CN107066968A - The vehicle-mounted pedestrian detection method of convergence strategy based on target recognition and tracking - Google Patents
The vehicle-mounted pedestrian detection method of convergence strategy based on target recognition and tracking Download PDFInfo
- Publication number
- CN107066968A CN107066968A CN201710236834.6A CN201710236834A CN107066968A CN 107066968 A CN107066968 A CN 107066968A CN 201710236834 A CN201710236834 A CN 201710236834A CN 107066968 A CN107066968 A CN 107066968A
- Authority
- CN
- China
- Prior art keywords
- mtd
- mrow
- msub
- target
- mtr
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
-
- 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
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a kind of vehicle-mounted pedestrian detection method of the convergence strategy based on target recognition and tracking, it is related to computer vision field, including：1 detects moving region using the background estimating method of the positive and negative subtractive of three frames, then is partitioned into moving target by Otsu, Canny edge detection operator and Morphological scale-space method；2 carry out motion estimate using the cascade classifier of fusion feature and combination Gentle Adaboost and SVM, judge whether target belongs to pedestrian；3 carry out pedestrian tracking using based on Kalman filter method；4 carry out the pedestrian detection of the convergence strategy based on identification and tracking, and use different inspection policies to the target of different conditions.The problem of moving object detection is disturbed by light in onboard system is the method overcome, convergence strategy avoids unnecessary computing, so as to shorten system operation time and improve discrimination, with good application value.
Description
Technical field
The present invention relates to computer vision field, a kind of vehicle-mounted row of the convergence strategy based on target recognition and tracking is refered in particular to
People's detection method.
Background technology
Pedestrian detection is an important component in automobile active safety technology, with road vehicle detection, trade line
Detection and detection of obstacles together form the major part that collision in driver assistance system avoids early warning system.Pedestrian tracking
There is very important meaning in intelligent transportation system with detection technique, it avoids driver by the warning to driver
Collision is produced with the pedestrian near it.In order to protect the pedestrian in road traffic, it is necessary to accurately detect and estimate du vehicule
The information such as particular location, the direction of motion and the movement velocity of pedestrian.Occur it is dangerous in the case of to driver give a warning with
Collision free pedestrian, so as to effectively raise the safety and reliability of Traffic Systems.
Pedestrian detection in onboard system includes moving target recognition and target identification, moving target recognition mainly for
Obtain there may be pedestrian target region, target identification is then to classify to candidate target, and whether judge it is pedestrian.Its
In, the algorithm of moving target recognition have the method for feature based, the method based on optical flow field, the method based on inter-frame difference, with
And based on background subtraction method.These methods all assume that video camera is actionless, then goes out target according to feature detection.
In practical application, especially in the video image of night infrared shooting, all there is the place of shortcoming, feature based in these methods
Method be vulnerable to the influence of the factors such as light and weather, it is necessary to be dynamically determined threshold value, the requirement handled in real time can not be met,
Neighbor frame difference point-score can be influenceed by light and weather.Traditional feature based and pedestrian's recognizer instruction of cascade classifier
The white silk time is long, and verification and measurement ratio is relatively low, the shortcomings of easily producing adaptation.Further, the posture of pedestrian is varied, between pedestrian
Appearance is even more far from each other；Outside illumination condition is changed greatly；When pedestrian is distant, pedestrian's shared pixel ratio in the picture
Less, it is low to add image resolution ratio, it is difficult to recognize；Pedestrian is during drawing near, and its size variation is very big；Pedestrian examines
Examining system also faces the challenge of real-time and robustness while accuracy is constantly pursued, and each of which increases in vehicular platform
The difficulty of pedestrian detection.
The content of the invention
The technical problem to be solved in the present invention is to propose that a kind of run time is short and discrimination is high based on target identification
With the vehicle-mounted pedestrian detection method of the convergence strategy of tracking.
In order to solve the above technical problems, the present invention is adopted the technical scheme that：A kind of melting based on object tracking and recognition
The vehicle-mounted pedestrian detection method of strategy is closed, the vehicle-mounted pedestrian detection method specifically includes following steps：S1 is utilizing adjacent three frame just
The method for estimating that minus tolerance subtracts builds initial background to detect moving region, then passes through Otsu, Canny edge detection operator
Moving target is partitioned into Morphological scale-space method；It includes：
S1.1 moving object detections；
The segmentation of S1.2 motion target areas；
S2 carries out motion estimate using the cascade classifier of fusion feature and combination Gentle Adaboost and SVM,
Classification and Identification is carried out to the moving target being partitioned into, judges whether the moving target belongs to pedestrian；It includes：
S2.1 combination Haar-like features and HOG features are to enhance the expressive force of feature set；
S2.2 real values Weak Classifier is designed, and includes the Weak Classifier design and the weak typing of HOG features of Haar-like features
Device is designed；
The design of S2.3 combination Gentle Adaboost and SVM cascade classifier；
S3 carries out pedestrian tracking using based on Kalman filter method；It includes：
S3.1 goal descriptions, are modeled to dbjective state and feature；
S3.2 predicts the motion state of target using Kalman filter；
Data correlation and template renewal that S3.3 is combined based on nearest neighbor method and template matches；
S4 carries out the pedestrian detection of the convergence strategy based on identification and tracking, and uses different to the target of different conditions
Inspection policies.
As the further optimization of technical solution of the present invention, S1.1 moving object detections described in this method are included according to initial
Background image, moving object detection is carried out using background subtraction method, during vehicle is moved, using the adjacent three frames figure
As the method for estimating of positive and negative subtractive dynamicallys update background image, the background image come out using current time is examined
Survey next frame moving target.
As the further optimization of technical solution of the present invention, the segmentation of S1.2 motion target areas described in this method is specifically wrapped
Include following steps：
Step1：According to difference of the following equation threshold method to background subtraction method in detection process of moving target in S1.1
Image is split：
In formula, dt(x, y), to need the image for carrying out extracting moving target, T is the target and background that Otsu algorithms are tried to achieve
Between inter-class variance it is maximum when threshold value；
Step2：Edge is detected using Canny operators；
Step3：By the expansion fillet in morphology, the edge that connecting detection goes out finally is partitioned into motion
Target.
As the further optimization of technical solution of the present invention, the Weak Classifier of Haar-like features is as follows described in this method
Shown in formula：
Wherein, fi(x) value, θ are characterizediFor the threshold value of the Weak Classifier of Haar-like features, α and β are interventions [- 1,1]
Between real number, the confidence level of presentation class result, negative value is expressed as non-pedestrian target, on the occasion of for pedestrian target；
The Weak Classifier of the HOG features includes each gradient orientation histogram for positive sample, utilizes sample
Weighted average m represents pedestrian target, shown in equation below：
Wherein, hiValue is characterized, n is the number of positive sample, ωiFor the weights of each sample；
The Weak Classifier of HOG features calculates each histogram to weighted average m Euclidean distance to be classified：
In above formula, as the Weak Classifier of Haar-like features, the confidence level of α and β difference presentation class results, just
Value is represented as pedestrian, and negative value is represented as non-pedestrian, d (hj(x) m) it is, Euclidean distance of the HOG characteristic values to weighted average m,
θjFor threshold value.
As the further optimization of technical solution of the present invention, S2.3 combinations Gentle Adaboost and SVM described in this method
The design of cascade classifier specifically include following steps：
S2.3.1 sets the maximum misclassification rate F of every layer of strong classifiermax∈ (0.0.5] and minimum detection rate Dmin∈[0.99,
1)；
When S2.3.2 carries out the training of Gentle Adaboost cascade classifiers using the feature set in S2.1, every layer is set
Maximum characteristic MAX, herein in maximum characteristic MAX, strong classifier, Gentle are trained using Gentle Adaboost algorithms
Adaboost training algorithm processes：
1) training sample (x is given1,y1),…,(xN,yN), set positive sample number P=1500, negative sample number K=
2500 for positive sample, makes yi=1；For negative sample, y is madei=1；
2) sample weights are initialized
In formula, P is positive sample number, and K is negative sample number；
3) for r=1 ..., R, R represents algorithm iteration number of times, that is, constitutes the Weak Classifier number of strong classifier；For every
One feature u, obtains Weak Classifier gu∈[-1,1]；Calculate Weak Classifier guClassification errorSelect optimal gt；Strong classifier result G (x) is updated to G (x)+gt(x)；Sample is weighed
Value is updated to ωi exp[yigt(xi)], wherein i is number of samples, i=1,2 ..., N；
4) strong classifier of output is：
S2.3.3 is more than MAX when the characteristic of S2.3.2 cascade graders layer, selects the MAX feature as SVM
The input of grader, it is hereafter, several layers of after cascade classifier that strong classifier training, the form of SVM classifier are all carried out with SVM algorithm
For：
In above formula, x is the characteristic vector of an observed sample, and y ∈ { -1,1 } are class label, xiIt is i-th of training
The characteristic vector of sample, N is the quantity of training sample, K (x, xi) it is a kernel function, α={ α1,α2,...,αNIt is to solve
Value, drawn by the quadratic programming problem for solving following：
Wherein,
In formula, C is a predefined parameter, is penalty factor, therefore can be converted into the problem of linearly inseparable soft
The linear separability problem at interval；xi,xjFor supporting vector, kernel function K (xi,xj) it is RBF；
K(xi,xj)=exp (- γ | | xi,xj||2),γ>0
In formula, xi,xjVector is characterized, γ is the parameter of kernel function；
When cascade classifier meets misclassification rate less than or equal to FmaxIt is more than or equal to minimum detection rate D with verification and measurement ratiominWhen,
Terminate the training of cascade classifier.
As the further optimization of technical solution of the present invention, in the S3.1, R={ L, F, x', y', h', w, △ are defined
X', △ y', △ h', △ w'} represent moving target, wherein, L is the label of moving target, and F represents the macroscopic features of target,
(x', y') is the center of target, and h', w' is respectively the height and width of target boundary rectangle, and △ x', △ y' represent target respectively
Center represents the pace of change of boundary rectangle height and width in x directions and y directions up conversion speed, △ h', △ w'；
In the S3.2, the motion state of target is predicted using Kalman filter, for quantity of state (x', y', h', w',
△ x', △ y', △ h', △ w'), Kalman filter equation can be expressed as：
Xk+1=AXk+Wk
Zk=H'Xk+Vk
In above formula, A is state-transition matrix, and H' is observing matrix, Wk、VkFor dynamic noise and observation noise, ZkIt is observation
Vector, state vector XkIt is an octuple vector：Xk=[x, ' y', h', w', △ x', △ y', △ h', △ w']T；
Assuming that pedestrian moves with uniform velocity in unit interval, state-transition matrix is represented by：
Take observation vector Zk=[x', y', h', w']T, selecting system observing matrix is：
Target state is predicted according to the Kalman filter equation and Kalman filter algorithm of structure,
The original state of the target of Kalman filter is the result that step S1 and step S2 are identified；
In the S3, shown in arest neighbors method equation below：
Dis(R1,R2)=| | x '1-x′1||2+||y′1-y′2||2+||h′1-h′2||2+||w′1-w′2||2
In above formula, R1、R2Prediction target and observed object, (x ' are represented respectively1,y′1) it is the center for predicting target,
(x′2,y′2) be observed object center, h '1、w′1To predict the height and width of target boundary rectangle, h '2、w′2For observation mesh
Mark the height and width of boundary rectangle；
If the distance between prediction target and observation for associating are considered as less than given threshold, the data correlation
Effectively, and with observed object now the template of respective objects is updated；If predicted between target and the observation associated
Distance be more than threshold value, then the association results be considered as it is invalid, now using template matches method prediction target it is attached
The new observed object of search near field；If template matches success, then the result of template matches is used as effective observation；If
It is the observation that template matches do not find respective objects, then the target may disappear or be blocked, so as to terminate tracking.
As the further optimization of technical solution of the present invention, this method utilizes Kalman multi-object tracking methods, will recognized
It is merged with tracking, different detection plans is used by multiframe recognition result comprehensive descision, and to the target of different conditions
Slightly, it includes：
1) for pedestrian target, the knowledge before pedestrian target appearance in 50 two field pictures every 10 testing property of frame
Not, only pedestrian target is tracked afterwards, without carrying out cascade classifier identification；
2) for non-pedestrian target, once recognized every 5 frames；
3) for the uncertain target of state, all it is identified per frame, and judge whether it is pedestrian.
Compared with prior art, the invention has the advantages that：
The vehicle-mounted pedestrian detection method of the convergence strategy based on target recognition and tracking of the present invention, overcomes onboard system
The problem of middle moving object detection is disturbed by light, improves based on AdaBoost pedestrian's recognizer, while will tracking and knowledge
Other convergence strategy is applied in vehicle-mounted pedestrian detecting system, it is to avoid unnecessary computing.
This method has good robustness and real-time, so that shorten system operation time and improve discrimination,
Therefore this method has good application value.
Brief description of the drawings
Fig. 1 is total algorithm flow chart described in the present embodiment；
Fig. 2 is moving target recognition flow chart described in the present embodiment；
Fig. 3 is the expression schematic diagram of rectangle in the present embodiment described image subwindow；
Fig. 4 is the Haar-like rectangular characteristic figures used in the present invention described in the present embodiment；
Fig. 5 is pedestrian's identification process figure described in the present embodiment.
Embodiment
The present invention is described in further details with reference to accompanying drawing, present embodiment provide it is a kind of based on target identification and
The vehicle-mounted pedestrian detection method of the convergence strategy of tracking, the vehicle-mounted pedestrian detection method are as shown in figure 1, mainly include following step
Suddenly：
S1 builds initial background to detect moving region using the method for estimating of the adjacent positive and negative subtractive of three two field pictures,
Moving target is partitioned into by Otsu, Canny edge detection operator and Morphological scale-space method again；
S1.1 moving object detections；
The image background gathered for vehicular platform is larger by the change of illumination condition and the interference of shade, using adjacent three
The method for estimating of the positive and negative subtractive of two field picture builds initial background to detect moving target.Assuming that three two field pictures being continuously shot
It is set to It-1(x,y)、It(x,y)、It+1(x, y), is divided into two groups of I by this three frames continuous videos imaget-1(x,y)、It(x,y)
And It(x,y)、It+1(x, y), positive and negative difference fortune is carried out according to formula (1) and (2) respectively by corresponding pixel in two groups of images
Calculate, and the positive minus tolerance absolute value of first group of two images is stored inIn, second group of two images
Positive minus tolerance absolute value is stored inIn.
Assuming that the background image and foreground image of t frames are respectivelyPreset a threshold
Value T0, for determining foreground image.If corresponding for a certain pixel (x, y)Gray value is both greater than T0, then can be determined that the pixel (x, y) is being worked as
All it is motion in preceding continuous three two field picture, therefore the pixel (x, y) can be classified as foreground imageWith reference to sentencing
The foreground image madeAll target pixel points are removed according to formula (3) in t two field pictures, staying part is
The background image extracted in t frames
Aforesaid operations are repeated to the n two field pictures of beginning, and the background image obtained every time is counted, you can one is obtained
Secondary complete initial background image.
According to initial background image, moving object detection is carried out using background subtraction method, during vehicle is moved, is adopted
Background image is dynamicallyd update with the method for estimating of the above-mentioned adjacent positive and negative subtractive of three two field pictures, is counted using current time
The background image come detects next frame moving target.
The segmentation of S1.2 motion target areas.
Detect after moving target, it is necessary to split to extract motion target area to image, to comprise the following steps that：
Step1：According to formula (4) difference diagram of the threshold method to background subtraction method in detection process of moving target in S1.1
As being split：
In formula, dt(x, y) for needs extract the image of moving target；T is the target and background that Otsu algorithms are tried to achieve
Between inter-class variance it is maximum when threshold value.
Step2：Edge is detected with Canny operators；
Step3：By the expansion fillet in morphology, the edge that connecting detection goes out finally is partitioned into motion
Target.
Moving target recognition flow is as shown in Figure 2.
S2 carries out motion estimate using the cascade classifier of fusion feature and combination Gentle Adaboost and SVM,
To the step 1) in the moving target that is partitioned into carry out Classification and Identification, judge whether the moving target belongs to pedestrian；
S2.1 combination Haar-like features and HOG features are comprised the following steps that with enhancing the expressive force of feature set：
S2.1.1 extracts Haar-like features, and Harr-like features are calculated using integrogram；
As shown in figure 3, any subwindow in image, size is that W × H, W and H represent long and wide respectively, any son
Any rectangular area of window can be represented by a four-tuple：R=(p, q, w, h), wherein (p, q) represents a left side for rectangle
Upper angular vertex coordinate, w and h are respectively the expression length and width of rectangle.The value of four-tuple is met：0≤x,x+w≤W；0≤y,y+h≤
H；x,y≥0,w,h≥0.
The calculation formula of Haar-like features is represented by：
Wherein, wiFor the weights of i-th of matrix, Re cSum (ri) be expressed as i-th of matrix all pixels gray value
It is composition feature with, NlThe number of the rectangle of feature.The Haar-like rectangular characteristics used in present embodiment are such as
Shown in Fig. 4.
Image is carried out in intensive scanning process, the pixel and meter of rectangular area will be calculated when calculating characteristic value every time
Calculation amount can be quite big, influences the real-time performance of system, therefore present embodiment calculates Haar-like spies using integral image
Levy.
S2.1.2 extracts HOG features.
The inside that Haar-like features are depended in target image, have ignored image detail part, and to target
Shape and illumination condition conversion ratio it is more sensitive.Therefore for the expressive force of Enhanced feature collection, introducing can be to local object appearance
The HOG features characterized very well with shape improve Detection results as the performance compensation of Haar-like features.Conventional HOG
Feature calculation method will carry out many read group totals, the problem of being computed repeatedly in order to avoid feature, present embodiment profit
HOG features are calculated with integration histogram, it is concretely comprised the following steps：
1) to the carry out gray processing and Gaussian smoothing of input picture；
2) gradient magnitude and gradient direction of each pixel in the gray level image after the calculating gray processing；
3) the image double integral of different directions is carried out to gradient image；
4) gradient image is projected into integration direction histogram of gradients.
S2.2 real values Weak Classifier is designed, and for different characteristic Design real value graders, it specifically includes following steps：
(1) the Weak Classifier design of Haar-like features；
In present embodiment shown in the form of Weak Classifier such as formula (6)：
Wherein, fi(x) it is characterized value；θiFor the threshold value of the Weak Classifier of Haar-like features；α and β are interventions [- 1,1]
Between real number, the confidence level of presentation class result, negative value is expressed as non-pedestrian target, on the occasion of for pedestrian target.
(2) the Weak Classifier design of HOG features.
For each gradient orientation histogram of positive sample, pedestrian target is represented using the weighted average m of sample：
Wherein, hiValue is characterized, n is the number of positive sample, ωiFor the weights of each sample.The weak typing of HOG features
Device calculates each histogram to weighted average m Euclidean distance to be classified：
It is the same with the Weak Classifier of Haar-like features in formula, the confidence level of α and β difference presentation class results, on the occasion of
Pedestrian is represented as, negative value is represented as non-pedestrian；d(hj(x) m) it is, Euclidean distance of the HOG characteristic values to weighted average m, θj
For threshold value.
The design of S2.3 combination Gentle Adaboost and SVM cascade classifier.
Cascade classifier is that different graders are trained for same training set, then these different grader collection
Altogether, a stronger final grader is constituted.Present embodiment is classified in Gentle Adaboost cascade structures
SVM classifier is introduced on device, i.e., is constituted in the last basic unit of cascade structure using the select feature of Adaboost algorithm
Characteristic vector is as the input of SVM classifier, and obtained strong classifier robustness more preferably, and reduces the training time, it is to avoid
Caused cross of multiple features was used to adapt to；It specifically includes following steps：
S2.3.1 sets the maximum misclassification rate F of every layer of strong classifiermax∈ (0.0.5] and minimum detection rate Dmin∈[0.99,
1)；
When S2.3.2 carries out the training of Gentle Adaboost cascade classifiers using the feature set in S2.1, every layer is set
Maximum characteristic MAX, herein in maximum characteristic MAX, strong classifier, Gentle are trained using Gentle Adaboost algorithms
Adaboost training algorithm processes：
1) training sample (x is given1,y1),…,(xN,yN), set positive sample number P=1500, negative sample number K=
2500 for positive sample, makes yi=1；For negative sample, y is madei=1；
2) sample weights are initialized
In formula, P is positive sample number, and K is negative sample number；
3) for r=1 ..., R, R represents algorithm iteration number of times, that is, constitutes the Weak Classifier number of strong classifier；It is right
In each feature u, Weak Classifier g is obtainedu∈[-1,1]；Calculate Weak Classifier guClassification errorSelect optimal gt；Strong classifier result G (x) is updated to G (x)+gt(x)；By sample
Right value update is ωi exp[yigt(xi)], wherein i is number of samples, i=1,2 ..., N.
4) strong classifier of output is：
The output of wherein Weak Classifier contains classification confidence, and the power of the Weak Classifier can be directly represented with confidence level
Value, final classification device is a sign function.
S2.3.3 takes when the characteristic of S2.3.2 cascade graders layer is more than MAX in MAX, present embodiment
100, the MAX feature is selected as the input of SVM classifier, it is hereafter, several layers of after cascade classifier all to be carried out by force with SVM algorithm
Classifier training, it is to avoid Gentle Adaboost graders need excessive feature to cause adaptation.SVM classifier
Form be：
In formula, x is the characteristic vector of an observed sample, and y ∈ { -1,1 } are class label, xiIt is i-th of training sample
This characteristic vector, N is the quantity of training sample, K (x, xi) it is a kernel function, α={ α1,α2,...,αNTo solve
Value, is drawn by the quadratic programming problem for solving following：
Wherein,
In formula, C is a predefined parameter, be it is this between a big interval and the classification error of peanut,
Namely penalty factor, therefore the linear separability problem of soft margin can be converted into the problem of linearly inseparable；xi,xjTo support
Vector.The kernel function of formula (12) is RBF：
K(xi,xj)=exp (- γ | | xi,xj||2),γ>0 (13)
In formula, xi,xjVector is characterized, γ is the parameter of kernel function.
Before training SVM classifier, C=0.15 in parameter C and γ, present embodiment is first determined, γ=
0.06.When cascade classifier meets misclassification rate less than or equal to FmaxIt is more than or equal to minimum detection rate D with verification and measurement ratiominWhen, knot
The training of beam cascade classifier.Utilize step 2) in cascade classifier to step 1) in moving target be identified, judgement is
No is pedestrian.The dbjective state of judgement has 3 classes：Pedestrian, non-pedestrian is not known.
Fig. 5 is pedestrian's identification process figure.
S3 pedestrian trackings：Pedestrian tracking is carried out using based on Kalman filter method；
Region is likely to occur in subsequent time using discrete Kalman filter prediction pedestrian, shortens the target search time, can
Realize that pedestrian quickly positions.Pedestrian tracking result can not only obtain pedestrian movement track, can also be provided for row human motion analysis
Reliable data source.
S3.1 goal descriptions；
Goal description is that dbjective state and feature are modeled, and is the basis of multiple target tracking.Present embodiment
In target with R=, { L, F, x', y', h', w, △ x', △ y', △ h', △ w'} are represented.Wherein, L is the label of moving target；
F represents the macroscopic features of target；(x', y') is the center of target；H', w' are respectively the height and width of target boundary rectangle；
△ x', △ y' represent target's center position in x directions and y directions up conversion speed respectively；△ h', △ w' represent that boundary rectangle is high
With wide pace of change.State (x', y', h', w', △ x', △ y', △ h', △ w') of the target in a later frame can be by
Kalman filter is predicted.
S3.2Kalman filter forecastings；
Moving target can be with uniform motion come approximate representation, and present embodiment predicts mesh using Kalman filter
Target motion state.For quantity of state (x', y', h', w', △ x', △ y', △ h', △ w'), Kalman filter equation can be with table
It is shown as：
Xk+1=AXk+Wk (14)
Zk=H'Xk+Vk (15)
In formula, A is state-transition matrix；H' is observing matrix；Wk、VkFor dynamic noise and observation noise；ZkBe observation to
Amount；State vector XkIt is an octuple vector：
Xk=[x, ' y', h', w', △ x', △ y', △ h', △ w']T (17)
Because pedestrian movement's speed is slower, the time interval between adjacent two field pictures is shorter, it will be assumed that pedestrian is in unit
Moved with uniform velocity in time interval, state-transition matrix is represented by：
Take observation vector Zk=[x', y', h', w']T, selecting system observing matrix is：
The forecast model and Kalman filter algorithm built according to formula (14)-(19) carries out pre- to target state
Survey, the original state of the target of Kalman filter is the result that step S1 and step S2 are identified.
Data correlation and model modification that S3.3 is combined based on nearest neighbor method and template matches.
Data correlation be processing prediction target and observation data between corresponding relation process, it be directly connected to target with
The accuracy of track.The method that present embodiment is combined using nearest neighbor method and template matches carries out data correlation, specifically
Process is as follows：
Data correlation is carried out using the arest neighbors method of formula (20),
Dis(R1,R2)=| | x '1-x′1||2+||y′1-y′2||2+||h′1-h′2||2+||w′1-w′2||2 (20)
In formula, R1、R2Prediction target and observed object are represented respectively；(x′1,y′1) it is the center for predicting target；
(x′2,y′2) be observed object center；h′1、w′1To predict the height and width of target boundary rectangle；h′2、w′2For observation mesh
Mark the height and width of boundary rectangle.
If the distance between prediction target and observation for associating are considered as less than given threshold, the data correlation
Effectively, and with observed object now the template of respective objects is updated；If predicted between target and the observation associated
Distance be more than threshold value, then the association results be considered as it is invalid, now using template matches method prediction target it is attached
The new observed object of search near field.If template matches success, then the result of template matches is used as effective observation；If
It is the observation that template matches do not find respective objects, then the target may disappear or be blocked, so as to terminate tracking.
S4 carries out the pedestrian detection of the convergence strategy based on identification and tracking, and uses different to the target of different conditions
Inspection policies.
In order to further reduce the amount of calculation in pedestrian's identification process, while pedestrian's recognition performance in onboard system is improved,
The present invention makes full use of Kalman multi-object tracking methods, and identification and tracking are merged, comprehensive by multiframe recognition result
Close and judge, and different inspection policies are used to the target of different conditions.
According to step 2) in target identification method obtain dbjective state and can be divided into 3 classes：Pedestrian, non-pedestrian does not know three
The state of kind.For pedestrian target, the identification before its appearance in 50 two field pictures every 10 testing property of frame prevents misrecognition
Phenomenon；Only it is tracked afterwards, without carrying out cascade classifier identification；For non-pedestrian target, one is carried out every 5 frames
Secondary identification, it is to avoid pedestrian is remotely by the phenomenon of leakage identification；For the uncertain target of state, all it is identified per frame, and
Whether judge it is pedestrian.
The inventive method can efficiently identify out pedestrian target, obtain pedestrian movement track, carried for row human motion analysis
For reliable data source.So as to give a warning to driver in the case where there is danger, it is effectively prevented from colliding pedestrian, this
It is effectively improved the safety and reliability of Traffic Systems.Car is reduced using the convergence strategy of target identification and identification
The amount of calculation of pedestrian detecting system is carried, requirement of real-time can be met.
The method proposed in the present invention can actually be embedded in FPGA realizations, phase of the exploitation with real time target recognitio and tracking
Machine or video camera.Above example only plays a part of explaining technical solution of the present invention, and protection domain of the presently claimed invention is simultaneously
It is not limited to realize system and specific implementation step described in above-described embodiment.Therefore, only to specific public in above-described embodiment
Formula and algorithm are simply replaced, but its substantive content still technical scheme consistent with the method for the invention, all should be belonged to
Protection scope of the present invention.
Claims (7)
1. a kind of vehicle-mounted pedestrian detection method of the convergence strategy based on target recognition and tracking, it is characterized in that, including following step
Suddenly,
S1 builds initial background to detect moving region using the method for estimating of the adjacent positive and negative subtractive of three frames, then passes through
Otsu, Canny edge detection operator and Morphological scale-space method are partitioned into moving target；It includes：
S1.1 moving object detections；
The segmentation of S1.2 motion target areas；
S2 carries out motion estimate using the cascade classifier of fusion feature and combination Gentle Adaboost and SVM, to institute
State the moving target being partitioned into and carry out Classification and Identification, judge whether the moving target belongs to pedestrian；It includes：
S2.1 combination Haar-like features and HOG features are to enhance the expressive force of feature set；
S2.2 real values Weak Classifier is designed, including the Weak Classifier design of Haar-like features and the Weak Classifier of HOG features are set
Meter；
The design of S2.3 combination Gentle Adaboost and SVM cascade classifier；
S3 carries out pedestrian tracking using based on Kalman filter method；It includes：
S3.1 goal descriptions, are modeled to dbjective state and feature；
S3.2 predicts the motion state of target using Kalman filter；
Data correlation and template renewal that S3.3 is combined based on nearest neighbor method and template matches；
S4 carries out the pedestrian detection of the convergence strategy based on identification and tracking, and uses different detections to the target of different conditions
Strategy.
2. the vehicle-mounted pedestrian detection method of the convergence strategy as claimed in claim 1 based on target recognition and tracking, its feature
It is, the S1.1 moving object detections include, according to initial background image, moving object detection being carried out using background subtraction method,
During vehicle is moved, Background is dynamicallyd update using the method for estimating of the adjacent positive and negative subtractive of three two field pictures
Picture, the background image come out using current time detects next frame moving target.
3. the vehicle-mounted pedestrian detection method of the convergence strategy as claimed in claim 2 based on target recognition and tracking, its feature
It is, the segmentation of the S1.2 motion target areas specifically includes following steps：
Step1：According to difference image of the following equation threshold method to background subtraction method in detection process of moving target in S1.1
Split：
<mrow>
<msub>
<mi>d</mi>
<mi>t</mi>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mrow>
<msub>
<mi>d</mi>
<mi>t</mi>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo>&GreaterEqual;</mo>
<mi>T</mi>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mrow>
<msub>
<mi>d</mi>
<mi>t</mi>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo><</mo>
<mi>T</mi>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
In formula, dt(x, y), to need the image for carrying out extracting moving target, T is class between the target and background that Otsu algorithms are tried to achieve
Between variance it is maximum when threshold value；
Step2：Edge is detected using Canny operators；
Step3：By the expansion fillet in morphology, the edge that connecting detection goes out finally is partitioned into the mesh of motion
Mark.
4. the vehicle-mounted pedestrian detection method of the convergence strategy as claimed in claim 3 based on target recognition and tracking, its feature
It is, shown in the Weak Classifier equation below of the Haar-like features：
<mrow>
<msub>
<mi>g</mi>
<mn>1</mn>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mi>&alpha;</mi>
</mtd>
<mtd>
<mrow>
<msub>
<mi>f</mi>
<mi>i</mi>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo><</mo>
<msub>
<mi>&theta;</mi>
<mi>i</mi>
</msub>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mi>&beta;</mi>
</mtd>
<mtd>
<mrow>
<mi>o</mi>
<mi>t</mi>
<mi>h</mi>
<mi>e</mi>
<mi>r</mi>
<mi>w</mi>
<mi>i</mi>
<mi>s</mi>
<mi>e</mi>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
Wherein, fi(x) value, θ are characterizediFor the threshold value of the Weak Classifier of Haar-like features, α and β are between intervention [- 1,1]
Real number, the confidence level of presentation class result, negative value is expressed as non-pedestrian target, on the occasion of for pedestrian target；
The Weak Classifier of the HOG features includes each gradient orientation histogram for positive sample, utilizes the weighting of sample
Average value m represents pedestrian target, shown in equation below：
<mrow>
<mi>m</mi>
<mo>=</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</munderover>
<msub>
<mi>&omega;</mi>
<mi>i</mi>
</msub>
<msub>
<mi>h</mi>
<mi>i</mi>
</msub>
</mrow>
Wherein, hiValue is characterized, n is the number of positive sample, ωiFor the weights of each sample；
The Weak Classifier of HOG features calculates each histogram to weighted average m Euclidean distance to be classified：
<mrow>
<msub>
<mi>g</mi>
<mn>2</mn>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mi>&alpha;</mi>
</mtd>
<mtd>
<mrow>
<mi>d</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>h</mi>
<mi>j</mi>
</msub>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo>,</mo>
<mi>m</mi>
<mo>)</mo>
<mo><</mo>
<msub>
<mi>&theta;</mi>
<mi>j</mi>
</msub>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mi>&beta;</mi>
</mtd>
<mtd>
<mrow>
<mi>o</mi>
<mi>t</mi>
<mi>h</mi>
<mi>e</mi>
<mi>r</mi>
<mi>w</mi>
<mi>i</mi>
<mi>s</mi>
<mi>e</mi>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
In above formula, as the Weak Classifier of Haar-like features, the confidence level of α and β difference presentation class results, Zheng Zhidai
Table is pedestrian, and negative value is represented as non-pedestrian, d (hj(x) m) it is, Euclidean distance of the HOG characteristic values to weighted average m, θjFor
Threshold value.
5. the vehicle-mounted pedestrian detection method of the convergence strategy as claimed in claim 4 based on target recognition and tracking, its feature
It is, the design of the cascade classifier of S2.3 combinations the Gentle Adaboost and SVM specifically includes following steps：
S2.3.1 sets the maximum misclassification rate F of every layer of strong classifiermax∈ (0.0.5] and minimum detection rate Dmin∈[0.99,1)；
When S2.3.2 carries out the training of Gentle Adaboost cascade classifiers using the feature set in S2.1, every layer of maximum is set
Characteristic MAX, herein in maximum characteristic MAX, strong classifier, Gentle are trained using Gentle Adaboost algorithms
Adaboost training algorithm processes：
1) training sample (x is given1,y1),…,(xN,yN), set positive sample number P=1500, K=2500 pairs of negative sample number
In positive sample, y is madei=1；For negative sample, y is madei=1；
2) sample weights are initialized
<mrow>
<msub>
<mi>&omega;</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mfrac>
<mn>1</mn>
<mrow>
<mn>2</mn>
<mi>P</mi>
</mrow>
</mfrac>
</mtd>
<mtd>
<mrow>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<mn>1</mn>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mfrac>
<mn>1</mn>
<mrow>
<mn>2</mn>
<mi>K</mi>
</mrow>
</mfrac>
</mtd>
<mtd>
<mrow>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<mo>-</mo>
<mn>1</mn>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
In formula, P is positive sample number, and K is negative sample number；
3) for r=1 ..., R, R represents algorithm iteration number of times, that is, constitutes the Weak Classifier number of strong classifier；For each
Feature u, obtains Weak Classifier gu∈[-1,1]；Calculate Weak Classifier guClassification errorChoosing
Go out optimal gt；Strong classifier result G (x) is updated to G (x)+gt(x)；Sample weights are updated to ωi exp[yigt
(xi)],
Wherein i is number of samples, i=1,2 ..., N；
4) strong classifier of output is：
<mrow>
<mi>G</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mi>s</mi>
<mi>i</mi>
<mi>g</mi>
<mi>n</mi>
<mo>&lsqb;</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>r</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>R</mi>
</munderover>
<msub>
<mi>g</mi>
<mi>t</mi>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo>&rsqb;</mo>
<mo>;</mo>
</mrow>
S2.3.3 is more than MAX when the characteristic of S2.3.2 cascade graders layer, selects the MAX feature as svm classifier
The input of device, hereafter, several layers of after cascade classifier that strong classifier training is all carried out with SVM algorithm, the form of SVM classifier is：
<mrow>
<mi>y</mi>
<mo>=</mo>
<mi>s</mi>
<mi>i</mi>
<mi>g</mi>
<mi>n</mi>
<mrow>
<mo>(</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
<msub>
<mi>&alpha;</mi>
<mi>i</mi>
</msub>
<mi>K</mi>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>+</mo>
<mi>b</mi>
<mo>)</mo>
</mrow>
In above formula, x is the characteristic vector of an observed sample, and y ∈ { -1,1 } are class label, xiIt is i-th of training sample
Characteristic vector, N is the quantity of training sample, K (x, xi) it is a kernel function, α={ α1,α2,...,αNIt is the value to be solved,
Drawn by the quadratic programming problem for solving following：
<mrow>
<mi>min</mi>
<mi>Q</mi>
<mrow>
<mo>(</mo>
<mi>&alpha;</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>-</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<msub>
<mi>&alpha;</mi>
<mi>i</mi>
</msub>
<mo>+</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<msub>
<mi>&alpha;</mi>
<mi>i</mi>
</msub>
<msub>
<mi>&alpha;</mi>
<mi>j</mi>
</msub>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mi>K</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
Wherein, C >=αi>=0, i=1 ... N,
In formula, C is a predefined parameter, is penalty factor, therefore soft margin can be converted into the problem of linearly inseparable
Linear separability problem；xi,xjFor supporting vector, kernel function K (xi,xj) it is RBF；
K(xi,xj)=exp (- γ | | xi,xj||2),γ>0
In formula, xi,xjVector is characterized, γ is the parameter of kernel function；
When cascade classifier meets misclassification rate less than or equal to FmaxIt is more than or equal to minimum detection rate D with verification and measurement ratiominWhen, terminate
The training of cascade classifier.
6. the vehicle-mounted pedestrian detection side of the convergence strategy based on target recognition and tracking as described in claim any one of 1-5
Method, it is characterised in that in the S3.1, defining R=, { L, F, x', y', h', w, △ x', △ y', △ h', △ w'} represent motion
Target, wherein, L is the label of moving target, and F represents the macroscopic features of target, and (x', y') is the center of target, h', w'
Respectively height and width of target boundary rectangle, △ x', △ y' represents target's center position in x directions and y directions up conversion respectively
Speed, △ h', △ w' represents the pace of change of boundary rectangle height and width；
In the S3.2, the motion state of target is predicted using Kalman filter, for quantity of state (x', y', h', w', △
X', △ y', △ h', △ w'), Kalman filter equation can be expressed as：
Xk+1=AXk+Wk
Zk=H'Xk+Vk
In above formula, A is state-transition matrix, and H' is observing matrix, Wk、VkFor dynamic noise and observation noise, ZkBe observation to
Amount, state vector XkIt is an octuple vector：Xk=[x, ' y', h', w', △ x', △ y', △ h', △ w']T；
Assuming that pedestrian moves with uniform velocity in unit interval, state-transition matrix is represented by：
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
</mtr>
</mtable>
</mfenced>
Take observation vector Zk=[x', y', h', w']T, selecting system observing matrix is：
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
</mtable>
</mfenced>
Target state is predicted according to the Kalman filter equation and Kalman filter algorithm of structure, Kalman filters
The original state of the target of ripple is the result that step S1 and step S2 are identified；
In the S3, shown in arest neighbors method equation below：
Dis(R1,R2)=| | x '1-x′1||2+||y′1-y′2||2+||h′1-h′2||2+||w′1-w′2||2
In above formula, R1、R2Prediction target and observed object, (x ' are represented respectively1,y′1) it is the center for predicting target, (x '2,
y′2) be observed object center, h '1、w′1To predict the height and width of target boundary rectangle, h '2、w′2Outside for observed object
Connect the height and width of rectangle；
If the distance between prediction target and observation for associating are less than given threshold, the data correlation is considered as effective
, and update with observed object now the template of respective objects；If predict between target and the observation associated away from
From more than threshold value, then the association results are considered as invalid, now using the method for template matches in prediction target proximity area
New observed object is searched in domain；If template matches success, then the result of template matches is used as effective observation；If mould
Plate matches the observation for not finding respective objects, then the target may disappear or be blocked, so as to terminate tracking.
7. the vehicle-mounted pedestrian detection method of the convergence strategy as claimed in claim 5 based on target recognition and tracking, its feature
It is, using Kalman multi-object tracking methods, identification and tracking is merged, is integrated and sentenced by multiframe recognition result
It is disconnected, and different inspection policies are used to the target of different conditions, it includes：
1) for pedestrian target, the identification before pedestrian target appearance in 50 two field pictures every 10 testing property of frame,
Only pedestrian target is tracked afterwards, without carrying out cascade classifier identification；
2) for non-pedestrian target, once recognized every 5 frames；
3) for the uncertain target of state, all it is identified per frame, and judge whether it is pedestrian.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710236834.6A CN107066968A (en) | 2017-04-12 | 2017-04-12 | The vehicle-mounted pedestrian detection method of convergence strategy based on target recognition and tracking |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710236834.6A CN107066968A (en) | 2017-04-12 | 2017-04-12 | The vehicle-mounted pedestrian detection method of convergence strategy based on target recognition and tracking |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107066968A true CN107066968A (en) | 2017-08-18 |
Family
ID=59602013
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710236834.6A Pending CN107066968A (en) | 2017-04-12 | 2017-04-12 | The vehicle-mounted pedestrian detection method of convergence strategy based on target recognition and tracking |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107066968A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107679528A (en) * | 2017-11-24 | 2018-02-09 | 广西师范大学 | A kind of pedestrian detection method based on AdaBoost SVM Ensemble Learning Algorithms |
| CN108109155A (en) * | 2017-11-28 | 2018-06-01 | 东北林业大学 | A kind of automatic threshold edge detection method based on improvement Canny |
| CN108197507A (en) * | 2017-12-30 | 2018-06-22 | 刘智 | A kind of privacy real-time protection method and system |
| CN108388947A (en) * | 2018-02-06 | 2018-08-10 | 大唐高鸿信息通信研究院（义乌）有限公司 | Improvement machine learning method suitable for vehicle-mounted short haul connection net safe early warning |
| CN108573244A (en) * | 2018-04-28 | 2018-09-25 | 济南浪潮高新科技投资发展有限公司 | A kind of vehicle checking method, apparatus and system |
| CN108805198A (en) * | 2018-06-08 | 2018-11-13 | Oppo广东移动通信有限公司 | Image processing method, device, computer readable storage medium and electronic equipment |
| CN108805901A (en) * | 2018-05-04 | 2018-11-13 | 北京航空航天大学 | A kind of quick detecting and tracking parallel computation of sensation target based on multi-core DSP and fusion method |
| CN108875668A (en) * | 2018-06-28 | 2018-11-23 | 东北大学 | A kind of pedestrian detection method based on multiple features and cascade classifier |
| CN109472218A (en) * | 2018-10-22 | 2019-03-15 | 北京建筑大学 | A kind of automobile assistant driving safety method based on pedestrian's direction discernment |
| CN109543610A (en) * | 2018-11-22 | 2019-03-29 | 中国科学院长春光学精密机械与物理研究所 | Vehicle detecting and tracking method, device, equipment and storage medium |
| CN112102280A (en) * | 2020-09-11 | 2020-12-18 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting loosening and loss faults of small part bearing key nut of railway wagon |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102222214A (en) * | 2011-05-09 | 2011-10-19 | 苏州易斯康信息科技有限公司 | Fast object recognition algorithm |
| CN103077539A (en) * | 2013-01-23 | 2013-05-01 | 上海交通大学 | Moving object tracking method under complicated background and sheltering condition |
| CN105335701A (en) * | 2015-09-30 | 2016-02-17 | 中国科学院合肥物质科学研究院 | Pedestrian detection method based on HOG and D-S evidence theory multi-information fusion |
-
2017
- 2017-04-12 CN CN201710236834.6A patent/CN107066968A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102222214A (en) * | 2011-05-09 | 2011-10-19 | 苏州易斯康信息科技有限公司 | Fast object recognition algorithm |
| CN103077539A (en) * | 2013-01-23 | 2013-05-01 | 上海交通大学 | Moving object tracking method under complicated background and sheltering condition |
| CN105335701A (en) * | 2015-09-30 | 2016-02-17 | 中国科学院合肥物质科学研究院 | Pedestrian detection method based on HOG and D-S evidence theory multi-information fusion |
Non-Patent Citations (4)
| Title |
|---|
| 李婷俐: "基于多特征级联分类器的道路车辆检测方法研究", 《中国优秀硕士学位论文全文数据库（电子期刊） 信息科技辑》 * |
| 葛俊锋 等: "一种改进的夜间行人检测算法", 《计算机工程》 * |
| 赵娜娜 等: "一种智能夜间行人检测算法", 《第19届全国煤矿自动化与信息化学术会议暨中国矿业大学(北京)百年校庆学术会议论文集》 * |
| 邹薇 等: "一种基于运动检测的行人多目标跟踪算法", 《计算机应用与软件》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107679528A (en) * | 2017-11-24 | 2018-02-09 | 广西师范大学 | A kind of pedestrian detection method based on AdaBoost SVM Ensemble Learning Algorithms |
| CN108109155A (en) * | 2017-11-28 | 2018-06-01 | 东北林业大学 | A kind of automatic threshold edge detection method based on improvement Canny |
| CN108197507A (en) * | 2017-12-30 | 2018-06-22 | 刘智 | A kind of privacy real-time protection method and system |
| CN108388947A (en) * | 2018-02-06 | 2018-08-10 | 大唐高鸿信息通信研究院（义乌）有限公司 | Improvement machine learning method suitable for vehicle-mounted short haul connection net safe early warning |
| CN108573244A (en) * | 2018-04-28 | 2018-09-25 | 济南浪潮高新科技投资发展有限公司 | A kind of vehicle checking method, apparatus and system |
| CN108573244B (en) * | 2018-04-28 | 2020-07-14 | 浪潮集团有限公司 | Vehicle detection method, device and system |
| CN108805901A (en) * | 2018-05-04 | 2018-11-13 | 北京航空航天大学 | A kind of quick detecting and tracking parallel computation of sensation target based on multi-core DSP and fusion method |
| CN108805901B (en) * | 2018-05-04 | 2022-02-22 | 北京航空航天大学 | Visual target rapid detection tracking parallel computing and fusion method based on multi-core DSP |
| CN108805198A (en) * | 2018-06-08 | 2018-11-13 | Oppo广东移动通信有限公司 | Image processing method, device, computer readable storage medium and electronic equipment |
| CN108875668A (en) * | 2018-06-28 | 2018-11-23 | 东北大学 | A kind of pedestrian detection method based on multiple features and cascade classifier |
| CN109472218A (en) * | 2018-10-22 | 2019-03-15 | 北京建筑大学 | A kind of automobile assistant driving safety method based on pedestrian's direction discernment |
| CN109543610A (en) * | 2018-11-22 | 2019-03-29 | 中国科学院长春光学精密机械与物理研究所 | Vehicle detecting and tracking method, device, equipment and storage medium |
| CN112102280A (en) * | 2020-09-11 | 2020-12-18 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting loosening and loss faults of small part bearing key nut of railway wagon |
| CN112102280B (en) * | 2020-09-11 | 2021-03-23 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting loosening and loss faults of small part bearing key nut of railway wagon |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107066968A (en) | The vehicle-mounted pedestrian detection method of convergence strategy based on target recognition and tracking | |
| CN107031623B (en) | A kind of road method for early warning based on vehicle-mounted blind area camera | |
| Adu-Gyamfi et al. | Automated vehicle recognition with deep convolutional neural networks | |
| Fang et al. | Road-sign detection and tracking | |
| CN107463890B (en) | A kind of Foregut fermenters and tracking based on monocular forward sight camera | |
| JP5124592B2 (en) | System and method for detecting and tracking a vehicle | |
| CN105260712B (en) | A kind of vehicle front pedestrian detection method and system | |
| CN106652465A (en) | Method and system for identifying abnormal driving behavior on road | |
| Huang et al. | Vehicle detection and inter-vehicle distance estimation using single-lens video camera on urban/suburb roads | |
| Sivaraman et al. | Combining monocular and stereo-vision for real-time vehicle ranging and tracking on multilane highways | |
| EP2061008B1 (en) | Method and device for continuous figure-ground segmentation in images from dynamic visual scenes | |
| US9626599B2 (en) | Reconfigurable clear path detection system | |
| US20090309966A1 (en) | Method of detecting moving objects | |
| CN106682586A (en) | Method for real-time lane line detection based on vision under complex lighting conditions | |
| CN103020986A (en) | Method for tracking moving object | |
| CN107031661A (en) | A kind of lane change method for early warning and system based on blind area camera input | |
| CN104951758A (en) | Vehicle-mounted method and vehicle-mounted system for detecting and tracking pedestrians based on vision under urban environment | |
| CN109697727A (en) | Method for tracking target, system and storage medium based on correlation filtering and metric learning | |
| Arthi et al. | Object detection of autonomous vehicles under adverse weather conditions | |
| BOURJA et al. | Real time vehicle detection, tracking, and inter-vehicle distance estimation based on stereovision and deep learning using YOLOv3 | |
| Rajalakshmi et al. | Traffic violation invigilation using transfer learning | |
| Lafuente-Arroyo et al. | Road sign tracking with a predictive filter solution | |
| Ramzan et al. | Intelligent pedestrian detection using optical flow and hog | |
| Tourani et al. | Challenges of video-based vehicle detection and tracking in intelligent transportation systems | |
| Kaur et al. | Local Regression Based Real-Time Traffic Sign Detection using YOLOv6 |
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: 20170818 |