CN106815562A - A kind of pedestrian detection tracking based on compressive features - Google Patents
A kind of pedestrian detection tracking based on compressive features Download PDFInfo
- Publication number
- CN106815562A CN106815562A CN201611203542.4A CN201611203542A CN106815562A CN 106815562 A CN106815562 A CN 106815562A CN 201611203542 A CN201611203542 A CN 201611203542A CN 106815562 A CN106815562 A CN 106815562A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- positive
- sample
- negative sample
- target
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
- G06F18/24155—Bayesian classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
Abstract
The invention provides the pedestrian detection tracking based on compressive features, its Gentle Adaboost algorithm for passing through to introduce misclassification cost-sensitive mechanism is trained to the positive/negative sample that training sample is concentrated, and ties up the random sparse Toeplitz matrixes Φ of spacing using m × nToepTracking and dimension-reduction treatment are compressed to the positive/negative sample of the tracking target as projection matrix, and the positive/negative sample characteristics compressed after dimensionality reduction are input into Bayes classifier and are trained jointly with the position weight of positive/negative sample, to generate the discriminant function for distinguishing pedestrian target and background, in next frame input picture, selection makes discriminant function take the sample of maximum as new tracking target, to realize the detecting and tracking to pedestrian target.By the present invention, improve exist complex jamming factor in public domain to the accuracy of pedestrian detection, enhance the robustness of algorithm, be especially suitable for having with pedestrian head and block or other abnormal disturbed conditions are detected and tracked to pedestrian.
Description
Technical field
The invention belongs to Computer Image Processing field, more particularly to a kind of pedestrian detection track side based on compressive features
Method.
Background technology
The accurate tracking of target location is the key problem of computer vision field.In man-machine interaction, video monitoring, enhancing
It is widely used in reality.In the understanding of task such as scene, the action recognition of higher level, target following also plays important
Role.
In the prior art, the pedestrian detection to public domain generally needs the positive sample and negative sample collection of positive sample concentration
In negative sample be trained to obtain pedestrian's grader, and by pedestrian's grader to Real-time Collection to image in pedestrian
Whether region decision is that pedestrian area is gone forward side by side pedestrian's counting number, so as to realize the detection to pedestrian.
Current most of track algorithms can regard the target detection problems of each frame as, i.e., based on tracking-by-
Problem under detection frameworks, the track algorithm under the framework is very sensitive to the accurate description of display model, once occur with
Track is drifted about, and the tracing positional of mistake is necessarily resulted in inaccurate display model, is difficult to give for change again based on inaccurate display model
Tracking target.Display model is inaccurate cause target following mistake when, if tracing positional can be corrected in time, what is tracked is accurate
Degree can be greatly promoted, and influence of the tracking drift to track algorithm also can accordingly decline.
Detected based on outward appearance and target can be predicted based on the track algorithm of prediction from different perspectives to a certain extent
Position, the new target tracking for proposing a kind of combination detection and prediction is can contemplate based on this thinking.
Traditional Gentle Adaboost algorithms are by optimal to realize by wrong point of sample weights in adjustment often wheel training
The selection of pedestrian's grader, the positive/negative sample for wrong point, its adjustment ratio is identical, therefore, cannot solve in the prior art
Lack of balance problem between certainly positive/negative sample size.In view of this, it is necessary to pedestrian's inspection in the prior art to public domain
Survey tracking to be improved, to solve above-mentioned technology flaw.Meanwhile, pedestrian detection method of the prior art in pedestrian deliberately
Head or shoulder are set during in improper form, then missing inspection can occur when extracting pedestrian dummy and carrying out detecting and tracking,
Therefore there is also certain defect.This defect there is in having occasion higher to Statistics Bar people number or safety check inspection
Larger potential safety hazard.
The content of the invention
It is an object of the invention to disclose a kind of pedestrian detection tracking based on compressive features,
To achieve the above object, the invention provides a kind of pedestrian detection tracking based on compressive features, including with
Lower step:
S1, the HOG characteristic vectors for calculating the positive/negative sample that training sample is concentrated, by introducing misclassification cost-sensitive machine
The Gentle Adaboost algorithms of system are trained to the positive/negative sample that training sample is concentrated, and obtain pedestrian's grader;
S2, the video streaming image of acquisition monitor area are made using pedestrian's grader as input picture to input picture
Pedestrian detection, to obtain pedestrian area;
S3, using the pedestrian area that detects as the tracking target of present frame input picture, sampled in target domain with
The positive/negative sample of track target, and the multiple dimensioned higher-dimension Haar-like characteristic vectors of the positive/negative sample of tracking target are extracted, and adopt
The random sparse Toeplitz matrixes Φ of spacing is tieed up with m × nToepThe positive/negative sample of the tracking target is entered as projection matrix
Row compression tracking and dimension-reduction treatment;
S4, by compress dimensionality reduction after positive/negative sample characteristics be input into Bayes's classification jointly with the position weight of positive/negative sample
It is trained in device, to generate the discriminant function for distinguishing pedestrian target and background, and uses double sigmoidal curve adaptive learning
Update the parameter of Bayes classifier;
S5, in next frame input picture, selection makes discriminant function take the sample of maximum as new tracking target, with
Realize the detecting and tracking to pedestrian target.
As a further improvement on the present invention, in the step S1, the positive/negative sample that training sample is concentrated for 32 × 64~
256 rank gray level images of 64 × 128 pixels.
As a further improvement on the present invention, the number of the positive sample that positive sample is concentrated is 4000 in the step S1, is born
The number of negative sample is 6000 in sample set.
As a further improvement on the present invention, the misclassification cost-sensitive mechanism in the step S1 is specially:
Calculate lack of balance cost loss function value;
Calculate initialization pedestrian's grader classification error rate, and calculate positive/negative sample initialize pedestrian's grader in
By the higher limit of mistake classification;
Current optimal Weak Classifier f (x) is chosen to initialization pedestrian's grader amendment, so as to obtain pedestrian's grader.
As a further improvement on the present invention, in the step S1, the calculating for calculating lack of balance cost loss function value is public
Formula is:
Wherein, C1It is that positive sample divides cost loss function value, C by mistake2It is that negative sample divides cost loss function value, y by mistakei
=1 represents sample xiActual is positive sample, yi=-1 represents sample xiActual is negative sample, H (xi)=sign (∑ fj(x)) it is strong
The classification results of grader, H (xi)=- 1 represents that sample is sentenced x by initialization pedestrian's graderiWei not negative sample, H (xiThe table of)=1
Show that sample is sentenced x by initialization pedestrian's graderiWei not positive sample, C1, C2∈ [0,1], and C1> C2;
In the step S1, the computing formula for calculating the classification error rate of initialization pedestrian's grader is:
Wherein, NFNBe positive sample by mistake classification quantity, NFPBe negative sample by mistake classification quantity;
In the step S1, positive/negative sample is calculated in pedestrian's grader is initialized by the meter of the higher limit of mistake classification
Calculating formula is:
Wherein, F (x) is cumulative model, and its expression formula is:
In the step S1, the computing formula of current optimal Weak Classifier f (x) is:
Wherein, Pw(y=1 | x) and Pw(y=-1 | x) represents the weight cumulative distribution of positive/negative sample respectively.
As a further improvement on the present invention, " every trade is entered to input picture using pedestrian's grader in the step S2
People detects " it is specially:Input picture is scanned using window, calculate fall in window altimetric image to be checked HOG features to
Amount, and the testing result of the altimetric image HOG characteristic vectors to be checked is input into pedestrian's grader, detect most probable row
People region;The specification of the window is 32 × 64~64 × 128 pixels.
As a further improvement on the present invention, the m × n in the step S3 ties up the random sparse Toeplitz matrixes of spacing
ΦToepIt is expressed as:
φI+1, j+1=φI, j, and the random sparse Toeplitz matrixes Φ of spacing is tieed up to m × nToepThe 1st row and the 1st row unit
Vectorial V={ the φ that element is constituted1, φ2... φN φN+1... φN+M-1Make the sparse change of random spacing, make element φiValue
Independent identically distributed random Gaussian distribution is obeyed, other elements are all entered as 0,
Wherein i ∈ κ, the κ are randomly selected from the index sequence of 1~N+M-1Individual index, institute
State the spacing that Δ is element.
As a further improvement on the present invention, " the positive/negative sample to the tracking target is compressed in the step S3
Track and dimension-reduction treatment " it is specially:
First, using the pedestrian area that detects as the tracking target of present frame input picture, in the neighborhood of tracking target
The positive/negative sample of interior sampling tracking target;
Then, the multiple dimensioned higher-dimension Haar-like characteristic vectors B of the positive/negative sample of tracking target is extractedk∈Rn×1;
Then, the random sparse Toeplitz matrixes Φ of spacing is tieed up using the m × nToepEvery a line it is positive/negative with k-th
The primitive character B of samplekIt is multiplied, obtains the ith feature f after k-th positive/negative sample compressioni;
Finally, m combinations of features is obtained into k-th low-dimensional feature F of positive/negative sample into m dimensional vectorsk, and Fk∈Rm×1。
As a further improvement on the present invention, the step S4 also includes:Using Mean shift algorithms prediction pedestrian area
The position of pedestrian target in domain in next frame input picture, and measure the positive/negative sample of pedestrian target and be input into present frame
Positional distance of the position and prediction pedestrian target of image between next frame input picture, to calculate pedestrian target in next frame
Position weight between the predicted position of the pedestrian target in input picture and the positive/negative sample position of pedestrian target.
Compared with prior art, the beneficial effects of the invention are as follows:By the present invention, improve and there is complex jamming factor
In public domain to the accuracy of pedestrian detection, enhance the robustness of algorithm, be especially suitable for pedestrian head have hide
Gear or other abnormal disturbed conditions are detected and tracked to pedestrian, and solve traditional Gentle Adaboost algorithms
The lack of balance between positive/negative sample size cannot be solved the problems, such as, making pedestrian's grader of final acquisition has more preferable classification performance.
Brief description of the drawings
Fig. 1 is a kind of schematic flow sheet of the pedestrian detection tracking based on compressive features of the present invention;
Fig. 2 is the operation principle schematic diagram of the video streaming image of the acquisition monitor area shown in Fig. 1;
Fig. 3 is the learning rate figure of double sigmoid function more young leaves Bayesian classifiers shown in the present invention.
Specific embodiment
The present invention is described in detail for shown each implementation method below in conjunction with the accompanying drawings, but it should explanation, these
Implementation method not limitation of the present invention, those of ordinary skill in the art are according to these implementation method institutes works energy, side
Equivalent transformation or replacement in method or structure, belong within protection scope of the present invention.
In this manual,Pedestrian detection is trackedWithPedestrian detectionOrThe acquisition of pedestrian areaHave in video detection
Equivalent technologies implication.Please join a kind of tool of Fig. 1 to pedestrian detection tracking of the present invention based on compressive features illustrated in fig. 3
Body implementation method.
In the present embodiment, the pedestrian detection tracking of compressive features should be based on, following steps were specifically included.
Step S1, the positive/negative sample HOG characteristic vectors for calculating training sample concentration, by introducing misclassification cost-sensitive
The positive/negative sample that the Gentle Adaboost algorithms training sample of mechanism is concentrated, obtains pedestrian's grader.
The gradient that HOG (Histogram of Oriented Gradient) characteristic vector passes through statistical picture regional area
Directional information characterizes the image local area profile.Because HOG characteristic vectors can keep geometry and optical conversion consistency,
Detection in especially suitable load environment to pedestrian, so the present invention characterizes each training sample set using HOG characteristic vectors.
In the present embodiment, training sample set is by the positive sample collection comprising some positive samples and comprising some negative samples
Negative sample collection is constituted.Due to pedestrian, head is smaller with the change of shoulder when walking, based on be easy to context of detection reliability with
, can be defined as training sample set by the consideration of accuracy:The positive sample collection of wardrobe portion and/or shoulder is only included, and is not included
The negative sample of pedestrian head and/or shoulder.
Specifically, positive/negative sample is 256 rank gray level images of 32 × 64~64 × 128 pixels in the positive/negative sample set.
Processing speed to image is carried out based on computer and efficiency considers, in the present embodiment, can be by positive/negative sample set
Positive/negative sample be preferably 256 rank gray level images of 32 × 64 pixels.Specifically, the positive sample in pedestrian's grader is initialized
The number of the positive sample of this concentration is 4000, and the number of the negative sample that negative sample is concentrated is 6000.
Traditional Gentle Adaboost algorithms often take turns the positive/negative sample weights divided by mistake in training by adjustment come real
The selection of existing optimum classifier.Positive/negative sample for wrong point, its adjustment ratio is identical, therefore, it is impossible to solve positive/negative
Lack of balance problem between sample size.
In the present embodiment, by introducing the misclassification cost-sensitive mechanism of positive/negative sample, traditional Gentle is solved
Adaboost algorithm cannot solve the problems, such as the lack of balance between positive/negative sample size." by introducing misclassification cost in step S1
The positive/negative sample that the Gentle Adaboost algorithms training sample of sensitive mechanism is concentrated, obtains initializing pedestrian's grader." tool
Body implementation process is as follows.
First, the cost loss function value of lack of balance is calculated, the computing formula of the lack of balance cost loss function value is:
Wherein, C1It is that positive sample divides cost loss function value, C by mistake2It is that negative sample divides cost loss function value, y by mistakei
=1 represents sample xiActual is positive sample, yi=-1 represents sample xiActual is negative sample, H (xi)=sign (∑ fj(x)) it is strong
The classification results of grader, H (xi)=- 1 represents that sample is sentenced x by initialization pedestrian's graderiWei not negative sample, H (xiThe table of)=1
Show that sample is sentenced x by initialization pedestrian's graderiWei not positive sample, C1, C2∈ [0,1], and C1> C2。
Then, the classification error rate of initialization pedestrian's grader is calculated, and calculates positive/negative sample in initialization pedestrian's classification
By the higher limit of mistake classification in device.
Specifically, the computing formula of the classification error rate of initialization pedestrian's grader is:
Wherein, NFNBe positive sample by mistake classification quantity, NFPBe negative sample by mistake classification quantity.
In the present embodiment, represent what positive/negative sample was classified in initialization pedestrian's grader mistake using exponential form
Higher limit, its computing formula is:
Wherein, F (x) is cumulative model, and its expression formula is
That is, the cumulative sum of the optimal Weak Classifier of each round.Then introduce the Gentle of misclassification cost-sensitive mechanism
The computing formula of a wheel loss function minimum value is in Adaboost algorithm model:
Wherein, I () is target function.
When the loss function of a new round is calculated, the value of model F (x) that adds up needs to be added up on the basis of last round of one and works as
Preceding optimal Weak Classifier f (x), computing formula is as follows:
Wherein,Ew[] is expected for weighting, weights desired computing formula and is:Ew[g (x, y)]
=E [w (x, y) g (x, y)]/E [w (x, y)].
Finally, choose current optimal Weak Classifier f (x) to be modified initialization pedestrian's grader, so as to obtain pedestrian
Grader.
Specifically, in the present embodiment, current optimal Weak Classifier f (x) specific formula for calculation is as follows:
Wherein, Pw(y=1 | x) and Pw(y=-1 | x) represents the weight cumulative distribution of positive/negative sample respectively, often takes turns iteration
Weight transfer formula is:
So far, the Gentle Adaboost algorithm amendments that misclassification cost-sensitive mechanism is introduced described in the present embodiment are complete
Finish, and finally obtain the pedestrian's grader for subsequently carrying out pedestrian detection to input picture.
Step S2, the video streaming image of acquisition monitor area are schemed using pedestrian's grader as input picture to input
As making pedestrian detection, to obtain pedestrian area.
Specifically, shown in ginseng Fig. 2, a kind of pedestrian detection tracking based on compressive features of the present invention is based on video camera
It is vertical to shoot and suitable for outdoor situations and indoor situations.In the present embodiment, step S2 is specially:By video camera 10
The video streaming image of monitor area 30 is obtained as input picture, the monitor area 30 is located at the underface of video camera 10.
Specifically, video camera 10 is arranged on the surface near gateway 20, pedestrian can along on the direction of arrow 201
Walked up and down in gateway 20.Monitor area 30 acquired in video camera 10 can be completely covered the Zone Full of gateway 20.
In the present embodiment, the monitor area 30 is rectangle, naturally it is also possible to be square or circular or other shapes
Shape.Video camera 10 is located at the surface of the central point 301 of monitor area 30, and thus we can derive, the monitor area 30
Positioned at the underface of video camera 10.
Use the window (unit that specification is 32 × 64~64 × 128:Pixel) to obtain input picture be scanned, count
Calculate the HOG characteristic vectors of the altimetric image to be checked in window, and the testing result of the altimetric image HOG characteristic vectors to be checked is defeated
Enter into pedestrian's grader, to detect most probable pedestrian area.This can be for the window being scanned to input picture
Rectangle, or oval or square.The top or oblique upper of monitor area 30 are arranged on due to video camera, therefore should
Pedestrian area generally rectangular or ellipse.
Step S3, using the pedestrian area that detects as the tracking target of present frame input picture, adopted in target domain
Sample tracks the positive/negative sample of target, and extracts the multiple dimensioned higher-dimension Haar-like characteristic vectors of the positive/negative sample of tracking target,
And using the sparse Toeplitz matrixes Φ of random spacing of m × n dimensionsToepAs projection matrix to it is described tracking target just/
Negative sample is compressed tracking and dimension-reduction treatment.
Compression tracking (Compressive Tracking, CT) algorithm is the fast robust proposed based on compressive sensing theory
Track algorithm.The algorithm using one meet limited equidistant (RIP, restricted isometry property) property with
Machine projection matrix E uses matrix product computing F=EX, E ∈ Rm×n(m < < n), by original dimensional signal X ∈ R highn×1It is compressed into low
Dimensional signal F ∈ Rm×1。
Wherein, parameter m, n represent the line number and columns of accidental projection matrix E respectively.
For the extraction of compressive features, the selection of accidental projection matrix E is extremely important, and different accidental projection matrix E lead
Different Haar-like characteristic vectors are caused, produced display model is also different.In the present embodiment, tieed up using m × n
The sparse Toeplitz matrixes Φ of random spacingToepFor signal reconstruction, the matrix is obtained in signal reconstruction preferably rebuilds effect
Really, and physically easily realization, carrying cost are low.The present invention uses it for tracking the compression expression of target signature, to obtain
Obtain more preferable tracking effect and efficiency.The random sparse Toeplitz matrixes Φ of spacingToepMathematic(al) representation be:
φ in matrixI+1, j+1=φI, j, the vectorial V={ φ constituted to its 1st row and the 1st column element1, φ2... φN
φN+1... φN+M-1The sparse change of random spacing is done, make element φi(i ∈ κ, κ are random from the index sequence of 1~N+M-1
ChooseIndividual index, Δ is the size of spacing) value obey the distribution of independent identically distributed random Gaussian, other yuan
It is plain to be all entered as 0.
The calculating of Haar-like characteristic vectors is simple, and has good recognition effect to the relatively-stationary target of structure.Knot
The calculating for closing the Haar-like characteristic vectors of integrogram is even more efficient, and this is extremely important for real-time modeling method.
Tracking target of the present invention using the pedestrian rectangular area that detects as present frame, tracking of being sampled in target neighborhood
The positive/negative sample of target, extracts the multiple dimensioned higher-dimension Haar-like characteristic vectors B of positive/negative samplek∈Rn×1, then using m × n
The sparse Toeplitz matrixes Φ of random spacing of dimensionToepEvery a line and k-th positive/negative sample primitive character BkIt is multiplied, obtains
Ith feature f after k-th positive/negative sample compressioni, m combinations of features is finally obtained into k-th positive/negative sample into m dimensional vectors
This low-dimensional feature Fk, Fk∈Rm×1。
Step S4, by compress dimensionality reduction after positive/negative sample characteristics be input into Bayes jointly with the position weight of positive/negative sample
It is trained in grader, to generate the discriminant function for distinguishing pedestrian target and background, and uses double sigmoidal curve self adaptation
Study updates the parameter of Bayes classifier.
When background characteristics is similar with target signature, this is unfavorable for that the classification of classification function is realized.According to closer to target
Actual position, detection algorithm is detected as bigger this thought of probability of target, in classification function add target actual position with
The influence of sample distance weighting, can increase the discrimination between target and sample, so as to improve the confidence level of classification.Target is true
Real position is in advance ignorant, if it is possible to predict target actual position well, so that it may substituted with the position of prediction true
Position.
The present invention proposes a simply and intuitively idea, using Mean shift algorithm future positions, surveys one by one
Amount sample position and future position distance, calculate predicted position and sample position distance weighting (sample position weight), so
Sample position weight input Bayes classifier is classified afterwards.
Assuming that lpIt is the target location of Mean Shift algorithms prediction, definitionFor
Positive/negative sample k range-to-gos, wherein,WithRepresent that the x coordinate and y of k-th positive/negative sample position are sat respectively
Mark.
In order to reduce considerable influence of the range noise to weight, by 1/dkCarry out tanh normalization, then normalize after just/
The position weight ω of negative sample kkComputing formula be:
In above formula, 1/d is chosenkThe reason for be:Positive/negative sample from predicted position more away from, positive/negative sample is detected as target
Probability it is smaller, positive/negative sample position weights omegakAlso should be smaller, vice versa.μGHAnd σGHIt is respectively 1/dkAverage and standard
Difference.
The computing formula of the discriminant function of combining target predicted position is as follows:
In above formula,WhereinWithPositive sample is represented respectively
The average and standard deviation of this ith feature,WithThe average and standard deviation of the ith feature of negative sample are represented respectively.
Y ∈ { 0,1 } are binary variables, and its value represents negative sample label and positive sample label respectively.
Assuming that prior probability is equal, i.e. p (y=1)=p (y=0).Current region is closer to target, ωkValue it is bigger.Again
Because closer to target, the probability that detection algorithm is detected as target is bigger, therefore p (fi| y=1) value also mutually strain it is big, and
(1-ωk) and p (fi| y=0) then diminish, therefore discriminant function M (f) can further expand the difference of background and target.Pass through
The normalized weights omegas of tanhkWhen=0.5, weight is to p (fi| y=1) and p (fi| y=0) influence consistent, weights omegakMore than 0.5
Then to p (fi| y=1) it is favourable, otherwise to p (fi| y=0) it is favourable, and weights omegakBigger influence is more obvious, therefore can good area
Partial objectives for and background.
Inventing discriminant function M (f) for proposing can combine sample position weight, preferably amplify the difference of background and target
It is different, enhancing classification confidence level.The present invention chooses target location of the maximum positive sample position of M (f) value as a new frame, and to phase
Related parameter is updated to adapt to the renewal of target and background.Shown in the following depicted of parameter proposed by the present invention more new model:
Wherein:
In above formula, variable x reflects current goal mean μ (or variances sigma2) with existing target mean μi(or variances sigmai 2)
Extent of deviation, its value shows that more greatly current goal is bigger with existing target departure degree, namely result is more insincere (because regarding
Change can not possibly be too big between two frames in frequency sequence), therefore to the μ of existing targeti(or) learning rate should be bigger.Therefore,
Only need appropriate arrange parameter t, r1And r2Good adaptive learning effect can be obtained.
Preferably, in the present embodiment, arrange parameter t=0.3, r1=0.3 and r2=0.2.
Parameter of the invention more new model can be according to the difference between the parameter of current goal and existing target component, automatically
The size of learning rate is adjusted, so as to have more adaptivity, the robustness of algorithm is enhanced.
Please join illustrated in fig. 3 pair of sigmoid function learning rate figure.
As seen from Figure 3, (passed through with the average of existing target or variance difference degree in the average or variance of current goal
Normalized), during value x=0.3, α=0.5, this shows that the speed for learning existing target with learning current goal is equal.Work as x
Value when being less than 0.3, the now average or variance of current goal and the average of existing target or variance difference degree (normalization)
Less than 0.3, the speed of the existing target of study should be less than 0.5 (see x=0.3 left-hand components);Conversely, when the value of x is more than 0.3,
Learning rate should be greater than 0.5;And due to r2< r1The curve ratio first half of latter half is more precipitous, namely is learned in x > 0.3
The change for practising speed α is rapider, and this is conducive to the influence for preventing noise or classification error to parameter.
Step S5, in next frame input picture, selection discriminant function is taken the sample of maximum as new tracking mesh
Mark, to realize the detecting and tracking to pedestrian target.
This step S5 is the circulation implementation procedure to step S4, implements process ginseng specifically explaining above for step S4
State, will not be repeated here.
Those listed above is a series of to be described in detail only for feasibility implementation method of the invention specifically
Bright, they simultaneously are not used to limit the scope of the invention, all equivalent implementations made without departing from skill spirit of the present invention
Or change should be included within the scope of the present invention.
It is obvious to a person skilled in the art that the invention is not restricted to the details of above-mentioned one exemplary embodiment, Er Qie
In the case of without departing substantially from spirit or essential attributes of the invention, the present invention can be in other specific forms realized.Therefore, no matter
From the point of view of which point, embodiment all should be regarded as exemplary, and be nonrestrictive, the scope of the present invention is by appended power
Profit requires to be limited rather than described above, it is intended that all in the implication and scope of the equivalency of claim by falling
Change is included in the present invention.Any reference in claim should not be considered as the claim involved by limitation.
Moreover, it will be appreciated that although the present specification is described in terms of embodiments, not each implementation method is only wrapped
Containing an independent technical scheme, this narrating mode of specification is only that for clarity, those skilled in the art should
Specification an as entirety, the technical scheme in each embodiment can also be formed into those skilled in the art through appropriately combined
May be appreciated other embodiment.
Claims (9)
1. a kind of pedestrian detection tracking based on compressive features, it is characterised in that comprise the following steps:
S1, the HOG characteristic vectors for calculating the positive/negative sample that training sample is concentrated, by introducing misclassification cost-sensitive mechanism
Gentle Adaboost algorithms are trained to the positive/negative sample that training sample is concentrated, and obtain pedestrian's grader;
S2, the video streaming image of acquisition monitor area make pedestrian using pedestrian's grader as input picture to input picture
Detection, to obtain pedestrian area;
S3, using the pedestrian area that detects as the tracking target of present frame input picture, tracking mesh of being sampled in target domain
The positive/negative sample of target, and the multiple dimensioned higher-dimension Haar-like characteristic vectors of the positive/negative sample of tracking target are extracted, and use m
× n ties up the random sparse Toeplitz matrixes Φ of spacingToepThe positive/negative sample of the tracking target is pressed as projection matrix
Contracting tracking and dimension-reduction treatment;
S4, by compress dimensionality reduction after positive/negative sample characteristics be input into Bayes classifier jointly with the position weight of positive/negative sample
It is trained, to generate the discriminant function for distinguishing pedestrian target and background, and is updated using double sigmoidal curve adaptive learning
The parameter of Bayes classifier;
S5, in next frame input picture, selection makes discriminant function take the sample of maximum as new tracking target, to realize
To the detecting and tracking of pedestrian target.
2. pedestrian detection tracking according to claim 1, it is characterised in that in the step S1, training sample set
In positive/negative sample for 32 × 64~64 × 128 pixels 256 rank gray level images.
3. pedestrian detection tracking according to claim 2, it is characterised in that positive sample is concentrated in the step S1
The number of positive sample is 4000, and it is 6000 that negative sample concentrates the number of negative sample.
4. pedestrian detection tracking according to claim 1, it is characterised in that the misclassification cost in the step S1
Sensitive mechanism is specially:
Calculate lack of balance cost loss function value;
The classification error rate of initialization pedestrian's grader is calculated, and it is wrong in pedestrian's grader is initialized to calculate positive/negative sample
The higher limit of misclassification;
Choose current optimal Weak Classifier f (x) to be modified initialization pedestrian's grader, so as to obtain pedestrian's grader.
5. pedestrian detection tracking according to claim 4, it is characterised in that in the step S1, calculates lack of balance
The computing formula of cost loss function value is:
Wherein, C1It is that positive sample divides cost loss function value, C by mistake2It is that negative sample divides cost loss function value, y by mistakei=1 table
Sample this xiActual is positive sample, yi=-1 represents sample xiActual is negative sample, H (xi)=sign (Σ fj(x)) it is strong classification
The classification results of device, H (xi)=- 1 represents that sample is sentenced x by initialization pedestrian's graderiWei not negative sample, H (xi)=1 represents just
Sample is sentenced x by beginningization pedestrian's graderiWei not positive sample, C1, C2∈ [0,1], and C1> C2;
In the step S1, the computing formula for calculating the classification error rate of initialization pedestrian's grader is:
Wherein, NFNBe positive sample by mistake classification quantity, NFPBe negative sample by mistake classification quantity;
In the step S1, positive/negative sample is calculated public by the calculating of the higher limit of mistake classification in pedestrian's grader is initialized
Formula is:
Wherein, F (x) is cumulative model, and its expression formula is:
In the step S1, the computing formula of current optimal Weak Classifier f (x) is:
Wherein, Pw(y=1 | x) and Pw(y=-1 | x) represents the weight cumulative distribution of positive/negative sample respectively.
6. pedestrian detection tracking according to claim 1, it is characterised in that " utilize pedestrian in the step S2
Grader carries out pedestrian detection to input picture " it is specially:Input picture is scanned using window, calculates in window
The HOG characteristic vectors of altimetric image to be checked, and the testing result of the altimetric image HOG characteristic vectors to be checked is input into pedestrian's classification
In device, most probable pedestrian area is detected;The specification of the window is 32 × 64~64 × 128 pixels.
7. pedestrian detection tracking according to claim 1, it is characterised in that the m × n dimensions in the step S3 are random
The sparse Toeplitz matrixes Φ of spacingToepIt is expressed as:
φI+1, j+1=φI, j, and the random sparse Toeplitz matrixes Φ of spacing is tieed up to m × nToepThe 1st row and the 1st column element institute
Vectorial V={ the φ of composition1, φ2... φN φN+1... φN+M-1Make the sparse change of random spacing, make element φiValue obey
Independent identically distributed random Gaussian distribution, other elements are all entered as 0,
Wherein i ∈ κ, the κ are randomly selected from the index sequence of 1~N+M-1Individual index, the Δ is
The spacing of element.
8. pedestrian detection tracking according to claim 7, it is characterised in that " to the tracking in the step S3
The positive/negative sample of target is compressed tracking and dimension-reduction treatment " it is specially:
First, adopted in the neighborhood of tracking target as the tracking target of present frame input picture using the pedestrian area that detects
Sample tracks the positive/negative sample of target;
Then, the multiple dimensioned higher-dimension Haar-like characteristic vectors B of the positive/negative sample of tracking target is extractedk∈Rn×1;
Then, the random sparse Toeplitz matrixes Φ of spacing is tieed up using the m × nToepEvery a line and k-th positive/negative sample
Primitive character BkIt is multiplied, obtains the ith feature f after k-th positive/negative sample compressioni;
Finally, m combinations of features is obtained into k-th low-dimensional feature F of positive/negative sample into m dimensional vectorsk, and Fk∈Rm×1。
9. pedestrian detection tracking according to claim 1, it is characterised in that the step S4 also includes:Using
Position of the pedestrian target in next frame input picture in Mean shift algorithms prediction pedestrian area, and measure pedestrian target
Positive/negative sample the position of present frame input picture and position of the prediction pedestrian target between next frame input picture away from
From to calculate the predicted position of pedestrian target of the pedestrian target in next frame input picture and the positive/negative sample of pedestrian target
Position weight between position.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611203542.4A CN106815562A (en) | 2016-12-19 | 2016-12-19 | A kind of pedestrian detection tracking based on compressive features |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611203542.4A CN106815562A (en) | 2016-12-19 | 2016-12-19 | A kind of pedestrian detection tracking based on compressive features |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106815562A true CN106815562A (en) | 2017-06-09 |
Family
ID=59109652
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611203542.4A Pending CN106815562A (en) | 2016-12-19 | 2016-12-19 | A kind of pedestrian detection tracking based on compressive features |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106815562A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109102519A (en) * | 2018-08-24 | 2018-12-28 | 周忠和 | Quick person tracking method under moving-target and feature templates change condition |
CN109859242A (en) * | 2019-01-16 | 2019-06-07 | 重庆邮电大学 | A kind of method for tracking target for predicting adaptive learning |
CN110287957A (en) * | 2019-06-24 | 2019-09-27 | 北京中电慧视科技有限公司 | A kind of localization method and positioning device of low slow Small object |
CN112434884A (en) * | 2020-12-12 | 2021-03-02 | 广东电力信息科技有限公司 | Method and device for establishing supplier classified portrait |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103632382A (en) * | 2013-12-19 | 2014-03-12 | 中国矿业大学(北京) | Compressive sensing-based real-time multi-scale target tracking method |
CN105426928A (en) * | 2014-09-19 | 2016-03-23 | 无锡慧眼电子科技有限公司 | Pedestrian detection method based on Haar characteristic and EOH characteristic |
-
2016
- 2016-12-19 CN CN201611203542.4A patent/CN106815562A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103632382A (en) * | 2013-12-19 | 2014-03-12 | 中国矿业大学(北京) | Compressive sensing-based real-time multi-scale target tracking method |
CN105426928A (en) * | 2014-09-19 | 2016-03-23 | 无锡慧眼电子科技有限公司 | Pedestrian detection method based on Haar characteristic and EOH characteristic |
Non-Patent Citations (1)
Title |
---|
钟宝康: "基于压缩感知的预测目标跟踪算法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109102519A (en) * | 2018-08-24 | 2018-12-28 | 周忠和 | Quick person tracking method under moving-target and feature templates change condition |
CN109102519B (en) * | 2018-08-24 | 2021-12-14 | 周忠和 | Quick person tracking method under moving target and characteristic template change conditions |
CN109859242A (en) * | 2019-01-16 | 2019-06-07 | 重庆邮电大学 | A kind of method for tracking target for predicting adaptive learning |
CN109859242B (en) * | 2019-01-16 | 2022-09-09 | 重庆邮电大学 | Target tracking method for prediction adaptive learning |
CN110287957A (en) * | 2019-06-24 | 2019-09-27 | 北京中电慧视科技有限公司 | A kind of localization method and positioning device of low slow Small object |
CN110287957B (en) * | 2019-06-24 | 2021-09-07 | 北京中电慧视科技有限公司 | Low-slow small target positioning method and positioning device |
CN112434884A (en) * | 2020-12-12 | 2021-03-02 | 广东电力信息科技有限公司 | Method and device for establishing supplier classified portrait |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108830252B (en) | Convolutional neural network human body action recognition method fusing global space-time characteristics | |
Tran et al. | Anomaly detection using a convolutional winner-take-all autoencoder | |
CN108416250B (en) | People counting method and device | |
Kwak et al. | Learning occlusion with likelihoods for visual tracking | |
Zhang et al. | Video anomaly detection and localization using motion-field shape description and homogeneity testing | |
CN111191667B (en) | Crowd counting method based on multiscale generation countermeasure network | |
CN109740419A (en) | A kind of video behavior recognition methods based on Attention-LSTM network | |
CN104680559B (en) | The indoor pedestrian tracting method of various visual angles based on motor behavior pattern | |
CN106874894A (en) | A kind of human body target detection method based on the full convolutional neural networks in region | |
CN106529419B (en) | The object automatic testing method of saliency stacking-type polymerization | |
CN102214309B (en) | Special human body recognition method based on head and shoulder model | |
CN110569772A (en) | Method for detecting state of personnel in swimming pool | |
CN110378208B (en) | Behavior identification method based on deep residual error network | |
CN103914854B (en) | A kind of image sequence target association and orbit generation method | |
CN106815562A (en) | A kind of pedestrian detection tracking based on compressive features | |
Chen et al. | LSTM with bio inspired algorithm for action recognition in sports videos | |
CN106156765A (en) | safety detection method based on computer vision | |
CN107833239B (en) | Optimization matching target tracking method based on weighting model constraint | |
CN110991274B (en) | Pedestrian tumbling detection method based on Gaussian mixture model and neural network | |
Huang et al. | Learning multimodal deep representations for crowd anomaly event detection | |
CN110633643A (en) | Abnormal behavior detection method and system for smart community | |
Luo et al. | Traffic analytics with low-frame-rate videos | |
CN103902989A (en) | Human body motion video recognition method based on non-negative matrix factorization | |
CN105809206A (en) | Pedestrian tracking method | |
Zhou et al. | Research on fast pedestrian detection algorithm based on autoencoding neural network and adaboost |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170609 |