CN109978851A - A kind of aerial weak moving target detection tracking of infrared video - Google Patents
A kind of aerial weak moving target detection tracking of infrared video Download PDFInfo
- Publication number
- CN109978851A CN109978851A CN201910220147.4A CN201910220147A CN109978851A CN 109978851 A CN109978851 A CN 109978851A CN 201910220147 A CN201910220147 A CN 201910220147A CN 109978851 A CN109978851 A CN 109978851A
- Authority
- CN
- China
- Prior art keywords
- target
- detection
- frame
- present frame
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/207—Analysis of motion for motion estimation over a hierarchy of resolutions
-
- 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
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10048—Infrared image
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention proposes a kind of aerial weak moving target detection trackings of infrared video.This method proposes fast and accurately skyline decision plan, perception obtains Sky Scene region, and then on high under the guidance of scene areas sensing results, aerial potential target is obtained from spatial filter by the significant difference of target and background, guarantees real goal not missing inspection;Target movement tendency, the removal false-alarm interference of multiple target similarity Comprehensive Evaluation are introduced, moving target is further confirmed that in conjunction with Background Motion Compensation, solves the problems, such as that moving target accurately tracks.The experimental results showed that this method robustness is good, be blocked in target, be overlapped, disappearing and the complex situations such as background variation and target are flickering under also can be carried out accurate, steady tracking.
Description
Technical field
The present invention relates to a kind of methods for the aerial weak moving target detection tracking of infrared video, belong to computer view
Feel technical field.
Background technique
With the rapid development of precision guided weapon, future war is to the overall performance of weapon system, especially adaptability
More stringent requirements are proposed.Infrared guidance technology is excellent with its extremely strong anti-interference ability, high guidance precision, good concealment etc.
Gesture is widely used in all kinds of guided missiles, aircraft and ground system, becomes the second generation guidance technology of controllable weapon system.In the people
With being also gradually widely applied in system.One infrared guidance system generally by target detection, target identification, target following and
The functional modules composition such as point of attack selection.Wherein, target detection and processing links of the target following as system front end are essences
The important component really guided.For the infrared video obtained using infrared guidance system, the base of target detection tracking
Present principles are captures and track the energy of target itself radiation to realize homing.However, working as shooting distance in practical application
Target imaging area often very little when farther out, or even point target is shown as, the information such as structure, shape can not be obtained, and can examine
It is relatively weak to survey signal, especially under the interference of non-stationary varying background, target is easily flooded by background clutter, so that infrared view
The accurate detecting and tracking of frequency Weak target is extremely difficult.
The missing of the diversity and Weak target information of target and background environment, so that being built based on infrared signature
The object detection method of mould is difficult to carry out.Therefore, existing infrared video moving object detection tracking is typically directly using traditional
The thinking of visible light video moving object detection tracking, is divided into two methods of Detect before Track and root-first search.
The method of Detect before Track usually first carries out background to input image sequence and inhibits pretreatment, to improve single frames figure
A possibility that detecting target as in;Then, using thresholding method, texture analysis method, morphological approach etc., in conjunction with best
The criterion such as Bayes judgement, minimax judgement or Neynan-Pearson judgement, make decisions pixel each in image,
Determine whether it is potential target point;Further according to the interframe high correlation of target image motion, the time product of multiple image is utilized
It is tired that majority voting is carried out to single frame detection result, Second Decision is formed to reject false-alarm point, retains the target finally confirmed.So
And when target signal to noise ratio (SIGNAL-NOISE RATIO, SNR) is lower, the method for this Detect before Track may be because
The missing inspection of Weak target, the erroneous detection of false-alarm targets and fail.
Thought of the method for root-first search according to multi frame detection, is incorporated to multi-frame processing process for room and time information
In.Possible target trajectory is tracked first, but whether is not that real target is entered a judgement to these tracks, but it is right
Trajectory calculation its posteriority probability function of every tracking, if the posterior probability functional value of certain track is more than a certain thresholding, just
Think this track represent a target (refer to " the quick detection of infrared small object under Celestial Background ", publish in " laser with it is red
Outside ", in April, 2017).Common root-first search method has Hough transform method, three-dimensional matched-filter approach (to refer to " base
In the small IR targets detection and tracking of Noise Variance Estimation ", publish in " optoelectronic laser ", in March, 2018), it is more
Grade hypothesis testing method, dynamic programming etc..The method of root-first search under complex jamming to Weak target, especially slowly
The adaptability of moving target is insufficient, and in platform movement, complex background and there are the interference of other homologues, targets to be blocked, mesh
When marking the motion of automobile, such method is then difficult to carry out continuous, accurate, steady long-time tracking.
For this purpose, present applicant proposes a kind of aerial weak moving target detection trackings of infrared video.It proposes quickly quasi-
True skyline decision plan, perception obtain Sky Scene region;And then pass through under the guidance of scene areas sensing results on high
The significant difference of target and background obtains aerial potential target from spatial filter, guarantees real goal not missing inspection;Introduce target
Movement tendency, the removal false-alarm interference of multiple target similarity Comprehensive Evaluation, further confirm that moving target in conjunction with Background Motion Compensation,
Solve the problems, such as that moving target accurately tracks.
Summary of the invention
The object of the present invention is to provide a kind of aerial weak moving target detection tracking based on infrared video.The party
Method not only considers that the otherness of real goal and background is detected in conventional method, the similitude of real goal and candidate target
It is tracked, and sky areas is obtained by scene perception, target detection, introducing multiple target similarity Comprehensive Evaluation is instructed to refer to
Tracking is led, the thought of Detect before Track, root-first search is fused together, realizes the rejecting of false target and true
Real target accurately tracks.
To achieve the above object, the technical solution adopted by the present invention process is as follows:
A kind of aerial weak moving target detection tracking of infrared video, it is characterised in that include the following steps:
(1) airspace target detection: skyline is extracted using scene areas perception, divides region in aerial region and non-empty;
And then region is based on TopHat transformation progress conspicuousness detection in the sky, realizes the pixel inspection of aerial conspicuousness Weak target
It surveys;And then several targets are obtained using target cluster segmentation to the pixel detected, and it is true to carry out signal-to-noise ratio to these targets
Recognize, final detection obtains well-marked target;It is tracked in initial frame using these well-marked target initialized targets that present frame detects
Chain;
(2) target following: target detection chain is constructed using well-marked target obtained above in present frame, with target following chain
In target carry out it is positive, reversed compare, establish the correspondence of the target in target detection chain and the target in target following chain, from
And target is positioned in the position of present frame;It is pre- using movement tendency for the target that do not establish in corresponding target following chain
The position target is measured in the position of present frame;For the target that do not establish in corresponding target detection chain, there is mesh as new
Mark is added in target following chain;
(3) time domain variation detection: time domain variation detection is carried out using present frame and preceding nth frame image, in two field pictures
After extracting SURF characteristic point respectively, affine Transform Model is calculated to estimate background motion, in compensation background based on Feature Points Matching
Picture frame after being compensated after movement, based on picture frame after present frame, compensation, to aforementioned obtained object judgement, whether it is transported
It is dynamic, target trajectory is formed to the target for being judged as movement, realizes target following.
Method as described above, it is characterised in that the detailed process of scene perception in step (1) are as follows: in present frame, successively
Its pixel value is traversed from bottom to up to each column image, the pixel value of previous pixel is i1, the pixel value of latter pixel is
i2, work as i2-i1≥Th1And i1<Th2When, show to initially enter sky areas at this time, wherein Th1, Th2Difference uncarved areas type
Variation and pixel value bright-dark degree, Th1Referred to as qualitative change threshold value, Th2Referred to as dark matter threshold value;The current of sky areas will be initially entered
Pixel coordinate is denoted as (k, l), and record persistently belongs to the pixel number Count, i.e. these pixels i of sky areas2With phase
Adjacent pixel i1Meet d=| i1-i2|<Th1And i2<Th2;When the condition is unsatisfactory for, that is, it is unsatisfactory for d=| i1-i2|<Th1And
i2<Th2When, current pixel point coordinate is (k+Count-1, l) to note at this time;This section of from (k, l) to (k+Count-1, l) is vertical
Line, pixel and neighbor pixel meet d=| i1-i2|<Th1And i2<Th2, belong to sky areas;It is each to present frame arrange into
The judgement of the above-mentioned sky areas of row, obtains several segments sky areas;These sky areas are traversed from bottom to up, most by duration ranges
Greatly, i.e. horizon line position of the vertical line lower end of the maximum sky areas of Count number as the image column;To each of present frame
Column detect horizon line position according to the method described above, then the horizon line position connection of each column can be obtained to the day of full frame image
Border line;Sky areas is used as on skyline.
Method as described above, it is characterised in that based on the well-marked target pixel detection of TopHat transformation in step (1)
Detailed process are as follows: in the sky areas that above-mentioned scene perception obtains, present frame and morphological operator are subjected to opening operation, used
It is that present frame subtracts opening operation as a result, obtain TopHat transformation as a result, i.e. Dst=TopHat (I, Element)=I-Open
(I,Element);Wherein I represents present frame, and Element represents morphological operator.
Method as described above, it is characterised in that the detailed process of target cluster segmentation in step (1) are as follows: become TopHat
The result changed carries out binary conversion treatment, and binarization threshold Th is arranged3, the result of TopHat transformation is greater than Th3It is then set to 1, is otherwise set
It is 0;Pixel after binary conversion treatment for 1 carries out neighborhood cluster, treatment process are as follows: random selection as suspected target point
For one suspected target point as initial clustering seed, 4 neighborhoods or 8 neighborhoods to this seed carry out breadth first search, if
Other suspected target points can be found then to illustrate that the two pixels distance is close, belong to same target, it should cluster as same mesh
Mark;Finally all cluster is attributed to some target to suspected target point after all binaryzations.
Method as described above, it is characterised in that the detailed process that target signal to noise ratio confirms in step (1) are as follows: Statistical Clustering Analysis
The Local modulus maxima of target, if its gray value is g0, to form 3 × 3 neighborhoods and 9 × 9 neighborhoods centered on it;In 9 × 9 neighborhoods
After removing 9 pixels in 3 × 3 contiguous ranges, remaining 72 pixels form the background area around target, gray value
It is denoted as gi, i=1 ... 72;According to Signal to Noise Ratio (SNR) formula (1)-(3), the signal-to-noise ratio of current goal is obtained;Work as target signal to noise ratio
SNR is less than snr threshold Th4, i.e. SNR < Th4, then the target is filtered;Noise is carried out to each cluster target in present frame
The calculating of ratio, the final real goal for confirming airspace conspicuousness and detecting.
Method as described above, it is characterised in that sky areas and non-sky area described in scene perception in step (1)
Transition method in, qualitative change threshold value Th1Preferably 25, dark matter threshold value Th2Preferably 50.
Method as described above, it is characterised in that the well-marked target pixel detection side that step (1) is converted based on TopHat
In method, morphological operator Element chooses size slightly larger than detection target, preferably
Method as described above, it is characterised in that TopHat transformation results are carried out when target cluster segmentation in step (1)
Binary conversion treatment, binarization threshold Th3Preferably 110.
Method as described above, it is characterised in that target signal to noise ratio threshold value Th in step (1)4Preferably 6, show signal-to-noise ratio
Target of the SNR greater than 6 is removed.
Method as described above, it is characterised in that target detection chain is compared with target following chain forward direction, reversely in step (2)
Detailed process are as follows: forward direction compares: taking out each of target detection chain detection target DObj, calculated separately in its M neighborhood
Distance detection target DObjTarget T in nearest target following chainObj, will distance DObjNearest TObjIt is considered as same target;Instead
To comparison: taking out each of target following chain tracking target TObj, in its M contiguous range, calculate target detection chain in
Track target TObjNearest detection target D 'Obj;Forward and reverse result compares: judging DObjWith D 'ObjIt whether is the same target, such as
Fruit is the same detection target, then with the target in the corresponding target following chain of the detection target update, forms the present frame mesh
Target pursuit path;Otherwise the pursuit path to the present frame target is formed according to the result of motion prediction.
Method as described above, it is characterised in that the detailed process that target movement tendency is predicted in step (2) are as follows: assuming that Δ
X, Δ y respectively indicate former frame of the target from initial value frame to present frame in moving distance horizontal, in vertical direction, Δ t table
Show the frame number of the former frame from initial frame to present frame, then target speed can use vx=Δ x/ Δ t, vy=Δ y/ Δ t
To indicate;The position of the target predicted in tracking in the current frame is x=x0+ Δ x, y=y0+ Δ y, wherein x0,y0Table
Show target in the position of the former frame of present frame;By 10. front-to-back ratio of claim to later, may be gone back in target following chain
Fail to match with the detection target in detection chain in the presence of tracking target, thinks that target is blocked or disappears at this time, take fortune
Dynamic prediction positioning present frame target position, target position of every prediction then predict that total degree adds 1 to it, when prediction total degree is super
It crosses after prediction frequency threshold value P and is believed that target has disappeared, at this time delete target.
Method as described above, it is characterised in that M neighborhood is preferably formed centered on selected target in step (2) 10 ×
10 neighborhood of pixels, prediction frequency threshold value P is preferably 10, indicates to also fail to position the target after 10 frames, then deletes the target, no
It is tracked again.
Method as described above, it is characterised in that step (3) time domain variation detection the specific process is as follows: current
SURF characteristic point is extracted respectively in frame, preceding nth frame image, calculates affine Transform Model based on Feature Points Matching, using RANSACTo estimate background motion;The picture frame after being compensated after compensating background motion, based on current
Picture frame after frame, compensation, to aforementioned obtained object judgement, whether it is moved, it is assumed that and preceding nth frame coordinates of targets is denoted as (x, y),
The coordinates of targets that corresponds in the current frame for estimating to obtain is (x', y'),And practical present frame mesh
Marking coordinate is (x ", y "), the Euclidean distance obtained by comparingThreshold distance threshold value is set
Th5If S < Th5, then it is assumed that active movement is not present in the target, belongs to stationary object, is not determined as moving target;Such as
Fruit S >=Th5, then it is assumed that it is moving target, without rejecting.
Method as described above, it is characterised in that the reference frame when variation detection of step (3) time domain selects preceding N=3 frame, Europe
Formula distance threshold Th5Preferably 3.
Compared with prior art, Sky Scene region provided by the present invention perceives, becomes in conjunction with airspace conspicuousness and time domain
Target detection, the method for movement tendency and multiple target similarity Comprehensive Evaluation removal false target for changing characteristic, avoid tradition
Method floods the limitation of target missing inspection caused by target due to Weak target loss of learning, ambient noise, solves unintentionally
Caused by adopted noise, multiple target interfere with each other the problem of target erroneous detection, tracking error, the real-time of detecting and tracking and accurate is improved
Property.Experimental result shows that, when the movement velocity of target relative scene is greater than 1 pixel/frame, this method is able to detect and tracing surface
Product is greater than 6 moving target in 1-30 pixel, target signal to noise ratio, with the spy that accuracy is high, robustness is good, practical
Point.
Detailed description of the invention
The present invention is further illustrated with reference to the accompanying drawings and detailed description.
Fig. 1 is the aerial weak moving target detection tracking overall framework of infrared video of the present invention;
Fig. 2 is the testing process of skyline in scene perception;
Fig. 3 is the final effect picture of the present invention;
Fig. 4 is target signal to noise ratio calculation method schematic diagram;
Fig. 5 is the spatial relation figure of target detection chain target corresponding with target following chain.
Specific embodiment
Preceding to have addressed, the present invention proposes a kind of aerial weak moving target detection tracking of infrared video, below with reference to
Detailed description of the invention a specific embodiment of the invention.
(1) general frame
Fig. 1 describes general frame of the invention, and the thinking of Detect before Track, root-first search is merged one
It rises, is divided into three parts: (1) airspace target detection;(2) time domain target detection;(3) motion target tracking.
Firstly, carrying out spatial filter in current frame image: extracting skyline using scene areas cognition technology, divide empty
Region in middle region and non-empty;And then region is detected to obtain the well-marked target highlighted in background based on conspicuousness in the sky, realizes
The pixel detection of aerial conspicuousness Weak target, lays the foundation for Weak target not missing inspection.And then based on target cluster segmentation,
Signal-to-noise ratio confirmation, detection obtain well-marked target.
Meanwhile tim e- domain detection is carried out using present frame and preceding nth frame image: extracting characteristic point respectively in two field pictures
Afterwards, affine Transform Model is calculated to estimate background motion based on Feature Points Matching, schemed after being compensated after compensating background motion
As frame judges whether it moves to aforementioned obtained well-marked target based on picture frame after present frame, compensation.Utilize the long period
Target motion change be slow moving target not missing inspection lay the foundation.
Motion target tracking: target detection chain is constructed using aforementioned obtained present frame well-marked target, with target following chain
In target be compared, guarantee that target is of short duration disappears in conjunction with motion prediction, multiple target similarity Comprehensive Evaluation, time domain target detection
Lose or multi-target track intersect when do not leak with accidentally with accurate detecting and tracking moving target.
(2) airspace target detection
The key of airspace target detection is perception scene areas type, knows sky areas, and then in the sky based on significant
Property detect to obtain the pixel highlighted in background, then by cluster etc. post-processings obtain target.
(2.1) scene perception.For the detecting and tracking of aerial target, if it is possible to which tracing area to be detected is limited to day
Empty part, the accuracy that significant increase is detected.The present invention proposes a kind of quick infrared image scene perception method, effectively mentions
Skyline is taken, the image-region on skyline is considered as sky areas, and the image-region under skyline is considered as non-sky area.
The detailed process of scene perception are as follows: to present frame, its pixel value is successively traversed from bottom to up to each column image, it is false
If the pixel value of previous pixel is i1, the pixel value of latter pixel is i2, then the pixel value variation of the two neighbor pixels
If it is less than threshold value Th1, i.e. d=| i1-i2|<Th1, then it is assumed that it belongs to the same area, referred to as homogeneous region;Otherwise d=| i1-
i2|≥Th1, it is believed that non-homogeneous region.Notice that the light and shade of sky and non-sky area on pixel value is different, sky areas phase
To darker, work as i1-i2≥Th1And i2<Th2When, it indicates that the transition from non-sky area to sky areas has occurred, works as i2-i1≥
Th1And i1<Th2When, it indicates that the transition from sky areas to non-sky area has occurred.Th1Referred to as qualitative change threshold value, Th2It is referred to as dark
Matter threshold value.Th1Preferably 25, Th2Preferably 50.
Since sky areas is often positioned in the top half of image, we pay close attention to i1-i2≥Th1And i2<Th2Feelings
Condition, occurs the transition from non-sky area to sky areas at this time, that is, shows that sky areas may be initially entered at this time.It will be current
Pixel coordinate is denoted as (k, l), and record persistently belongs to the pixel number Count, i.e. these pixels i of sky areas2With phase
Adjacent pixel i1Meet d=| i1-i2|<Th1And i2<Th2;When the condition is unsatisfactory for, that is, it is unsatisfactory for d=| i1-i2|<Th1And
i2<Th2When, current pixel point coordinate is (k+Count-1, l) to note at this time.This section of coordinate from (k, l) to (k+Count-1, l)
Vertical line, pixel and neighbor pixel meet d=| i1-i2|<Th1And i2<Th2, this section of vertical line belong to sky areas.Figure
As the lower left corner is as coordinate origin.
The judgement that each column of present frame are carried out with above-mentioned sky areas, obtains several segments sky areas.It traverses from bottom to up
These sky areas, by duration ranges maximum, i.e., the vertical line lower end of the maximum sky areas of Count number is as the image column
Horizon line position.
Horizon line position is detected according to the method described above for each image column of each frame, then by the day of each image column
Border line position connects the skyline that full frame image can be obtained, the i.e. line of demarcation of sky and ground, realizes scene perception.
(2.2) it is based on TopHat change detection Weak target.The purpose of the present invention is detecting aerial Weak target, especially
Point target.In the sky areas of above-mentioned acquisition, using morphology TopHat change detection Weak target.The tool of TopHat transformation
Body treatment process are as follows: present frame and morphological operator are subjected to opening operation, with present frame subtract opening operation as a result, obtaining
TopHat transformation as a result, i.e. Dst=TopHat (I, Element)=I-Open (I, Element).Wherein I represents present frame,
Element represents morphological operator, chooses size slightly larger than detection target.
It is preferred that
The micro bright point in full frame image can be detected by TopHat transformation, reach and detect that highlight background weak
The purpose of Small object.
(2.3) neighborhood clusters.It, will in order to preferably highlight target although TopHat transformation highlights bright spot
The result of TopHat transformation carries out binary conversion treatment: setting binarization threshold Th3, the result of TopHat transformation is greater than Th3Then it is set to
1, otherwise it is set to 0.Pixel after binary conversion treatment for 1 carries out neighborhood cluster as suspected target point.The purpose of neighborhood cluster
It is the very close suspected target point aggregation of Euclidean distance to be become into a biggish target, and generate target minimum circumscribed rectangle work
For target frame.This have the advantage that the operation after being allowed to is carried out based on target, without to a object pixels up to ten thousand
Point is screened, and the efficiency of algorithm is improved.The treatment process of neighborhood cluster are as follows: at random that suspected target o'clock is initial poly- as one
Class seed, 4 neighborhoods or 8 neighborhoods to this seed carry out breadth first search, if it is possible to find other suspected target points then
Illustrate that the two pixels distance is close, belong to same target, it should cluster as same target.What is obtained after all binaryzations is doubtful
Target point should all belong to the target after a certain cluster.
(2.4) target signal to noise ratio confirms, filters noise.The target that neighborhood clusters is likely present erroneous detection, need by
It is distinguished with real goal, therefore is filtered using signal-to-noise ratio.The Local modulus maxima of Statistical Clustering Analysis target first, with
3 × 3 neighborhoods and 9 × 9 neighborhoods are formed centered on it.Assuming that as shown in Fig. 4, wherein black color dots represent cluster and obtain the office of target
Portion's maximum point, if its gray value is g0.White area is 3 × 3 neighborhoods around Local modulus maxima, and shadow region is local pole
9 × 9 neighborhood of pixels (removal white area) around big value point, the gray value of shadow region is gi, i=1 ... 72;.Under
The signal-to-noise ratio that formula (1)-(3) calculate target and peripheral region is stated, if signal-to-noise ratio is less than target signal to noise ratio threshold value Th4, then say
The signal of its bright opposite background is weak, is considered noise with great probability, it should be filtered.Selected objective target snr threshold Th4
=6, show that target of the Signal to Noise Ratio (SNR) greater than 6 is removed.By the meter for carrying out signal-to-noise ratio to each cluster target in picture frame
It calculates, the final real goal for confirming airspace conspicuousness and detecting.
(3) motion target tracking
(3.1) initial frame target following
In initial frame, the object initialization that spatial filter is obtained is the target in target following chain, and initialized target
The information such as position, size.
(3.2) subsequent each frame target following
In present frame, the target that spatial filter is obtained is as the target in target detection chain.In this way, in subsequent processing
Front-to-back ratio carried out according to multiple target similarity Comprehensive Evaluation to the prediction of, target movement tendency, newly there is target detection safeguarding
Target detection chain, target following chain.
The front-to-back ratio pair of (3.2.1) based on multiple target similarity Comprehensive Evaluation
In present frame, to the target in target detection chain and target following chain, positive comparison is carried out first: taking out target inspection
Each of surveyor's chain detects target DObj, distance detection target D is calculated separately in its M neighborhoodObjNearest target following chain
In target TObj, since Weak target signal is weak, it is positioned proximate to account for deciding factor in target following, in principle will
Distance DObjNearest TObjIt is considered as same target.However, it is above-mentioned it is positive compare be obtained by target detection chain with detection
The nearest tracking target of target, and may exist in actual conditions and tracking target TObjNearest detection target not be
DObj, spatial relation is as shown in Fig. 5.It is influenced by target traversal order in target detection chain, the mesh when forward direction compares
No. 1 target in mark detection chain can be nearest with No. 2 targets in target following chain, to establish the mistake of " detection 1- tracking 2 "
Match, and " tracking 2 " after target is occupied, No. 2 targets in subsequent target detection link can only establish " detection 2- tracking 1 "
Erroneous matching.And from No. 1 target in target following chain can be the discovery that No. 1 target in target detection chain and its most
Match.Therefore need reversely to be compared after forward direction compares successfully: reversed comparison from target following chain, taking-up target with
Each of track chain tracks target TObj, in its M contiguous range, calculate target detection chain in tracking target TObjNearest
Detect target D 'Obj.Judge DObjWith D 'ObjIt whether is the same target, if it is the same detection target, then with the detection mesh
Mark updates the target in corresponding target following chain, forms the pursuit path of the present frame target;Otherwise pre- according to following movements
The result of survey forms the pursuit path to the present frame target.
10 × 10 neighborhood of pixels that M neighborhood is preferably formed centered on selected target.
The prediction of (3.2.2) target movement tendency
According to the characteristics of motion before target, the movement tendency of target can be predicted.Assuming that Δ x, Δ y distinguish table
Show former frame of the target from initial value frame to present frame in moving distance horizontal, in vertical direction, Δ t indicate from initial frame to
The frame number of the former frame of present frame, then target speed can simply use vx=Δ x/ Δ t, vy=Δ y/ Δ t indicates,
vx,vyUnit be pixel/frame, i.e., the mobile distance of average each frame target.The target predicted in tracking is being worked as
Position in previous frame is x=x0+ Δ x, y=y0+ Δ y, wherein x0,y0Indicate target in the position of the former frame of present frame.
, to later, fail and detect it is possible that also tracking target in target following chain by (3.2.1) front-to-back ratio
Detection target in chain matches, and thinks that target is blocked or disappears at this time, and motion prediction is taken to position present frame target position
It sets.Target position of every prediction then predicts that total degree adds 1 to it, i.e. it is believed that mesh after predicting that total degree is more than threshold value P
Mark has disappeared, at this time delete target.P referred to as prediction frequency threshold value, preferably 10, indicate to also fail to position the mesh after 10 frames
Mark, then delete the target, no longer track to it.
Newly there is target detection in (3.2.3)
By (3.2.1) front-to-back ratio to later, in addition to there is tracking target to fail in target following chain mentioned above
Match with the detection target in detection chain, corresponds to target at this time and be blocked or disappear outside, there are also another situations, i.e. mesh
Mark detection chain in exist with the unpaired target of target following chain, think at this time these detection targets be target newly occur, as
New target is inserted into target following chain tail of the queue.
After completing above-mentioned (3.2.1), (3.2.2), (3.2.3), according to the position of each target of present frame and attribute information
Update target following chain.
(4) time domain variation detection
The detection of moving target, the kinetic characteristic of target are one of key factors, and the present invention judges mesh using variation detection
Whether mark moves, and carries out time domain variation detection to target obtained above.
Nth frame extracted SURF characteristic point to reference frame and present frame as reference frame respectively in the past, and based on RANSAC
The affine Transform Model for calculating two field pictures, obtains transformation parameter matrix Matrix=Assuming that preceding nth frame image
In some pixel coordinate be (x, y), then the coordinate of the pixel is denoted as (x', y') in the current frame, hasTransformation parameter matrix description is from preceding nth frame to the background motion of present frame.
Thus, it is supposed that the preceding nth frame coordinates of targets that aforementioned airspace Object Detecting and Tracking obtains is (x, y), carrying on the back
Under the action of scape movement, the corresponding coordinates of targets of present frame is denoted as (x', y'), hasIts
In, the coordinate of target uses the center point coordinate of target rectangle frame.
The case where aforesaid operations all exist just for corresponding target in two frames.When the current frame, to the target of preceding nth frame
Coordinate obtains it in the estimation of current frame coordinate by transformation parameter matrix, then with coordinate that target is corresponded in practical present frame
The calculating of Euclidean distance is carried out, to judge target with the presence or absence of active movement.Assuming that preceding nth frame coordinates of targets is denoted as (x, y), estimate
Counting the obtained coordinates of targets that corresponds in the current frame is (x', y'),And practical present frame target
Coordinate is (x ", y ").The Euclidean distance obtained by comparingDistance threshold Th is set5, such as
Fruit S < Th5, then it is assumed that active movement is not present in the target, belongs to stationary object, is not determined as moving target;Else if
S≥Th5, then it is assumed that it is moving target, without rejecting.
Th5Preferably 3.In view of the time delay of detection, and the slowly speed of moving target, N is preferably 3.
Disclosed above is only specific example of the invention, the thought provided according to the present invention, those skilled in the art
Can think and variation, should all fall within the scope of protection of the present invention.
Claims (14)
1. a kind of aerial weak moving target detection tracking of infrared video, it is characterised in that include the following steps:
(1) airspace target detection: skyline is extracted using scene areas perception, divides region in aerial region and non-empty;In turn
Region is based on TopHat transformation and carries out conspicuousness detection in the sky, realizes the pixel detection of aerial conspicuousness Weak target;Into
And several targets are obtained using target cluster segmentation to the pixel detected, and signal-to-noise ratio confirmation is carried out to these targets, most
Final inspection measures well-marked target;Chain is tracked using these well-marked target initialized targets that present frame detects in initial frame;
(2) target following: in present frame using in well-marked target obtained above building target detection chain, with target following chain
Target carries out positive, reversed comparison, the correspondence of the target in target detection chain and the target in target following chain is established, depending on
Position target is in the position of present frame;For the target that do not establish in corresponding target following chain, measured in advance using movement tendency
The position target is in the position of present frame;For the target that do not establish in corresponding target detection chain, add as newly there is target
It is added in target following chain;
(3) time domain variation detection: time domain variation detection is carried out using present frame and preceding nth frame image, in two field pictures respectively
After extracting SURF characteristic point, affine Transform Model is calculated to estimate background motion, in compensation background motion based on Feature Points Matching
After compensated after picture frame, based on present frame, picture frame after compensation, to aforementioned obtained object judgement, whether it is moved, right
It is judged as that the target of movement forms target trajectory, realizes target following.
2. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly in (1) scene perception detailed process are as follows:
(2.1) in present frame, its pixel value is successively traversed from bottom to up to each column image, the pixel value of previous pixel is i1,
The pixel value of latter pixel is i2, work as i1-i2≥Th1And i2<Th2When, show to initially enter sky areas at this time, wherein Th1,
Th2The variation of uncarved areas type and pixel value bright-dark degree respectively, Th1Referred to as qualitative change threshold value, Th2Referred to as dark matter threshold value;
(2.2) the current pixel point coordinate for initially entering sky areas is denoted as (k, l), record persistently belongs to the picture of sky areas
Vegetarian refreshments number Count, i.e. these pixels i2With neighbor pixel i1Meet d=| i1-i2|<Th1And i2<Th2;When the condition
When being unsatisfactory for, that is, it is unsatisfactory for d=| i1-i2|<Th1And i2<Th2When, current pixel point coordinate is (k+Count-1, l) to note at this time;
This section of vertical line of from (k, l) to (k+Count-1, l), pixel and neighbor pixel meet d=| i1-i2|<Th1And i2<
Th2, belong to sky areas;
(2.3) judgement that each column of present frame are carried out with above-mentioned sky areas, obtains several segments sky areas;It traverses from bottom to up
These sky areas, by duration ranges maximum, i.e., the vertical line lower end of the maximum sky areas of Count number is as the image column
Horizon line position;
(2.4) horizon line position is detected according to the method described above to each column of present frame, then by the horizon line position of each column
The skyline of full frame image can be obtained in connection;Sky areas is used as on skyline.
3. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly the detailed process of the well-marked target pixel detection in (1) based on TopHat transformation are as follows: in the day that above-mentioned scene perception obtains
In empty region, present frame and morphological operator are subjected to opening operation, subtract becoming as a result, obtaining TopHat for opening operation with present frame
It is changing as a result, i.e. Dst=TopHat (I, Element)=I-Open (I, Element);Wherein I represents present frame, Element
Represent morphological operator.
4. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly in (1) target cluster segmentation detailed process are as follows:
(4.1) result by TopHat transformation carries out binary conversion treatment, and binarization threshold Th is arranged3, TopHat transformation result it is big
In Th3It is then set to 1, is otherwise set to 0;
(4.2) pixel after binary conversion treatment for 1 carries out neighborhood cluster, treatment process are as follows: random as suspected target point
Select a suspected target point as initial clustering seed, 4 neighborhoods or 8 neighborhoods to this seed carry out breadth first search,
Illustrate that the two pixels distance is close if it can find other suspected target points, belong to same target, it should which it is same for clustering
One target;Finally all cluster is attributed to some target to suspected target point after all binaryzations.
5. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly the detailed process that target signal to noise ratio confirms in (1) are as follows:
(5.1) Local modulus maxima of Statistical Clustering Analysis target, if its gray value is g0, by formed centered on it 3 × 3 neighborhoods and 9 ×
9 neighborhoods;
After 9 pixels in (5.2) 9 × 9 neighborhoods in 3 × 3 contiguous ranges of removal, remaining 72 pixels composition target week
The background area enclosed, gray value are denoted as gi, i=1 ... 72;
(5.3) according to Signal to Noise Ratio (SNR) formula, the signal-to-noise ratio of current goal is obtained;
(5.4) when target signal to noise ratio SNR is less than snr threshold Th4, i.e. SNR < Th4, then the target is filtered;To in present frame
Each cluster target carries out the calculating of signal-to-noise ratio, the final real goal for confirming airspace conspicuousness and detecting.
6. a kind of aerial weak moving target detection tracking of infrared video as claimed in claim 2, it is characterised in that: institute
In the transition for stating sky areas and non-sky area, qualitative change threshold value Th1Preferably 25, dark matter threshold value Th2Preferably 50.
7. the aerial weak moving target detection tracking of a kind of infrared video as described in claim 3, it is characterised in that: described
Morphological operator Element chooses size slightly larger than detection target, preferably
8. a kind of aerial weak moving target detection tracking of infrared video as claimed in claim 4, it is characterised in that: institute
When stating to TopHat transformation results progress binary conversion treatment, binarization threshold Th3Preferably 110.
9. a kind of aerial weak moving target detection tracking of infrared video as claimed in claim 5, it is characterised in that: institute
State target signal to noise ratio threshold value Th4Preferably 6, show that target of the Signal to Noise Ratio (SNR) greater than 6 is removed.
10. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly the detailed process that target detection chain and target following chain are positive in (2), reversely compare are as follows:
(10.1) positive to compare: to take out each of target detection chain detection target DObj, calculated separately in its M neighborhood away from
From detection target DObjTarget T in nearest target following chainObj, will distance DObjNearest TObjIt is considered as same target;
(10.2) reversed to compare: to take out each of target following chain tracking target TObj, in its M contiguous range, calculate mesh
Mark detection chain in tracking target TObjNearest detection target D 'Obj;
(10.3) forward and reverse result compares: judging DObjWith D 'ObjIt whether is the same target, if it is the same detection target,
Then with the target in the corresponding target following chain of the detection target update, the pursuit path of the present frame target is formed;Otherwise it presses
The pursuit path to the present frame target is formed according to the result of motion prediction.
11. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly the detailed process that target movement tendency is predicted in (2) are as follows: Δ x, before Δ y respectively indicates target from initial value frame to present frame
One frame indicates the frame number of the former frame from initial frame to present frame in moving distance horizontal, in vertical direction, Δ t, then target
Movement velocity can use vx=Δ x/ Δ t, vy=Δ y/ Δ t is indicated;The target predicted in tracking is in the current frame
Position be x=x0+ Δ x, y=y0+ Δ y, wherein x0,y0Indicate target in the position of the former frame of present frame;It is wanted by power
10. front-to-back ratios are asked to fail and the detection target phase in detection chain to tracking target later, is likely present in target following chain
Matching, thinks that target is blocked or disappears at this time, and motion prediction is taken to position present frame target position, target of every prediction
Position then predicts that total degree adds 1 to it, i.e. it is believed that target has disappeared after predicting that total degree is more than prediction frequency threshold value P
It loses, at this time delete target.
12. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly 10 × 10 neighborhood of pixels that M neighborhood is preferably formed centered on selected target in (2), prediction frequency threshold value P is preferably 10, table
Show that 10 frames also fail to position the target later, then deletes the target, no longer it is tracked.
13. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly the detailed process that time domain variation detects in (3) are as follows:
(13.1) SURF characteristic point is extracted respectively in present frame, preceding nth frame image, based on Feature Points Matching, using RANSAC
Calculate affine Transform ModelTo estimate background motion;
(13.2) picture frame after being compensated after compensating background motion is obtained based on picture frame after present frame, compensation to aforementioned
Object judgement its whether move, detailed process are as follows: assuming that preceding nth frame coordinates of targets is denoted as (x, y), estimate current
It is (x', y') that coordinates of targets is corresponded in frame,And practical present frame coordinates of targets is (x ", y "),
The Euclidean distance obtained by comparingDistance threshold Th is set5If S < Th5, then think
Active movement is not present in the target, belongs to stationary object, is not determined as moving target;If S >=Th5, then it is considered to transport
Moving-target, without rejecting.
14. a kind of aerial weak moving target detection tracking of infrared video as described in claim 1, it is characterised in that step
Suddenly nth frame before the reference frame in (3) for comparing, N is preferably 3, the Euclidean distance threshold value Th of time domain variation detection5Preferably 3.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910220147.4A CN109978851B (en) | 2019-03-22 | 2019-03-22 | Method for detecting and tracking small and medium moving target in air by using infrared video |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910220147.4A CN109978851B (en) | 2019-03-22 | 2019-03-22 | Method for detecting and tracking small and medium moving target in air by using infrared video |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109978851A true CN109978851A (en) | 2019-07-05 |
CN109978851B CN109978851B (en) | 2021-01-15 |
Family
ID=67079999
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910220147.4A Active CN109978851B (en) | 2019-03-22 | 2019-03-22 | Method for detecting and tracking small and medium moving target in air by using infrared video |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109978851B (en) |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110399823A (en) * | 2019-07-18 | 2019-11-01 | Oppo广东移动通信有限公司 | Main body tracking and device, electronic equipment, computer readable storage medium |
CN110490904A (en) * | 2019-08-12 | 2019-11-22 | 中国科学院光电技术研究所 | A kind of Dim targets detection and tracking |
CN110490903A (en) * | 2019-08-12 | 2019-11-22 | 中国科学院光电技术研究所 | Multiple target fast Acquisition and tracking in a kind of Binocular vision photogrammetry |
CN110599516A (en) * | 2019-08-22 | 2019-12-20 | 普联技术有限公司 | Moving target detection method and device, storage medium and terminal equipment |
CN110991313A (en) * | 2019-11-28 | 2020-04-10 | 华中科技大学 | Moving small target detection method and system based on background classification |
CN111353991A (en) * | 2020-03-10 | 2020-06-30 | 北京市商汤科技开发有限公司 | Target detection method and device, electronic equipment and storage medium |
CN111652151A (en) * | 2020-06-04 | 2020-09-11 | 北京环境特性研究所 | Infrared weak and small target tracking method |
CN111899512A (en) * | 2020-08-05 | 2020-11-06 | 深圳大学 | Vehicle track extraction method and system combined with skyline observation and storage medium |
CN112381076A (en) * | 2021-01-18 | 2021-02-19 | 西南石油大学 | Method for preprocessing picture in video significance detection task |
CN112686895A (en) * | 2021-03-11 | 2021-04-20 | 成都点泽智能科技有限公司 | Method and device for detecting infrared dim target and server |
CN113344963A (en) * | 2021-05-27 | 2021-09-03 | 绍兴市北大信息技术科创中心 | Seed point self-adaptive target tracking system based on image segmentation |
CN113822279A (en) * | 2021-11-22 | 2021-12-21 | 中国空气动力研究与发展中心计算空气动力研究所 | Infrared target detection method, device, equipment and medium based on multi-feature fusion |
TWI767596B (en) * | 2020-04-28 | 2022-06-11 | 大陸商深圳市商湯科技有限公司 | Scene depth and camera motion prediction method, electronic equipment and computer readable storage medium |
CN115205327A (en) * | 2022-05-18 | 2022-10-18 | 哈尔滨工业大学 | Infrared small target tracking method fusing historical library information |
CN115222775A (en) * | 2022-09-15 | 2022-10-21 | 中国科学院长春光学精密机械与物理研究所 | Weak and small target detection tracking device and detection tracking method thereof |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1570957A (en) * | 2004-05-13 | 2005-01-26 | 上海交通大学 | Method for detecting infrared weak and small object under antiaerial imaging fluctuation background |
CN1606033A (en) * | 2004-11-18 | 2005-04-13 | 上海交通大学 | Weak target detecting and tracking method in infrared image sequence |
CN101567087A (en) * | 2009-05-25 | 2009-10-28 | 北京航空航天大学 | Method for detecting and tracking small and weak target of infrared sequence image under complex sky background |
CN102074022A (en) * | 2011-01-10 | 2011-05-25 | 南京理工大学 | Infrared image-based weak and small moving target detecting method |
CN102103748A (en) * | 2010-12-14 | 2011-06-22 | 西北工业大学 | Method for detecting and tracking infrared small target in complex background |
CN104268852A (en) * | 2014-09-05 | 2015-01-07 | 西安电子科技大学 | Detection method and device for small and dim targets in infrared sequence images |
CN104299229A (en) * | 2014-09-23 | 2015-01-21 | 西安电子科技大学 | Infrared weak and small target detection method based on time-space domain background suppression |
KR20150109083A (en) * | 2014-03-19 | 2015-10-01 | 국방과학연구소 | Method of Distance Adaptive Detection Threshold Value for Infrared Target Detection System |
CN107705319A (en) * | 2017-10-16 | 2018-02-16 | 中国电子科技集团公司第二十八研究所 | One kind is based on the real-time detecting and tracking method of empty day background Small object |
CN109003277A (en) * | 2017-06-07 | 2018-12-14 | 中国航空工业集团公司洛阳电光设备研究所 | A kind of infrared small target in complex background detection method and device |
CN109446553A (en) * | 2018-09-14 | 2019-03-08 | 清华大学 | A kind of aerial Dynamic IR Scene analogue system based on Unity3D |
-
2019
- 2019-03-22 CN CN201910220147.4A patent/CN109978851B/en active Active
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1570957A (en) * | 2004-05-13 | 2005-01-26 | 上海交通大学 | Method for detecting infrared weak and small object under antiaerial imaging fluctuation background |
CN1606033A (en) * | 2004-11-18 | 2005-04-13 | 上海交通大学 | Weak target detecting and tracking method in infrared image sequence |
CN101567087A (en) * | 2009-05-25 | 2009-10-28 | 北京航空航天大学 | Method for detecting and tracking small and weak target of infrared sequence image under complex sky background |
CN102103748A (en) * | 2010-12-14 | 2011-06-22 | 西北工业大学 | Method for detecting and tracking infrared small target in complex background |
CN102074022A (en) * | 2011-01-10 | 2011-05-25 | 南京理工大学 | Infrared image-based weak and small moving target detecting method |
KR20150109083A (en) * | 2014-03-19 | 2015-10-01 | 국방과학연구소 | Method of Distance Adaptive Detection Threshold Value for Infrared Target Detection System |
CN104268852A (en) * | 2014-09-05 | 2015-01-07 | 西安电子科技大学 | Detection method and device for small and dim targets in infrared sequence images |
CN104299229A (en) * | 2014-09-23 | 2015-01-21 | 西安电子科技大学 | Infrared weak and small target detection method based on time-space domain background suppression |
CN109003277A (en) * | 2017-06-07 | 2018-12-14 | 中国航空工业集团公司洛阳电光设备研究所 | A kind of infrared small target in complex background detection method and device |
CN107705319A (en) * | 2017-10-16 | 2018-02-16 | 中国电子科技集团公司第二十八研究所 | One kind is based on the real-time detecting and tracking method of empty day background Small object |
CN109446553A (en) * | 2018-09-14 | 2019-03-08 | 清华大学 | A kind of aerial Dynamic IR Scene analogue system based on Unity3D |
Non-Patent Citations (4)
Title |
---|
LIJUN XIE ET AL.: "An Effective TBD Algorithm for the Detection of Infrared Dim-Small Moving Target in the Sky Scene", 《COMMUNICATIONS IN COMPUTER AND INFORMATION SCIENCE》 * |
XIANGZHI BAI ET AL.: "Enhancement of dim small top-hat transformation under the condition of heavy clutter", 《SIGNAL PROCESSING》 * |
XIAOLIANG SUN ET AL.: "Real-time visual enhancement for infrared small dim targets in video", 《INFRARED PHYSICS & TECHNOLOGY》 * |
郑锦 等: "基于三阶中心矩区域分类的视频运动目标检测", 《计算机辅助设计与图形学学报》 * |
Cited By (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110399823B (en) * | 2019-07-18 | 2021-07-09 | Oppo广东移动通信有限公司 | Subject tracking method and apparatus, electronic device, and computer-readable storage medium |
CN110399823A (en) * | 2019-07-18 | 2019-11-01 | Oppo广东移动通信有限公司 | Main body tracking and device, electronic equipment, computer readable storage medium |
CN110490904A (en) * | 2019-08-12 | 2019-11-22 | 中国科学院光电技术研究所 | A kind of Dim targets detection and tracking |
CN110490903A (en) * | 2019-08-12 | 2019-11-22 | 中国科学院光电技术研究所 | Multiple target fast Acquisition and tracking in a kind of Binocular vision photogrammetry |
CN110490904B (en) * | 2019-08-12 | 2022-11-11 | 中国科学院光电技术研究所 | Weak and small target detection and tracking method |
CN110490903B (en) * | 2019-08-12 | 2022-11-11 | 中国科学院光电技术研究所 | Multi-target rapid capturing and tracking method in binocular vision measurement |
CN110599516A (en) * | 2019-08-22 | 2019-12-20 | 普联技术有限公司 | Moving target detection method and device, storage medium and terminal equipment |
CN110991313A (en) * | 2019-11-28 | 2020-04-10 | 华中科技大学 | Moving small target detection method and system based on background classification |
CN110991313B (en) * | 2019-11-28 | 2022-02-15 | 华中科技大学 | Moving small target detection method and system based on background classification |
CN111353991A (en) * | 2020-03-10 | 2020-06-30 | 北京市商汤科技开发有限公司 | Target detection method and device, electronic equipment and storage medium |
TWI767596B (en) * | 2020-04-28 | 2022-06-11 | 大陸商深圳市商湯科技有限公司 | Scene depth and camera motion prediction method, electronic equipment and computer readable storage medium |
CN111652151A (en) * | 2020-06-04 | 2020-09-11 | 北京环境特性研究所 | Infrared weak and small target tracking method |
CN111652151B (en) * | 2020-06-04 | 2023-12-19 | 北京环境特性研究所 | Infrared dim target tracking method |
CN111899512A (en) * | 2020-08-05 | 2020-11-06 | 深圳大学 | Vehicle track extraction method and system combined with skyline observation and storage medium |
CN112381076A (en) * | 2021-01-18 | 2021-02-19 | 西南石油大学 | Method for preprocessing picture in video significance detection task |
CN112686895B (en) * | 2021-03-11 | 2021-05-18 | 成都点泽智能科技有限公司 | Method and device for detecting infrared dim target and server |
CN112686895A (en) * | 2021-03-11 | 2021-04-20 | 成都点泽智能科技有限公司 | Method and device for detecting infrared dim target and server |
CN113344963A (en) * | 2021-05-27 | 2021-09-03 | 绍兴市北大信息技术科创中心 | Seed point self-adaptive target tracking system based on image segmentation |
CN113822279A (en) * | 2021-11-22 | 2021-12-21 | 中国空气动力研究与发展中心计算空气动力研究所 | Infrared target detection method, device, equipment and medium based on multi-feature fusion |
CN113822279B (en) * | 2021-11-22 | 2022-02-11 | 中国空气动力研究与发展中心计算空气动力研究所 | Infrared target detection method, device, equipment and medium based on multi-feature fusion |
CN115205327A (en) * | 2022-05-18 | 2022-10-18 | 哈尔滨工业大学 | Infrared small target tracking method fusing historical library information |
CN115222775A (en) * | 2022-09-15 | 2022-10-21 | 中国科学院长春光学精密机械与物理研究所 | Weak and small target detection tracking device and detection tracking method thereof |
Also Published As
Publication number | Publication date |
---|---|
CN109978851B (en) | 2021-01-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109978851A (en) | A kind of aerial weak moving target detection tracking of infrared video | |
CN107993245B (en) | Aerospace background multi-target detection and tracking method | |
CN104217428B (en) | A kind of fusion feature matching and the video monitoring multi-object tracking method of data correlation | |
CN108038415B (en) | Unmanned aerial vehicle automatic detection and tracking method based on machine vision | |
US6298143B1 (en) | Moving target detecting system and moving target detecting method | |
CN108537829B (en) | Monitoring video personnel state identification method | |
CN103077539A (en) | Moving object tracking method under complicated background and sheltering condition | |
CN104766079B (en) | A kind of remote method for detecting infrared puniness target | |
CN110400294B (en) | Infrared target detection system and detection method | |
Lian et al. | A novel method on moving-objects detection based on background subtraction and three frames differencing | |
CN110070560B (en) | Object motion direction identification method based on target detection | |
Jiang et al. | Multiple pedestrian tracking using colour and motion models | |
CN111311644A (en) | Moving target detection method based on video SAR | |
Dehghani et al. | Small target detection and tracking based on the background elimination and Kalman filter | |
Liu et al. | Space target extraction and detection for wide-field surveillance | |
Jiang et al. | An infrared small target detection algorithm based on peak aggregation and Gaussian discrimination | |
Najafzadeh et al. | Object tracking using Kalman filter with adaptive sampled histogram | |
Kim et al. | Three plot correlation-based small infrared target detection in dense sun-glint environment for infrared search and track | |
Zhang et al. | Study on moving-objects detection technique in video surveillance system | |
Li et al. | Sort with depth image based pedestrian tracking robots | |
Chen et al. | Summary about detection and tracking of infrared small targets | |
Liu et al. | Shadow Elimination in Traffic Video Segmentation. | |
Wu et al. | Robust object tracking with bidirectional corner matching and trajectory smoothness algorithm | |
Faramarzi et al. | Infrared dim and small targets detection based on multi-scale local contrast measure utilizing efficient spatial filters | |
Micheloni et al. | Fast good features selection for wide area monitoring |
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 |