for the present invention, CT and CNT track another result schematic diagram.
Another Fig. 6 the method for the present invention, CT and CNT tracking result schematic diagram.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawing to the application into
One step is described in detail.
Such as Fig. 1, the compression tracking based on motion prediction of the present embodiment, comprising: first frame image is chosen in initialization
Middle tracking target area, the sampled images block near first frame image target area carry out feature extraction and dimensionality reduction to image block
The feature vector of each image block is obtained, classifier is established.It tracks target and motion prediction is carried out to target: for t+1 frame
Image, using front cross frame, the value of t is greater than the integer equal to 2.Target position in image obtains target motion vectors and pre-
Direction is surveyed, and obtains the predicted position of target according to target motion vectors and prediction direction, in the predicted position that t frame traces into
Near sampled images block, and then obtain the feature vector of each image block, classifier is by the corresponding image block of maximum classifier
The target traced into as present frame.It determines the final tracking result of present frame: adaptive optimization being carried out to tracking window, in mesh
It near mark, is sampled again, and selected by classifier, using the corresponding image block of maximum classifier as final tracking
Target;And present frame target position is recorded, target position is traced into, classifier parameters, such as Fig. 3 are updated.
In the present embodiment, classifier uses Bayes classifier H (v), such as formula (1),
Wherein, p (y=1)=p (y=0), y ∈ { 0,1 } are positive negative sample marks, and y=0 represents negative sample, and y=1 generation
Table positive sample, low dimensional space v=(v1,....,vn)T,viFor i-th of element in v.
Further, the condition distribution p (v in classifier H (v) is seti| y=1) and p (vi| y=0) it is Gaussian Profile, and
MeetWhereinWithThe respectively phase of positive sample probability
Prestige and standard deviation, andWithThe respectively expectation and standard deviation of negative sample probability,To be desired forAnd standard deviation
ForGaussian Profile,To be desired forIt is with standard deviationGaussian Profile, N represents the symbol of Gaussian Profile,
Bayes classifier parameter is updated such as formula (2), formula (3):
Wherein, μ1、σ1Indicate the expectation and standard deviation of positive sample after updating, λ > 0 is Studying factors, n representative sample number, vi(k) kth sample is represented in lower dimensional space expression, k
Represent k-th of sample.
In the present embodiment, the sampling of image block realizes acquisition, positive sample selection gist by the positive and negative samples selection of optimization
Success rate formula, success rate formula such as formula (4):
Wherein, CI is the calculated target frame region of algorithm, and RI is realistic objective region;When the value of Rt in a frame is greater than
0.5, then it is assumed that tracking result is correct.
Further, such as Fig. 2, the optimization of positive and negative samples selection, according to success rate formula, the optimization of positive and negative samples selection is public
Formula such as formula (5):
Wherein, w, h are the width and height of target window, and l is offset, and n and m are horizontal and vertical offset components,
Positive sample selection region is success rate in 0.9 or more region, and negative sample is success rate in 0.5 region below.
Motion prediction specifically comprises the following steps: in the present embodiment
Step 1) sets former frame target position as (x1,y1), present frame target position is (x2,y2), then motion vector β
Are as follows:
β=(x2,y2)-(x1,y1);
Step 2) defines a matrix A, the matrix size and search model according to target direction of motion, definition of search region
It is identical to enclose size, it is divided into four quadrants according to rectangular coordinate system, the quadrant where β is set as candidate region, target
Candidate region carries out with matrix A and after operation, is just only left to rely on the candidate target in the direction of motion of former frame prediction.
Step 3) defines another matrix B, copes with the tracking in present frame in the case of the unexpected deflecting of target:
Use matrix A and matrix B progress or operation as sampling matrix;
Step 4) calculates the distance d of front cross frame target movement, is adaptively adjusted according to the distance according to motion vector
Weight is arranged by the distance and target window size of movement, then weight ω in search range are as follows:
Wherein, w, h are the width and height of target window;Initial search frequency range is adaptively adjusted using weight ω.
In order to improve the robustness of tracking, the adaptive optimization of tracking window, specifically: every time with the distance of two pixels
Window size is adjusted, feature extraction is carried out in these windows, and classify using Bayes classifier, finds best
With as a result, obtaining the adaptive optimization of tracking window.
The effect to the present embodiment method of the application has carried out experimental verification, using CT [document 1] and CNT [document 4],
It is compared to the effect with the method for the present invention, in conjunction with Fig. 4, Fig. 5 and tracking result figure shown in fig. 6.The tracking of CT algorithm
As a result it is marked with No. 1 frame solid line of red, CNT arithmetic result is marked with green No. 2 frame solid lines, and result of the present invention uses blue 3
Dotted box marks.As seen from the figure, the accuracy of the tracking of the present embodiment either target following is still in tracking window
Target position be all better than other two methods.Experiment is verified in 15 kinds of challenging sequences, is tested in total
7531 frame images.This 15 kinds of videos are public video library [document 5] (http://cvlab.hanyang.ac.kr/tracker_
Benchmark/ it is randomly selected in).Experimental facilities is configured to, 2.5Ghz dominant frequency four core Core i5CPU, memory 4GB,
Windows7 32-bit operating system, and run in MATLAB 2014a development platform.
For evaluation algorithms performance of target tracking, inventor using tracking success rate Rt and center offset two indices come
Measure the accuracy rate of tracking.Wherein, success rate Rt is defined as:
Wherein, CI is the calculated target frame region of algorithm, and RI is realistic objective region.So, when the value of Rt in a frame
Greater than 0.5, then it is assumed that tracking result is correct.
The definition of center offset is the central point of tracking result frame and the Euclidean distance of actual frames central point.As a result as follows
Shown in table.
1 time of table and success rate
2 center offset of table
As seen from the above table, the method for the present invention improves the robustness and real-time of tracking.Compared to CNT algorithm, the present invention
Method has very big advantage in terms of real-time, and the time used is more than one the percent of CNT algorithm;Come in tracking success rate
It sees, CNT algorithm doing well in individual video, but entirety will be slightly worse than the method for the present invention.Meanwhile the method for the present invention
Mean center offset is minimum, the center offset highest of CT algorithm.Therefore, in terms of the degree of off-centring, side of the present invention
There has also been biggish improvement for method.
The compression tracking based on motion prediction of the present embodiment is first to setting the goal in the first frame of target video sequence
In the case where beginning state, motion prediction is carried out to video object according to present frame, obtains the direction of motion of target according to movement arrow
Amount, the distance for calculating the movement of front cross frame target greatly reduce search time according to this apart from adjust automatically search range,
Reduce the acquisition of candidate samples;Using adaptive tracing window optimization, to improve compression tracking in some complicated fields
Real-time and robustness under scape.
The foregoing is only a preferred embodiment of the present invention, but protection scope of the present invention it is not limited to this, appoint
What those familiar with the art in the technical scope disclosed by the present invention, according to the technique and scheme of the present invention and its this
Inventive concept is subject to equivalent substitution or change, is all included within the scope of the present invention.
Claims (6)
1. a kind of compression tracking based on motion prediction, characterized by comprising:
Initialization is chosen in first frame image and tracks target area, and sampled images block, right near first frame image target area
Described image block carries out feature extraction and dimensionality reduction obtains the feature vector of each image block, establishes classifier;
Track target and carry out motion prediction to target: for t+1 frame image, t is the integer more than or equal to 2, uses preceding two
Target position in frame image obtains target motion vectors and prediction direction, and according to the target motion vectors and prediction direction
The predicted position for obtaining target, the sampled images block near the predicted position that t frame traces into, and then obtain each image block
Feature vector, the target that the classifier traces into the corresponding image block of maximum classifier as present frame;
It determines the final tracking result of present frame: adaptive optimization being carried out to tracking window and is carried out again in the target proximity
Sampling, and selected by classifier, using the corresponding image block of maximum classifier as final tracking target;And it records current
Frame target position traces into target position, updates classifier parameters;
The motion prediction specifically comprises the following steps:
Step 1) sets former frame target position as (x1,y1), present frame target position is (x2,y2), then motion vector β are as follows:
β=(x2,y2)-(x1,y1);
Step 2) defines a matrix A, the matrix A size and search range according to target direction of motion, definition of search region
Size is identical, it is divided into four quadrants according to rectangular coordinate system, the quadrant where β is set as candidate region, target is waited
Favored area carries out with matrix A and after operation, is just only left to rely on the candidate target in the direction of motion of former frame prediction;
Step 3) defines another matrix B, copes with the tracking in present frame in the case of the unexpected deflecting of target:
Use matrix A and matrix B progress or operation as sampling matrix;
Step 4) calculates the distance d of front cross frame target movement, is adaptively adjusted search according to the distance according to motion vector
Weight is arranged by the distance and target window size of movement, then weight ω in range are as follows:
Wherein, w, h are the width and height of target window;Initial search frequency range is adaptively adjusted using weight ω.
2. the compression tracking according to claim 1 based on motion prediction, it is characterised in that the classifier uses
Bayes classifier H (v), such as formula (1),
Wherein, p (y=1)=p (y=0), y ∈ { 0,1 } are positive negative sample marks, and y=0 represents negative sample, and y=1 is represented just
Sample, low dimensional space v=(v1,....,vn)T,viFor i-th of element in v.
3. the compression tracking according to claim 2 based on motion prediction, it is characterised in that setting classifier H (v)
In condition distribution p (vi| y=1) and p (vi| y=0) it is Gaussian Profile, and meetWhereinWithRespectively the expectation of positive sample probability and
Standard deviation, andWithThe respectively expectation and standard deviation of negative sample probability,To be desired forIt is with standard deviation
Gaussian Profile,To be desired forIt is with standard deviationGaussian Profile, N represents the symbol of Gaussian Profile, pattra leaves
This classifier parameters is updated such as formula (2), formula (3):
Wherein, μ1、σ1Indicate the expectation and standard deviation of positive sample after updating, λ > 0 is Studying factors, n representative sample number, vi(k) kth sample is represented in lower dimensional space expression, k
Represent k-th of sample.
4. the compression tracking according to claim 1 based on motion prediction, it is characterised in that described image block is adopted
Positive and negative samples selection realization acquisition of the sample by optimization, positive sample selection gist success rate formula, success rate formula such as formula (4):
Wherein, CI is the calculated target frame region of algorithm, and RI is realistic objective region;When Rt in a frame value be greater than 0.5, then
Think that tracking result is correct.
5. the compression tracking according to claim 4 based on motion prediction, it is characterised in that positive and negative samples selection
Optimization, according to the success rate formula, the optimization formula such as formula (5) of positive and negative samples selection:
Wherein, w, h are the width and height of target window, and l is offset, and n and m are horizontal and vertical offset component, positive sample
This selection region is success rate in 0.9 or more region, and negative sample is success rate in 0.5 region below.
6. the compression tracking according to claim 2 based on motion prediction, it is characterised in that tracking window it is adaptive
It should optimize, specifically: window size is adjusted with the distance of two pixels every time, feature extraction is carried out in these windows, and
Classified using Bayes classifier, finds best matching result, that is, obtain the adaptive optimization of tracking window.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610701199.XA CN106372650B (en) | 2016-08-19 | 2016-08-19 | A kind of compression tracking based on motion prediction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610701199.XA CN106372650B (en) | 2016-08-19 | 2016-08-19 | A kind of compression tracking based on motion prediction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106372650A CN106372650A (en) | 2017-02-01 |
CN106372650B true CN106372650B (en) | 2019-03-19 |
Family
ID=57879528
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610701199.XA Active CN106372650B (en) | 2016-08-19 | 2016-08-19 | A kind of compression tracking based on motion prediction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106372650B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108304808B (en) * | 2018-02-06 | 2021-08-17 | 广东顺德西安交通大学研究院 | Monitoring video object detection method based on temporal-spatial information and deep network |
CN109859242B (en) * | 2019-01-16 | 2022-09-09 | 重庆邮电大学 | Target tracking method for prediction adaptive learning |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101867798A (en) * | 2010-05-18 | 2010-10-20 | 武汉大学 | Mean shift moving object tracking method based on compressed domain analysis |
CN102881024A (en) * | 2012-08-24 | 2013-01-16 | 南京航空航天大学 | Tracking-learning-detection (TLD)-based video object tracking method |
CN104683802A (en) * | 2015-03-24 | 2015-06-03 | 江南大学 | H.264/AVC compressed domain based moving target tracking method |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2411229B (en) * | 2003-07-22 | 2006-04-12 | Hitachi Int Electric Inc | Object tracking method and object tracing apparatus |
-
2016
- 2016-08-19 CN CN201610701199.XA patent/CN106372650B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101867798A (en) * | 2010-05-18 | 2010-10-20 | 武汉大学 | Mean shift moving object tracking method based on compressed domain analysis |
CN102881024A (en) * | 2012-08-24 | 2013-01-16 | 南京航空航天大学 | Tracking-learning-detection (TLD)-based video object tracking method |
CN104683802A (en) * | 2015-03-24 | 2015-06-03 | 江南大学 | H.264/AVC compressed domain based moving target tracking method |
Non-Patent Citations (1)
Title |
---|
"基于压缩感知的运动目标跟踪算法的研究";王哲;《中国优秀硕士学位论文全文数据库 信息科技辑》;20140215;第29-30、41、43-44、46页 |
Also Published As
Publication number | Publication date |
---|---|
CN106372650A (en) | 2017-02-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Jiao et al. | New generation deep learning for video object detection: A survey | |
CN108470354B (en) | Video target tracking method and device and implementation device | |
WO2020108362A1 (en) | Body posture detection method, apparatus and device, and storage medium | |
Yuan et al. | Learning target-focusing convolutional regression model for visual object tracking | |
CN107358623B (en) | Relevant filtering tracking method based on significance detection and robustness scale estimation | |
CN103325112B (en) | Moving target method for quick in dynamic scene | |
Han et al. | Visual object tracking via sample-based Adaptive Sparse Representation (AdaSR) | |
CN102307274B (en) | Motion detection method based on edge detection and frame difference | |
CN107590427B (en) | Method for detecting abnormal events of surveillance video based on space-time interest point noise reduction | |
CN109685045A (en) | A kind of Moving Targets Based on Video Streams tracking and system | |
Zhang et al. | A swarm intelligence based searching strategy for articulated 3D human body tracking | |
Bedagkar-Gala et al. | Gait-assisted person re-identification in wide area surveillance | |
CN112329784A (en) | Correlation filtering tracking method based on space-time perception and multimodal response | |
CN112616023A (en) | Multi-camera video target tracking method in complex environment | |
CN106372650B (en) | A kind of compression tracking based on motion prediction | |
Zhang et al. | Residual memory inference network for regression tracking with weighted gradient harmonized loss | |
CN113129332A (en) | Method and apparatus for performing target object tracking | |
Cai et al. | Anti-occlusion multi-object surveillance based on improved deep learning approach and multi-feature enhancement for unmanned smart grid safety | |
Chen et al. | A unified model sharing framework for moving object detection | |
CN103971100A (en) | Video-based camouflage and peeping behavior detection method for automated teller machine | |
CN101877135B (en) | Moving target detecting method based on background reconstruction | |
CN112733770A (en) | Regional intrusion monitoring method and device | |
CN104463910B (en) | High-speed motion target extraction method based on motion vector | |
CN110570450A (en) | Target tracking method based on cascade context-aware framework | |
CN110232314A (en) | A kind of image pedestrian's detection method based on improved Hog feature combination neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | 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 |