Object tracking using particle filter презентация

Object Tracking using Particle Filter Nandini Easwar Jogen Shah CIS 601, Fall 2003

Overview Background Information Basic Particle Filter Theory Rao Blackwellised Particle Filter Color Based Probabilistic Tracking

Object Tracking Tracking objects in video involves the modeling of non-linear and non-gaussian systems. Non-Linear Non-Gaussian

Background In order to model accurately the underlying dynamics of a physical system, it is important to include elements of non-linearity and non-gaussianity in many application areas. Particle Filters can be used to achieve this. They are sequential Monte Carlo methods based on point mass representations of probability densities, which are applied to any state model.

The Particle Filter Particle Filter is concerned with the problem of tracking single and multiple objects. Particle Filter is a hypothesis tracker, that approximates the filtered posterior distribution by a set of weighted particles. It weights particles based on a likelihood score and then propagates these particles according to a motion model.

Mathematical Background Particle Filtering estimates the state of the system, x t, as time t as the Posterior distribution: P( x t | y 0-t ) Let, Est (t) = P( x t | y 0-t ) Est(1) can be initialized using prior knowledge

Mathematical Background Particle filtering assumes a Markov Model for system state estimation. Markov model states that past and future states are conditionally independent given current state. Thus, observations are dependent only on current state.

Mathematical Background Est(t) = P( x t | y 0 - t ) = p(y t | x t, y 0 – t-1).P(x t | y 0 – t-1) (Using Baye’s Theorem) = p(y t | x t ). P(x t | y 0 – t-1) (Using Markov model) = p(y t | x t ). P(x t |x t-1).P(x t-1 | y 0 – t-1) = p(y t | x t ). P(x t |x t-1).Est(t-1)

Mathematical Background Final Result: Est(t) = p(y t | x t ). P(x t |x t-1).Est(t-1) Where: p(y t | x t ): Observation Model P(x t |x t-1).Est(t-1): Proposal distribution

Mathematical Background To implement Particle Filter we need State Motion model: P(x t |x t-1) Observation Model: p(y t | x t ): Initial State: Est(1)

Mathematical Background We sample from the proposal and not the posterior for estimation. To take into account that we will be sampling from wrong distribution, the samples have to be likelihood weighed by ratio of posterior and proposal distribution: W t = Posterior i.e.Est (t) / proposal Distribution = p(y t | x t ) Thus, weight of particle should be changed depending on observation for current frame.

Basic Particle Filter Theory A discrete set of samples or particles represents the object-state and evolves over time driven by the means of "survival of the fittest". Nonlinear motion models can be used to predict object-states.

Basic Particle Filter Theory (Cont.) Particle Filter is concerned with the estimation of the distribution of a stochastic process at any time instant, given some partial information up to that time. The basic model usually consists of a Markov chain X and a possibly nonlinear observation Y with observational noise V independent of the signal X.

Basic Particle Filter Theory (Cont.) System Dynamics ie.Motion Model: p(x t| x 0:t-1) Observation Model: p(y t | x t) Posterior Distribution: p(x t | y o..t) Proposal Distribution is the Motion Model Weight, w t = Posterior / Proposal = observation

Basic Particle Filter Theory (Cont.) Given N particles (samples) {x(i)0:t-1,z(i)0:t-1}Ni=1 at time t-1, approximately distributed according to the distribution P(dx(i)0:t-1,z(i)0:t-1|y1:t-1), particle filters enable us to compute N particles {x(i)0:t,z(i)0:t}Ni=1 approximately distributed according to the posterior distribution P(dx(i)0:t,z(i)0:t|y1:t)

Basic Particle Filter Theory (Cont.) The basic Particle Filter algorithm consists of 2 steps: Sequential importance sampling step Selection step

Particle Filter Algorithm Sequential importance sampling Uses Sequential Monte Carlo simulation. For each particle at time t, we sample from the transition priors For each particle we then evaluate and normalize the importance weights.

Particle Filter Algorithm Selection Step Multiply or discard particles with respect to high or low importance weights w(i)t to obtain N particles. This selection step is what allows us to track moving objects efficiently.

Rao-Blackwellised Particle Filter RBPF is an extension on PF. It uses PF to compute the distribution of discrete state with Kalman Filter to compute the distribution of continuous state. For each sample of the discrete states, the mean and covariance of the continuous state are updated using the exact computations. We have implemented the particle filter algorithm and not the RBPF.

RBPF Approach RBPF models the states as <Ct,Dt> Ct is the continuous state representation Dt is the discrete state representation The aim of this approach is to predict the discrete state Dt. However, for our object tracking application, the above approach was unsuitable.

Implementation We have implemented the Particle Filter algorithm in Matlab. Our approach towards this project: Reading research papers on PF given to us by Dr.Latecki. Trying to implement PF-RBPF algorithm written by Nando de Freitas.

Implementation Color Based Probabilistic Tracking These trackers rely on the deterministic search of a window, whose color content matches a reference histogram color model. Uses principle of color histogram distance. This color based tracking is very flexible and can be extended in many ways.

Color Based Probabilistic Tracking The combination of tools used to accomplish a given tracking task depends on whether one tries to track: Objects of a given nature eg.cars,faces Objects of a given nature with a specific attribute eg.moving cars, face of specific person Objects of unknown nature, but of specific interest to us eg.moving objects.

Color Based Probabilistic Tracking Reference Color Window The target object to be tracked forms the reference color window. Its histogram is calculated, which is used to compute the histogram distance while performing a deterministic search for a matching window.

Color Based Probabilistic Tracking State Space We have modeled the states, as its location in each frame of the video. The state space is represented in the spatial domain as: X = ( x , y ) We have initialized the state space for the first frame manually.

Color Based Probabilistic Tracking System Dynamics A second-order auto-regressive dynamics is chosen on the parameters used to represent our state space i.e (x,y). The dynamics is given as: Xt+1 = Axt + Bxt-1 Matrices A and B could be learned from a set of sequences where correct tracks have been obtained. We have used an ad-hoc model for our implementation.

Color Based Probabilistic Tracking Observation yt The observation yt is proportional to the histogram distance between the color window of the predicted location in the frame and the reference color window. Yt α Dist(q,qx), Where q = reference color histogram. qx = color histogram of predicted location.

Color Based Probabilistic Tracking Particle Filter Iteration Steps: Initialize xt for first frame Generate a particle set of N particles {xmt}m=1..N Prediction for each particle using second order auto-regressive dynamics. Compute histogram distance Weigh each particle based on histogram distance Select the location of target as a particle with minimum histogram distance. Sampling the particles for next iteration.

Color Based Probabilistic Tracking An step by step look at our code, highlighting the concepts applied: Initialization of state space for the first frame and calculating the reference histogram: reference = imread('reference.jpg'); [ref_count,ref_bin] = imhist(reference); x1= 45; y1= 45; Describing the N particles within a specified window: for i = 1:N x(1,i,1) = x1 + 50 * rand(1) - 50 *rand(1); x(2,i,1) = y1 + 50 * rand(1) - 50 *rand(1); end

Color Based Probabilistic Tracking For each particle, we apply the second order dynamics equation to predict new states: if (j==2) x(:,i,j) = A * x(:,i,j-1); else x(:,i,j)=rand(n_x)*x(:,i,j-1)+rand(n_x)*x(:,i,j-2); The color window is defined and the histogram is calculated: rect = [(x(1,i,j)-15),(x(2,i,j)-15),30,30]; [count,binnumber] = imhist(imcrop(I(:,:,:,j),rect));

Color Based Probabilistic Tracking Calculate the histogram distance: for k = 1:255 d( I , j ) = d( i, j ) + (double ( count ( k ) ) - double(ref_count( k ) ) ) ^ 2; end Calculating the normalized weight for each particle: w(:,j) = w(:,j)./sum(w(:,j)); w(:,j) = one(:,1) - w(:,j);

Color Based Probabilistic Tracking Re-sampling step, where the new particle set is chosen: for i = 1:N x(1,i,j) = state(1,j) + 50 * rand(1) - 50 *rand(1); x(2,i,j) = state(2,j) + 50 * rand(1) - 50 *rand(1); end

Color Based Probabilistic Tracking Functions Used: Track_final1.m : PF tracking code multinomialR.m : Resampling function.

Color Based Probabilistic Tracking: Results

Applications Video Surveillance Gesture HCI Reality and Visual Effects Medical Imaging State estimation of Rovers in outer-space.

Future Work Automatic initialization of reference window. Multi part color window. Multi-object tracking.

References M. Isard and A. Blake. Condensation–conditional density propagation for visual tracking. Int. J. Computer Vision, 29(1):5–28, 1998. D. Reid, “An algorithm for tracking multiple targets,” IEEE Trans. on Automation and Control, vol. AC-24,pp. 84–90, December 1979. N. Gordon, D. Salmond, and A. Smith, “Novel approach to nonlinear/non-Gaussian Bayesian state estimation,” IEEE Procedings F, vol. 140, no. 2, pp. 107–113, 1993. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A tutorial on particle filters for on-line non-linear/non-Gaussian Bayesian tracking,” IEEE Transactions on Signal Processing, vol. 50, pp. 174–188, Feb. 2002.

Thank You