CN108537826A - A kind of Ship Target tracking based on manual intervention - Google Patents
A kind of Ship Target tracking based on manual intervention Download PDFInfo
- Publication number
- CN108537826A CN108537826A CN201810273119.4A CN201810273119A CN108537826A CN 108537826 A CN108537826 A CN 108537826A CN 201810273119 A CN201810273119 A CN 201810273119A CN 108537826 A CN108537826 A CN 108537826A
- Authority
- CN
- China
- Prior art keywords
- frame
- region
- naval vessel
- area
- tracking
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- 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
- G06T7/251—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving models
-
- 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
- G06T7/75—Determining position or orientation of objects or cameras using feature-based methods involving models
-
- 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
- G06T2207/20081—Training; Learning
-
- 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
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of Ship Target tracking based on manual intervention, including step:S1. the naval vessel area information in first frame on certain video carrier is sketched the contours of manually;S2. the video carrier after normalization size is input in Fast RCNN networks, carries out the tracking monitor in naval vessel region to the frame after first frame by Fast RCNN networks, obtains the location information on the naval vessel in each frame;S3. by the location information on naval vessel in human eye observation's present frame, and judge whether annotation results are qualified;Then include after completing manual intervention operation, continuing step S2 by naval vessel region by manually drawing a box with mouse if S4. unqualified;S5. above step is all executed to each frame in video carrier, until tracking operation terminates.The present invention uses the deep learning frame based on Fast RCNN networks, realizes and is operated to the tracking of Ship Target, to prevent occurring during tracking the method that tracking error artificially corrects tracing area using manual intervention again, substantially increases the accuracy of tracking result.
Description
Technical field
The present invention relates to a kind of Ship Target tracking more particularly to a kind of Ship Target tracking based on manual intervention
Method.
Background technology
There are its identification and the detection of movement position in naval vessel as seaborne main body and important military target
Very important meaning.It is most of greatly to use the study based on non-depth for current target detection technique
Method realize detection to moving target, such as:The optical flow algorithm etc. of particle filter, Meanshift and feature based point,
More or less there is the shortcomings of certain precision accuracy for these methods.
Invention content
The main purpose of the present invention is to provide a kind of naval vessel tracking based on manual intervention, it is intended to solve existing at present
The technical issues of shortcoming of some naval vessel tracking precision accuracys.
To achieve the goals above, the technical solution that the present invention takes is to provide a kind of naval vessel tracking based on manual intervention
Method includes the following steps:
S1. the naval vessel area information in first frame on certain video carrier is sketched the contours of manually;
S2. the video carrier after normalization size is input in Fast-RCNN networks, passes through Fast-RCNN networks pair
Frame after first frame carries out the tracking monitor in naval vessel region, obtains the location information on the naval vessel in each frame;
S3. by the location information on naval vessel in human eye observation's present frame, and judge whether annotation results are qualified;
Then include completing naval vessel region by manually drawing a box with mouse if S4. unqualified
After manual intervention operation, continue step S2;
S5. above step is all executed to each frame in video carrier, until tracking operation terminates.
Optionally, the step S2 is further comprising the steps of:
S21. target training is carried out to Fast-RCNN networks;
S22. Fast-RCNN networks are used to carry out target following.
Optionally, the step S21 is further comprising the steps of:
A. several times are therefrom obtained by selective search methods for each pictures in training set
Favored area, and record, the location coordinate information of each candidate region;
B. for each candidate region in step A, it is labeled label, i.e., 0 or 1,1 represents the candidate region
Including naval vessel part, 0 indicates that the candidate region does not include naval vessel region part;
C. for each region in step A, all there is its correction position coordinate, i.e.,:
If the markup information in current candidate region is 1, illustrate current candidate region RiIn include naval vessel region, we
Record its correct co-ordinate position information, the amendment coordinate information, including the corresponding top left corner apex coordinate of a box with
And the width and length information of box, and naval vessel part is then surrounded by the box just;
D. then several candidate regions constitute the input of network, mark corresponding to each candidate region and its is right
Correction position information is answered to constitute the output of network;
E. Fast-RCNN networks are trained according to the network inputs and output that are obtained in step D, are calculated by using BP
Method updates weight and the biasing of neuron, finally so that Fast-RCNN networks reach convergence state.
Optionally, the step A includes the following steps:
S211., the size and aspect ratio range of candidate region are set;
S212. it chooses in the video frame a bit at random, is denoted as M, then calculate in video frame between any point N and point M
Color distortion degree point N is slid into the range of convergence of point M, the meter of color distortion degree X if color distortion degree is less than K therewith
Calculating formula is:
Wherein, wherein MR、MG、MBPixel values of the point M on tri- components of R, G and B is indicated respectively, similarly NR、NGAnd NG
Indicate that pixel values of the point N on tri- components of R, G and B, point N are not overlapped with point M respectively, K=150;
S213. selected element N is repeated, and executes diversity factor calculating, expands the range of convergence of point M, the range of convergence of point M is final
An image-region will be constituted, then the image-region just constitutes a part of several above-mentioned candidate regions;
If S214. above-mentioned candidate region meets the size and length-width ratio model for the candidate region being arranged in step S211
It encloses, then terminates step S213;
S215. different point M is randomly selected, and executes step, and executes step S213-S214, and meets and obtains every time
Candidate area size within the scope of size condition in step S211, but candidate area size does not repeat.
Optionally, the step S22 includes the following steps:
Each frame of the video carrier of input is all exported multiple images region by S221.Fast-RCNN networks, for each
Image-region is indicated with a pair of of coordinate points and two parameters w and h;Using this to coordinate points as the top left corner apex of rectangle, w and h
The width and length of rectangle are indicated respectively, then the rectangle area encompassed is the naval vessel region that network is thought;
S222. the location information on naval vessel final in present frame is determined by identified tracing area in previous frame.
Optionally, the step S222 includes the following steps:
S2221. for a later frame of the initialization tracing area of video carrier, it is assumed that in the frame, neural network
Output multiple images region Areai, wherein 1≤i≤n, n are the number of image-region in present frame;
S2222. pass through formula:
It calculates in present frame, each image-region AreaiWith determined tracing area in previous frame, coincidence ratio Li, 1
≤ i≤n, wherein Sp-1Indicate the area of the determined tracing area in -1 frame of pth,It indicates in pth frame, image-region
AreaiCorresponding area, symbol ∩ represent intersection operation, if -1 frame of pth indicates that frame of initialization tracing area,
Sp-1Just the area of initialization tracing area is indicated;
S2223. L is choseniIn the case of obtaining maximum value, corresponding to image-region update the tracking area of present frame
Domain.
Optionally, in the step S3, the tracking of naval vessel regional location whether qualification determination is according to being:In Fast-RCNN
In network, the naval vessel regional location thought is surrounded with a box, if do not had in box area encompassed
There is the naval vessel region by practical 80% or more to be surrounded, therefore thinks that the tracking result accuracy of present frame is relatively low, it is unqualified, then
Execute step S4.
Optionally, the step S4 includes:Tracing area manually is reinitialized with mouse, i.e., draws a side with mouse
Then naval vessel region includes by frame.
Optionally, in the step A, by selective search methods, 2000 candidate regions are therefrom obtained.
Optionally, the range of the size of the candidate region is [5000,10000] pixel size, length than model
Enclose is [5,10].
The beneficial effects of the invention are as follows:The present invention proposes a kind of Ship Target technology based on manual intervention.It is special at this
In profit, the deep learning frame based on Fast-RCNN is used first, to realize that the tracking to Ship Target operates, in addition, being
It prevents during tracking, the case where tracking error (since there are shelters so that BREAK TRACK) of appearance, we
Manual intervention is used, tracing area is artificially corrected, to improve the accuracy of tracking result.
Description of the drawings
Fig. 1 is flow chart of steps provided by the invention;
Fig. 2 is substantially flow chart of steps provided by the invention;
Fig. 3 is human intervention flow chart of steps provided by the invention;
Fig. 4 is the specific steps flow chart of step S21 provided by the invention;
Fig. 5 is the single neuronal structure figure of BP algorithm provided by the invention;
The embodiments will be further described with reference to the accompanying drawings for the realization, the function and the advantages of the object of the present invention.
Specific implementation mode
The present invention is further described for explanation and specific implementation mode below in conjunction with the accompanying drawings.It should be appreciated that described herein
Specific embodiment be only used to explain the present invention, be not intended to limit the present invention.
The method of deep learning is combined by the present invention with visual pursuit, in addition, in order to prevent using deep learning into
During row vision tracks, errors (loss etc. of target tracking) of some trackings of appearance, present invention employs into pedestrian
The operation of work intervention, into when line trace, is missed when in the method using deep learning if human eye finds that target following exists
Difference is then manually corrected tracing area again, it is ensured that the normal progress of tracking largely ensure that the accurate of tracking
Rate.
Using the method for deep learning, that is, use the tracking frame based on Fast-RCNN, realize to Ship Target with
Track not only increases the accuracy rate of tracking, equally also meets the requirement of real-time;By the way of manual intervention, for tracking
As a result there is error, then can timely be corrected tracing area, improve the accuracy rate of tracking effect to a greater extent.
Embodiment 1, the present invention propose a kind of Ship Target technology based on manual intervention.In the present invention, it adopts first
With the deep learning frame based on Fast-RCNN, to realize that the tracking to Ship Target operates, in addition, in order to prevent with
During track, the case where tracking error (since there are shelters so that BREAK TRACK) of appearance, present invention employs people
Work intervention, artificially corrects tracing area, to improve the accuracy of tracking result.Its concrete operations flow is as shown in figure 3, tool
Body includes three big steps:
Step 1 initializes tracing area;
For target following technology, the cover object implemented is often video carrier.First, it would be desirable to selected
One target to be tracked, to complete the operation of initialization tracing area.In the present invention, initialization tracing area be
In the present frame of video carrier, a box is manually used, sketches the contours of naval vessel region part, then the region in box is for we
Required tracing area.
Step 2, the update of tracing area;
It is actually each of the video after initializing tracing area for the update of tracing area operation
In frame, naval vessel region part is equally searched out.For the part, the present invention uses the Fast-RCNN frames based on deep learning
To realize relevant operation.
Step 3, manual intervention;
The specific implementation step of manual intervention, as shown in figure 3, during Ship Target carries out mobile, if background
Greatly either in the current frame, there is shelter, then very in the color similarity in (in video frame, non-naval vessel region) and naval vessel region
During target following, it may appear that there is mistake (i.e. in current, identifying that other regions are naval vessel region) in tracking result,
Further, since during target following, the continuity with target movement, i.e., in the next frame, the target confirmed
The position of tracking is in previous frame, and near identified target location, the range moved there is no target is excessive, because
This, if in the current frame, the target location of tracking is malfunctioned, then can be influenced in the next frame, the result of target following.
Based on this, in this step, the present invention corrects tracing area, utmostly in real time by the way of manual intervention
On ensure the accuracy of tracking result.
Embodiment 2, is based on above-described embodiment, and the present invention provides a kind of naval vessel tracking based on manual intervention, this reality
Applying example and above-described embodiment, difference lies in as shown in Figure 1, include the following steps:
S1. the naval vessel area information in first frame on certain video carrier is sketched the contours of manually;The wherein described naval vessel region letter
Breath be exactly naval vessel in the first frame shared by region, naval vessel area information as described below is also similarly, and to pass through Fast-RCNN nets
The naval vessel area information that network predicts is identical as the concept for the naval vessel area information manually sketched the contours of, the difference is that, manually
It sketches the contours variant with the accuracy rate of the naval vessel area information of Fast-RCNN neural network forecasts.
S2. the video carrier after normalization size is input in Fast-RCNN networks, passes through Fast-RCNN networks pair
Video frame after first frame carries out the tracking monitor in naval vessel region, obtains the location information on the naval vessel in each frame;
Step S3- step S4, as shown in Figure 3:S3. by the location information on naval vessel in human eye observation's present frame, and judge
Whether annotation results are qualified;In eye-observation present frame, the accuracy of naval vessel regional location tracking, the judgment basis of accuracy is such as
Under:In Fast-RCNN networks, removing thought naval vessel regional location is surrounded with a box.If in side
About 80% or more the region on practical naval vessel is not surrounded by frame area encompassed, therefore thinks the tracking knot of present frame
Fruit accuracy is relatively low, needs the manual intervention operation for executing step S4, otherwise continues to execute the update in the naval vessel region of step S2
Operation.
Then include completing naval vessel region by manually drawing a box with mouse if S4. unqualified
After manual intervention operation, continue step S2;Because during Ship Target carries out mobile, if background (in video frame,
Non- naval vessel region) with the color similarity in naval vessel region very it is big either in the current frame, there is shelter, then in target following
During, it may appear that there is mistake (i.e. in current, identifying that other regions are naval vessel region) in tracking result, further, since
During target following, the continuity with target movement, i.e., in the next frame, the position of the target following confirmed
It is in previous frame, near identified target location, the range moved there is no target is excessive, therefore, if working as
In previous frame, the target location of tracking is malfunctioned, then can influence the result of target following in the next frame.So step S3 and step
S4 is essential.Wherein, the specific implementation operation of step S4 is as follows:Tracing area manually is reinitialized with mouse, i.e.,
A box is drawn with mouse, then includes by naval vessel region, to complete manual intervention operation.
S5. above step is all executed to each frame in video carrier, until tracking operation terminates.
The precision accuracy of naval vessel tracking can be greatly promoted by the method for manual intervention.
Embodiment 3, is based on above-described embodiment, and the present invention provides a kind of naval vessel tracking based on manual intervention, this reality
The distinguishing feature for applying example and above-described embodiment is, for the update of tracing area operation, actually exists
In each frame for initializing the video after tracing area, naval vessel region part is equally searched out.For the part, our this patents
Realize that relevant operation, concrete operations are as follows using the Fast-rcnn frames based on deep learning:
S21. target training is carried out to Fast-RCNN networks;Fast-RCNN network models after training are more accurate, institute
The accuracy rate higher of obtained result;Finally carried out again in video carrier by the Fast-RCNN networks after the completion of training
Naval vessel carries out target following.
Embodiment 3, as shown in figure 4, the specific implementation step of the step S21 is:
A. several times are therefrom obtained by selective search methods for each pictures in training set
Favored area, and record, the location coordinate information of each candidate region;Wherein, this several candidate region may exist
Overlapping phenomenon, each region are to be surrounded by a box, therefore location coordinate information is the corresponding upper left of the box
The width and length of angular vertex coordinate and the box, by selective search algorithms (selective search algorithm) we
Relatively coarse naval vessel region can be obtained, then by being trained to neural network, thus from rough naval vessel region,
Obtain out more accurately naval vessel regional location.
Specifically, we use color similarity as the standard for obtaining candidate region, specific implementation method is as follows:
S211., the size and aspect ratio range of candidate region are set;Due in the video frame, the area surface on naval vessel
Product and size be meet it is a certain range of, it would therefore be desirable to first be arranged candidate region relative dimensions.In this step
In, naval vessel region is arranged in we areal extent is [5000,10000] pixel size, length than range be [5,10].And this
In embodiment, by taking 2000 candidate regions as an example.
S212. it chooses in the video frame a bit at random, is denoted as M, then calculate in video frame between any point N and point M
Color distortion degree point N is slid into the range of convergence of point M, the meter of color distortion degree X if color distortion degree is less than K therewith
Calculating formula is:
Wherein, MR、MG、MBPixel values of the point M on tri- components of R, G and B is indicated respectively, similarly NR、NGAnd NGRespectively
Indicate that pixel values of the point N on tri- components of R, G and B, point N are not overlapped with point M, K=150;
S213. selected element N is repeated, and executes diversity factor calculating, expands the range of convergence of point M, the range of convergence of point M is final
An image-region will be constituted, then the image-region just constitutes a part of above-mentioned 2000 candidate regions;
If S214. above-mentioned candidate region meets the size and length-width ratio model for the candidate region being arranged in step S211
It encloses, then terminates step S213;
S215. different point M is randomly selected, and executes step, and executes step S213-S214, and meets and obtains every time
Candidate area size within the scope of size condition in step S211, but candidate area size does not repeat.Such as:The
The image-region once obtained, area are 5001, and length ratio is 5, and the image-region obtained after me can be that area is
5100, length-width ratio is 5.
B. for each candidate region in step A, it is labeled label, i.e., 0 or 1,1 represents the candidate region
Including naval vessel part, 0 indicates that the candidate region does not include naval vessel region part;Correct coordinate information, including a box pair
The information such as the width of the top left corner apex coordinate and box answered and length, and the box then just surrounds naval vessel part
Come.
C. for each region in step A, all there is its correction position coordinate, i.e.,:
If the markup information in current candidate region is 1, illustrate current candidate region RiIn include naval vessel region, we
Record its correct co-ordinate position information, the amendment coordinate information, including the corresponding top left corner apex coordinate of a box with
And the width and length information of box, and naval vessel part is then surrounded by the box just;Wherein, if current region Ri's
Mark is 0, then need not carry out any operation to it, and correction position information is exactly current location coordinate information.
D. several candidate regions constitute the input of network, the mark corresponding to each candidate region and its correspondence
Correction position information constitutes the output of network;
E. Fast-RCNN networks are trained according to the network inputs and output that are obtained in step D, are calculated by using BP
Method updates weight and the biasing of neuron, finally so that Fast-RCNN networks reach convergence state.
Specifically, the structure of simple nervelet network can be as shown in figure 5, wherein each circle represents a nerve
Member, w1And w2The weight between neuron is represented, b indicates biasing, and g (z) is activation primitive, so that output becomes non-linear,
A indicates output, x1And x2It indicates input, is then directed to current structure, output is represented by:
A=g (x1*w1+x2*w2+1*b) (2)
It can be obtained by formula (2), in the case of input data and constant activation primitive, the value a of the output of neural network is
Related with weight and biasing, by adjusting different weight and biasing, the output of neural network also has different results.
The value (predicted value) of known neural network output is a, it is assumed that its corresponding actual value is a'.
For Fig. 5, BP algorithm executes as follows:
It a, can every connecting line weight (w of first random initializtion in BP algorithm1And w2) and biasing b;
B, for input data x1, x2, BP algorithm can all first carry out fl transmission and obtain predicted value a;
C, and then according to the error between actual value a' and predicted value aReverse feedback update nerve net
The weight of every connecting line and every layer of biasing in network.
Shown in weight and the update method of biasing such as formula (3)-(5), i.e., w is asked respectively to E1, w2, the local derviation of b.Wherein η tables
What is shown is learning rate, is a parameter set in this formula.
D, step a-c is constantly repeated, until the value of network convergence, i.e. E is minimum or is held essentially constant.This moment, it indicates
Network is trained to be finished.
By step S21 to Fast-RCNN networks after training, we can use the network carry out target with
The specific implementation step of track, i.e. step S22, the step S22 includes:
Each frame of the video carrier of input is all exported multiple images region (in the frame by S221.Fast-RCNN networks
In, may include multiple Ship Targets), equally it is with a pair of of coordinate points and two parameter w for each candidate region
It is indicated with h.Using this to coordinate points as the top left corner apex of rectangle, w and h indicate the width and length of rectangle respectively, then the rectangle
Area encompassed is the naval vessel region that Fast-RCNN networks are thought.
S222. since the movement on naval vessel has continuity, we can pass through identified tracking area in previous frame
Domain determines in present frame, final Ship Target, and specific implementation is as follows:
S2221. for video initialization tracing area a later frame for, it is assumed that we in the frame, neural network
Output multiple images region Areai, 1≤i≤n, n are the number of image-region in present frame.
S2222. it calculates in present frame, each image-region AreaiWith determined tracing area in previous frame (if upper one
Frame is the video frame carried out corresponding to initialization tracing area, then the tracing area just refer to initialize tracing area) coincidence
Ratio is denoted as Li, 1≤i≤n.Specific such as formula (6) is shown, wherein Sp-1Indicate the determined tracing area in -1 frame of pth
Area,It indicates in pth frame, image-region AreaiCorresponding area, symbol ∩ represents intersection operation, if pth -1
Frame indicates that frame of initialization tracing area, then Sp-1Just the area of initialization tracing area is indicated.
S2223. L is choseniObtain maximum value in the case of, corresponding to image-region carry out the tracking area with new present frame
Domain.
In the present invention, the deep learning frame based on Fast-RCNN networks is used first, to realize to Ship Target
Tracking operation, in addition, in order to prevent during tracking, the tracking error of appearance (since there are shelters so that target with
Track is lost) the case where, we use manual intervention, the artificial method for correcting tracing area, to improve tracking result
Accuracy.
The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be said that
The specific implementation of the present invention is confined to these explanations.For those of ordinary skill in the art to which the present invention belongs, exist
Under the premise of not departing from present inventive concept, a number of simple deductions or replacements can also be made, all shall be regarded as belonging to the present invention's
Protection domain.
Claims (10)
1. a kind of Ship Target tracking based on manual intervention, which is characterized in that include the following steps:
S1. the naval vessel area information in first frame on certain video carrier is sketched the contours of manually;
S2. the video carrier after normalization size is input in Fast-RCNN networks, passes through Fast-RCNN networks pair first
Frame after frame carries out the tracking monitor in naval vessel region, obtains the location information on the naval vessel in each frame;
S3. by the location information on naval vessel in human eye observation's present frame, and judge whether annotation results are qualified;
Then include completing naval vessel region manually by manually drawing a box with mouse if S4. unqualified
After intervention operation, continue step S2;
S5. above step is all executed to each frame in video carrier, until tracking operation terminates.
2. according to the method described in claim 1, it is characterized in that, the step S2 is further comprising the steps of:
S21. target training is carried out to Fast-RCNN networks;
S22. Fast-RCNN networks are used to carry out target following.
3. according to the method described in claim 2, it is characterized in that, the step S21 is further comprising the steps of:
A. several candidate regions are therefrom obtained by selective search methods for each pictures in training set
Domain, and record, the location coordinate information of each candidate region;
B. for each candidate region in step A, it is labeled label, i.e., 0 or 1,1, which represents the candidate region, includes
Naval vessel part, 0 indicates that the candidate region does not include naval vessel region part;
C. for each region in step A, all there is its correction position coordinate, i.e.,:
If the markup information in current candidate region is 1, illustrate current candidate region RiIn include naval vessel region, we record
Lower its corrects co-ordinate position information, the amendment coordinate information, including the corresponding top left corner apex coordinate of a box and side
The width and length information of frame, and naval vessel part is then surrounded by the box just;
D. several candidate regions constitute the input of network, and the mark and its correspondence corresponding to each candidate region are corrected
Location information constitutes the output of network;
E. Fast-RCNN networks are trained according to the network inputs and output that are obtained in step D, are come by using BP algorithm
Weight and the biasing of neuron are updated, finally so that Fast-RCNN networks reach convergence state.
4. according to the method described in claim 3, it is characterized in that, the step A includes the following steps:
S211., the size and aspect ratio range of candidate region are set;
S212. it chooses in the video frame a bit at random, is denoted as M, then calculate the face in video frame between any point N and point M
Color diversity factor slides into point N if color distortion degree is less than K therewith in the range of convergence of point M, and the calculating of color distortion degree X is public
Formula is:
Wherein, wherein MR、MG、MBPixel values of the point M on tri- components of R, G and B is indicated respectively, similarly NR、NGAnd NGRespectively
Indicate that pixel values of the point N on tri- components of R, G and B, point N are not overlapped with point M, K=150;
S213. selected element N is repeated, and executes diversity factor calculating, expands the range of convergence of point M, the range of convergence of point M is finally by structure
At an image-region, then the image-region just constitutes a part of several above-mentioned candidate regions;
If S214. above-mentioned candidate region meets the size and aspect ratio range for the candidate region being arranged in step S211,
Then terminate step S213;
S215. different point M is randomly selected, and executes step, and executes step S213-S214, and meets the time obtained every time
Constituency domain sizes are within the scope of the size condition in step S211, but candidate area size does not repeat.
5. according to the method described in claim 2, it is characterized in that, the step S22 includes the following steps:
Each frame of the video carrier of input is all exported multiple images region by S221.Fast-RCNN networks, for each image
Region is indicated with a pair of of coordinate points and two parameters w and h;Using this to coordinate points as the top left corner apex of rectangle, w and h difference
Indicate the width and length of rectangle, then the rectangle area encompassed is the naval vessel region that network is thought;
S222. the location information on naval vessel final in present frame is determined by identified tracing area in previous frame.
6. according to the method described in claim 5, it is characterized in that, the step S222 includes the following steps:
S2221. for a later frame of the initialization tracing area of video carrier, it is assumed that in the frame, neural network output
Multiple images region Areai, wherein 1≤i≤n, n are the number of image-region in present frame;
S2222. pass through formula:
It calculates in present frame, each image-region AreaiWith determined tracing area in previous frame, coincidence ratio Li, 1≤i≤
N, wherein Sp-1Indicate the area of the determined tracing area in -1 frame of pth,It indicates in pth frame, image-region AreaiInstitute
Corresponding area, symbol ∩ represent intersection operation, if -1 frame of pth indicates that frame of initialization tracing area, Sp-1Just table
Show the area of initialization tracing area;
S2223. L is choseniIn the case of obtaining maximum value, corresponding to image-region update the tracing area of present frame.
7. according to the method described in claim 1, it is characterized in that, in the step S3, whether the tracking of naval vessel regional location
Qualification determination is according to being:In Fast-RCNN networks, the naval vessel regional location thought has been surrounded with a box
Come, if be not surrounded practical 80% or more naval vessel region in box area encompassed, therefore thinks present frame
Tracking result accuracy is relatively low, unqualified, thens follow the steps S4.
8. according to the method described in claim 1, it is characterized in that, the step S4 includes:Manually reinitialized with mouse
Tracing area draws a box with mouse, then includes by naval vessel region.
9. according to the method described in claim 3, it is characterized in that, in the step A, pass through the sides search selective
Method therefrom obtains 2000 candidate regions.
10. according to the method described in claim 4, it is characterized in that, the range of the size of the candidate region is
[5000,10000] pixel size, length than range be [5,10].
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810273119.4A CN108537826A (en) | 2018-05-28 | 2018-05-28 | A kind of Ship Target tracking based on manual intervention |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810273119.4A CN108537826A (en) | 2018-05-28 | 2018-05-28 | A kind of Ship Target tracking based on manual intervention |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108537826A true CN108537826A (en) | 2018-09-14 |
Family
ID=63481615
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810273119.4A Pending CN108537826A (en) | 2018-05-28 | 2018-05-28 | A kind of Ship Target tracking based on manual intervention |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108537826A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110222632A (en) * | 2019-06-04 | 2019-09-10 | 哈尔滨工程大学 | A kind of waterborne target detection method of gray prediction auxiliary area suggestion |
CN110348356A (en) * | 2019-07-03 | 2019-10-18 | 北京遥感设备研究所 | A kind of successive frame RD images steganalysis method based on depth light stream network |
CN110413166A (en) * | 2019-07-02 | 2019-11-05 | 上海熙菱信息技术有限公司 | A kind of method of history video real time linear tracking |
CN111062298A (en) * | 2019-12-11 | 2020-04-24 | 深圳供电局有限公司 | Power distribution network power equipment target identification method and system |
CN116343049A (en) * | 2023-05-24 | 2023-06-27 | 四川创意科技有限公司 | Method, device, equipment and storage medium for monitoring abnormal behavior of offshore target |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106683117A (en) * | 2016-12-30 | 2017-05-17 | 佳都新太科技股份有限公司 | Target grasping algorithm based on kinematics behavior analysis |
CN106934332A (en) * | 2015-12-31 | 2017-07-07 | 中国科学院深圳先进技术研究院 | A kind of method of multiple target tracking |
CN106960446A (en) * | 2017-04-01 | 2017-07-18 | 广东华中科技大学工业技术研究院 | A kind of waterborne target detecting and tracking integral method applied towards unmanned boat |
US20170206431A1 (en) * | 2016-01-20 | 2017-07-20 | Microsoft Technology Licensing, Llc | Object detection and classification in images |
CN107292297A (en) * | 2017-08-09 | 2017-10-24 | 电子科技大学 | A kind of video car flow quantity measuring method tracked based on deep learning and Duplication |
CN107564004A (en) * | 2017-09-21 | 2018-01-09 | 杭州电子科技大学 | It is a kind of that video labeling method is distorted based on computer auxiliary tracking |
CN107818571A (en) * | 2017-12-11 | 2018-03-20 | 珠海大横琴科技发展有限公司 | Ship automatic tracking method and system based on deep learning network and average drifting |
-
2018
- 2018-05-28 CN CN201810273119.4A patent/CN108537826A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106934332A (en) * | 2015-12-31 | 2017-07-07 | 中国科学院深圳先进技术研究院 | A kind of method of multiple target tracking |
US20170206431A1 (en) * | 2016-01-20 | 2017-07-20 | Microsoft Technology Licensing, Llc | Object detection and classification in images |
CN106683117A (en) * | 2016-12-30 | 2017-05-17 | 佳都新太科技股份有限公司 | Target grasping algorithm based on kinematics behavior analysis |
CN106960446A (en) * | 2017-04-01 | 2017-07-18 | 广东华中科技大学工业技术研究院 | A kind of waterborne target detecting and tracking integral method applied towards unmanned boat |
CN107292297A (en) * | 2017-08-09 | 2017-10-24 | 电子科技大学 | A kind of video car flow quantity measuring method tracked based on deep learning and Duplication |
CN107564004A (en) * | 2017-09-21 | 2018-01-09 | 杭州电子科技大学 | It is a kind of that video labeling method is distorted based on computer auxiliary tracking |
CN107818571A (en) * | 2017-12-11 | 2018-03-20 | 珠海大横琴科技发展有限公司 | Ship automatic tracking method and system based on deep learning network and average drifting |
Non-Patent Citations (7)
Title |
---|
J.R.R.UIJLINGS,AND ETC: "Selective Search for Object Recognition", 《SPRINGERSCIENCE》 * |
周云成等: "基于深度卷积神经网络的番茄主要器官分类识别方法", 《农业工程学报》 * |
徐超等: "改进的卷积神经网络行人检测方法", 《计算机应用》 * |
曹诗雨等: "基于Fast R-CNN的车辆目标检测", 《中国图象图形学报》 * |
曾向阳著: "《智能水中目标识别》", 31 March 2016, 北京:国防工业出版社 * |
王一丁等: "《数字图像处理》", 31 August 2015, 西安:西安电子科技大学出版社 * |
邱莉榕等: "《算法设计与优化》", 31 December 2016, 北京:中央民族大学出版社 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110222632A (en) * | 2019-06-04 | 2019-09-10 | 哈尔滨工程大学 | A kind of waterborne target detection method of gray prediction auxiliary area suggestion |
CN110413166A (en) * | 2019-07-02 | 2019-11-05 | 上海熙菱信息技术有限公司 | A kind of method of history video real time linear tracking |
CN110413166B (en) * | 2019-07-02 | 2022-11-25 | 上海熙菱信息技术有限公司 | Real-time linear tracking method for historical video |
CN110348356A (en) * | 2019-07-03 | 2019-10-18 | 北京遥感设备研究所 | A kind of successive frame RD images steganalysis method based on depth light stream network |
CN111062298A (en) * | 2019-12-11 | 2020-04-24 | 深圳供电局有限公司 | Power distribution network power equipment target identification method and system |
CN116343049A (en) * | 2023-05-24 | 2023-06-27 | 四川创意科技有限公司 | Method, device, equipment and storage medium for monitoring abnormal behavior of offshore target |
CN116343049B (en) * | 2023-05-24 | 2023-08-15 | 四川创意科技有限公司 | Method, device, equipment and storage medium for monitoring abnormal behavior of offshore target |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108537826A (en) | A kind of Ship Target tracking based on manual intervention | |
US20210023720A1 (en) | Method for detecting grasping position of robot in grasping object | |
CN110472554B (en) | Table tennis action recognition method and system based on attitude segmentation and key point features | |
CN106650630B (en) | A kind of method for tracking target and electronic equipment | |
CN108549876A (en) | The sitting posture detecting method estimated based on target detection and human body attitude | |
CN109271888A (en) | Personal identification method, device, electronic equipment based on gait | |
CN105205453B (en) | Human eye detection and localization method based on depth self-encoding encoder | |
CN108229268A (en) | Expression Recognition and convolutional neural networks model training method, device and electronic equipment | |
CN109285179A (en) | A kind of motion target tracking method based on multi-feature fusion | |
CN108803617A (en) | Trajectory predictions method and device | |
CN109176512A (en) | A kind of method, robot and the control device of motion sensing control robot | |
CN104751466B (en) | A kind of changing object tracking and its system based on conspicuousness | |
CN107886089A (en) | A kind of method of the 3 D human body Attitude estimation returned based on skeleton drawing | |
CN108171141A (en) | The video target tracking method of cascade multi-pattern Fusion based on attention model | |
CN111192294B (en) | Target tracking method and system based on target detection | |
CN108197584A (en) | A kind of recognition methods again of the pedestrian based on triple deep neural network | |
CN106373160A (en) | Active camera target positioning method based on depth reinforcement learning | |
CN104299245A (en) | Augmented reality tracking method based on neural network | |
CN108520218A (en) | A kind of naval vessel sample collection method based on target tracking algorism | |
CN108765468A (en) | A kind of method for tracking target and device of feature based fusion | |
CN106803084A (en) | A kind of facial characteristics independent positioning method based on end-to-end recirculating network | |
CN108334878A (en) | Video images detection method and apparatus | |
CN109087337A (en) | Long-time method for tracking target and system based on layering convolution feature | |
CN110472628A (en) | A kind of improvement Faster R-CNN network detection floating material method based on video features | |
CN107274437A (en) | A kind of visual tracking method based on convolutional neural networks |
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 |
Application publication date: 20180914 |
|
RJ01 | Rejection of invention patent application after publication |