Abstract-When multiple objects are occluded they are obtruded from human vision. A new agency has been proposed to track such objects. This paper proposes a fresh attack to observe and track occluded traveling objects. A filter bank attack has been used to foretell the gesture of the traveling marks utilizing the current province and the old few provinces. The additive Kalman filter, the drawn-out Kalman filter ( EKF ) and the unscented Kalman filter ( UKF ) have been used in the filter bank. Kalman Filter uses measurings that are observed over clip and bring forth values that tend to be closer to the existent values of the measurings. The drawn-out Kalman filter ( EKF ) is the nonlinear filter which linearizes about the current mean and covariance utilizing a Taylor series based transformation.The Unscented Kalman Filters utilizes a set of sample points, which guarantees truth with the posterior mean and covariance to the 2nd order for any nonlinearity For an experimental rating, sequences consisting of uninterrupted informations are used. The algorithm is used to prove existent clip informations.

Keywords- Occlusion handling, mark trailing, filter Bankss,

Kalman filter, EKF, UKF.

Introduction

Trailing is the building of correspondence of relationships between “ tracked objects ” in old frames

and “ detected objects ” in the current frame [ 7 ] . When such objects move behind some other objects they become undetected due to occlusion and tracking such objects becomes hard. Occlusion consequences in obtrusion of vision. In this instance, building of correspondence becomes hard.

Tracking is particularly of import in surveillance. In general, the processing model of ocular surveillance in dynamic scenes includes the undermentioned phases: ( 1 ) objects sensing,

( 2 ) trailing, ( 3 ) apprehension and description of behaviours, and ( 4 ) designation [ 9 ]

The paper is organized as follows. In Section 2, the related work is described. In Section 3 the new algorithm is proposed. In Section 4 the algorithm is evaluated on image sequences. In Section 5 conclusive comments are provided.

Related Work

An incorporate pilotage information system is an embedded system installed in a auto which must cognize continuously the current place with a good preciseness. The drawn-out Kalman filter can be used for positioning the pilotage information system. When the anticipation and update maps are extremely non additive computation becomes complex, so UKF has been proposed. This paper compares the public presentation of UKF and EKF.

Comparing the discrepancy of both the filters, it is found that the unscented Kalman filter ( UKF ) has somewhat better consequences for positioning than the drawn-out Kalman filter ( EKF ) .It uses a complex merger method which improves the preciseness of estimation.When there is no GPS solution available there is no addition in public presentation. It is found that the computational clip of the UKF is greater than the EKF. So a Bank of filter has been proposed which takes attention of the computational clip. [ 8 ]

A merger of two non additive filters ( SOEKF and UKF ) have been introduced to assist the trailing of steering mark. UKF overcomes the occasional divergency of SOEKF and SOEKF overcomes the job of covering with extremely non additive equations. It is found that the merger algorithms enhances the truth.

A method for tracking traveling objects earlier, during and after occlusion has been proposed. A combination of visual aspect and gesture information is used [ 2 ] .By dynamically altering covariance of procedure noise and measuring noise, it is possible to turn up objects accurately in the image plane even when they are partly or to the full occluded.

Kalman filter is used to pattern the places and speeds of objects and is so combined with colour histogram which describes the visual aspects of objects.Using colour histogram makes it unreliable.When different objects have similar colored visual aspect it produces incorrect correspondence. [ 2 ]

Steering marks do non hold a consecutive gesture and a changeless speed. This maneuvering causes the procedure equation to be nonlinear. The Kalman filter being a additive filter can non be applied for non additive systems. so the system have to be linearised utilizing first-order Taylor series enlargement, ensuing an drawn-out Kalman filter. [ 6 ]

In the instance of steering mark, this shortness of Taylor series may take to divergence. Second-order extended Kalman filter has beeen used which computes Hessian matrix for sing the second-order term of Taylor series in add-on to the first-order term which is computed from the Jacobian matrix. [ 6 ] Sing the higher order footings will do the computation so complicated in ( SOEKF ) . In this instance, a nonlinear province calculator like UKF should be used.

PROPOSED Work

Occlusion can be classified into three classs: ego occlusion, interobject occlusion, and occlusion by the background scene construction. Self occlusion occurs when one portion of the object occludes another. This state of affairs most often arises while tracking articulated objects. Interobject occlusion occurs when two objects being tracked occlude each other. Similarly, occlusion by the background occurs when a construction in the background occludes the tracked objects. Here a method is proposed to manage interobject occlusion.

This method can be used to track changing figure of traveling objects in presence of occlusions. Both the visual aspect and the gesture information are considered while tracking. In additive and Gaussian conditions, the optimum solution about gesture province of traveling objects can be obtained by utilizing the Kalman filter.

Therefore, a method is proposed to track the object based on both visual aspect and gesture information. In non additive conditions, the Extended Kalman filter and the Unscented Kalman filter can be used. Thus the Kalman filter bank contains these three filters.

SYSTEM ARCHITECTURE

The overview of the system is given in Fig.1 The stairss involved are 1 ) Organization of Kalman Filter Bank

2 ) Design of Filter Bank 3 ) Appraisal of filter parametric quantities.

The first measure involved is to plan the filter bank. The filters can either be arranged in consecutive order or can be arranged to organize a feedback cringle. In the following phase, we estimate the Kalman filter parametric quantities.

The assorted stairss involved in the building of the Kalman filter bank are described in the subdivisions below.

Fig.1 System Overview

Organization of Kalman Filter Bank

The Kalman filter bank describes a system with a system province theoretical account and a measurement theoretical account as in Equation 1:

s ( K ) = ( k a?’ 1 ) s ( k a?’ 1 ) + tungsten ( K ) — — – ( 1 )

m ( K ) = H ( K ) s ( K ) + V ( K ) — — — — — — – ( 2 )

m ( K ) – measuring theoretical account

s ( K ) – system province theoretical account

tungsten ( K ) -state noise

V ( K ) – measuring noise

The system province s ( K ) at the kth clip frame is linearly associated with the province at the ( k a?’ 1 ) -th clip frame, and there is besides a additive relationship between the measuring m ( K ) and the system province s ( K ) . The province noise and the measuring noise are assumed to be independent of each other, and have a white Gaussian distribution ( i.e ) P ( tungsten ) ~ N ( 0, Q ) and P ( V ) ~ N ( 0, R ) .

In Equation 1, s ( K ) is called the province passage matrix that relates the province at clip frame K to the province at frame K + 1, and H ( K ) is called the observation matrix that relates the province to the measuring. The Kalman filter bank can be used to successfully carry through gauging gesture information under a deteriorating status such as occlusion. The Bank of Kalman filter utilizes the relational information such as place, speed and acceleration among sub parts of a traveling object. The relational information is used to supplement the undependable measurings on a partly occluded sub part, so that an a priori estimation of the following province of the sub-region might be obtained based on the relational information every bit good as the existent measurings.

The filter estimates the system province at a peculiar clip and so obtains its feedback in the signifier of measurings. The undertakings of the Kalman filter autumn into two stages:

Prediction measure:

The anticipation measure is responsible for projecting frontward the current province and priori estimations.

Correction measure:

The rectification measure is responsible for the feedback. It incorporates an existent measuring into the priori estimation to obtain an improved posteriori estimation. The improved posterior estimation in bend provenders into the anticipation measure and the anticipation, rectification rhythm is repeated.

The Kalman filter bank estimates the system province by utilizing a signifier of feedback control. Figure 2 shows the feedback rhythm.

Fig.2 The Feedback rhythm

Design OF FILTER BANK

A filter bank is an array of band-pass filters that separates the input signal into multiple constituents, each one transporting a individual frequence bomber set of the original signal. To accomplish an accurate estimation with a sufficiently simple dynamic theoretical account and filter method, several filters can be used, each adopted to depict a specific characteristic.

The Filter Bank has the undermentioned filters:

1 ) Kalman filter

Its intent is to utilize measurings that are observed over clip that contain noise and other inaccuracies, and bring forth values that tend to be closer to the existent values of the measurings.

Kalman filter has two stairss:

Prediction measure: The following province of the system is predicted utilizing the old measurings in this measure.

Update measure: The current province of the system is estimated utilizing the measuring.

The Kalman Filter addresses the basic job of appraisal of the province of a discrete-time controlled procedure that is governed by the additive difference equation,

Ten K = A Ten k-1 + B U k-1 + W k-1 — — — ( 3 )

Z K = H X k + V K — — — ( 4 )

F k – province passage theoretical account

H k – observation theoretical account

Q k – covariance of the procedure noise ( Wk )

R k – covariance of the observation noise ( Vk )

Bk -control-input theoretical account

The posteriori estimation mistake can be calculated as,

Ek = Xk – X^ K — — – ( 5 )

The posteriori estimation mistake covariance is calculated utilizing the equation,

Pk = E [ Ek ETk ] — — — ( 6 )

Prediction:

Predicted province utilizing the kalman filter is:

Ten k|k-1 = F K X^ k-1|k-1 + BK UK — — ( 7 )

Predicted estimation covariance is calculated as,

P k|k-1 = Fk P k-1|k-1 FTK + Qk — — ( 8 )

The two variables that can stand for the filter:

X^ k|k – posteriori province estimation at clip K

P k|k – posteriori mistake covariance matrix

Though Kalman filters provides a convenient step of appraisal truth and fuses information from multiple-sensors, the Kalman filter buzzword be applied to non additive systems.

The advantages of Kalman Filters are that it provides a convenient step of appraisal ccuracy ( via the covariance matrix P ) and it can blend information from multiple-sensors.

2 ) Extended Kalman Filter ( EKF )

The drawn-out Kalman filter ( EKF ) is the nonlinear version of the Kalman filter which linearizes about the current mean and covariance.

The drawn-out Kalman filter extends the range of Kalman filter to nonlinear optimum filtrating jobs by organizing a Gaussian estimate to the joint distribution of province tens and measurings y utilizing a Taylor series based transmutation.The advantages of the drawn-out Kalman filter are

The province passage and observation theoretical accounts need non be additive maps of the province but may alternatively be differentiable maps.

Simple, intuitive and computationally efficient.

The province can be estimated utilizing the Extended Kalman Filter as given in the equation 9,

Xk = degree Fahrenheit ( Ten k-1, U k-1 ) + W k-1 — — ( 9 )

Zk = H ( X K ) + Vk — — ( 10 )

degree Fahrenheit, h – maps used for anticipation

Prediction Equation:

The Predicted province is calculated as,

X^ k|k-1 = degree Fahrenheit ( x^ k-1|k-1, U k-1 ) — – ( 11 )

The Predicted estimation covariance is,

3 ) Unscented Kalman Filters

In the UKF, the chance denseness is approximated by the nonlinear transmutation of a random variable, which returns much more accurate consequences than the first-order Taylor enlargement of the nonlinear maps in the EKF.

The estimate utilizes a set of sample points, which guarantees truth with the posterior mean and covariance to the 2nd order for any nonlinearity. The UKF tends to be more robust and more accurate than the EKF in its appraisal of mistake.

The mean and covariance of the procedure noise is used in increasing the estimated province and covariance as follows,

X a k-1|k-1 = [ X^ T k-1|k-1 E [ wTk ] ] T — — ( 13 )

P a k-1|k-1 = [ P k-1|k-1 Qk ] — — – ( 14 )

L – The dimension of the augmented province

The province and covariance aid in derivation of a set of ( 2L+1 ) sigma points and

X 0 k-1|k-1 = X a k-1|k-1 — — ( 15 )

Ten one k-1|k-1 = X a k-1|k-1 + ( a?s ( L+I» ) Pa k-1|k-1 ) I

i=1aˆ¦.L — — – ( 16 )

I» = I± 2 ( L+k ) -L — — ( 17 )

I» – grading factor

I± and k – control the spread of the sigma points

The predicted province and covariance are produced by recombination of the leaden sigma points as follows,

X^ K|k-1 = a?‘ 2Li=0 W is Ten ik|k-1 — — ( 18 )

P k|k-1 = a?‘2L i=0 Wic [ Xi k|k-1 – X^ k|k-1 ] [ X I k|k-1 – X^ k|k-1 _ — — ( 19 )

And the weights for the province and covariance are given by,

W Intelligence Community = 1/ ( 2 ( L+ I» ) ) — — ( 20 )

Appraisal OF FILTER PARAMETERS

Fig.3 Estimation of filter parametric quantities

The anticipation measure is responsible for projecting frontward the current province ( xk ) and measuring ( mk ) for the following clip frame.

The rectification measure of the Kalman filter bank combines the procedure of ciphering the relational measuring vector and updating the filter measuring. It computes a posteriori estimation of the system province as a additive combination of a priori estimation and a leaden amount of differences between existent measurings and predicted measurings.

OUR METHOD

Low-level formatting with X^ ( 0|0 ) , P^ ( 0|0 ) , k=1

Deterministically compute 2n+1 sample points eleven ( k-1|k-1 ) and the weights Wisconsin ( m ) , wi ( degree Celsius ) , i=0, aˆ¦.2n where N is the dimension of the province.

X0 ( k-1|k-1 ) =X^ ( k-1|k-1 )

Eleven ( k-1|k-1 ) =X^ ( k-1|k-1 ) + ( a?s ( n+I» ) P^ ( k-1|k-1 ) ) I, i=1, … N

Eleven ( k-1|k-1 ) =X^ ( k-1|k-1 ) – ( a?s ( n+I» ) P^ ( k-1|k-1 ) ) I -n, i=n+1, … 2n

( 3 ) Prediction:

Propagate the samples in footings of the province and calculate the predicted mean province

Eleven ( k|k-1 ) =AXi ( k-1|k-1 ) + ( I-A ) X^ i=0, aˆ¦2n

X^ ( k|k-1 ) =a?‘i=02n Wi ( m ) Eleven ( k|k-1 )

Make anticipation to obtain Yi ( k|k-1 ) and calculate its point denseness pi ( k|k-1 ) =a??j=1m pj, so calculate the predicted observation mean

Y^ ( k|k-1 ) =a?‘i=02n Wi ( m ) Yi ( k|k-1 )

Calculate the predicted covariance

P^ ( k|k-1 ) =a?‘i=02n Wi ( degree Celsius ) [ Xi ( k|k-1 ) -X^ ( k|k-1 ) ] [ Xi ( k|k-1 ) -X^ ( k|k-1 ) ] T +Rw

( 4 ) Measurement update:

Calculate the car correlativity of measuring and correlativity of province and measuring as

P^Yk Yk =a?‘i=02n Wi ( degree Celsius ) [ Yi ( k|k-1 ) -Y^ ( k|k-1 ) ] [ Yi ( k|k-1 ) -Y^ ( k|k-1 ) ] T +Rv

P^xk xk =a?‘i=02n Wi ( degree Celsius ) [ Xi ( k|k-1 ) -X^ ( k|k-1 ) ] [ Yi ( k|k-1 ) -Y^ ( k|k-1 ) ] T

Determine the true measuring as Y ( K ) =Yj ( k|k-1 ) fow which pi ( K ) =maxi=0, aˆ¦..2n pi ( k|k-1 ) , so update mean and covariance

K^ ( K ) =P^xk ykP^yk yk-1

X^ ( k|k ) =X^ ( k|k-1|+k^ ( K ) ( Y ( K ) -Y^ ( k|k-1 )

P^ ( k|k ) =P^ ( k|k-1|+K^ ( K ) P^ yk ykK^ ( K ) Thymine

( 5 ) k=k+1.Goto measure 2.

Experimental consequences

To verify the proposed method assorted experiments were carried out on both uninterrupted informations and existent images. Figures 4,5 and 6 depict the mistake secret plans of utilizing the Kalman filter, EKF and UKF singularly.

Fig.4 Mean mistake secret plan utilizing Kalman filter

Fig.5 Mean Error Plot utilizing Extended Kalman Filter

Fig.6 Mean Error secret plan utilizing Unscented Kalman filter

The same set of information is given as input to the Kalman filter bank ( KFB ) .The error secret plans obtained are given as Figures 7 and 8.

Fig.7 Before utilizing KFB

Fig.8 After utilizing KFB

It is clearly seen that the Kalman Filter Bank outperforms each of the single filters.

To farther prove the proposed method, .PNG and.PPM images obtained from a.AVI picture file were given as inputs. The tracked object is represented by jumping objects. The green bounding box denotes the boundary of the tracked object. The ruddy cross denotes the estimated place ( utilizing KFB ) of the tracked object.

The sequence of the images used and the consequences are given in Figure 9. It is seen that the traveling objects can be tracked accurately utilizing this method.

Fig 9. Tracking utilizing Kalman Filter Bank

Decision

This paper proposes a new method to accurately track traveling objects under both normal and occluded conditions. Both the visual aspect and gesture information have been used for tracking in the above mentioned complex environments. The experimental consequences show the truth and hardiness of the method.

This work chiefly focuses on inter object occlusion. We further hope to heighten the work to cover both self occlusion and occlusion due to play down minus.