三维计算机视觉 3D computer vision（基于卡尔曼滤波的运动结构）

3D computer vision Structure from motion using Kalman filter SFM Kalman V9a 1

SFM Kalman V9a 1 3D computer vision Structure from motion using Kalman filter

Aims To obtain structure and camera motion from an image sequence Utilize inter-picture dynamics of a sequence Such as constant speed acceleration of the camera etc SFMKalman vga 2

SFM Kalman V9a 2 Aims • To obtain structure and camera motion from an image sequence • Utilize inter-picture dynamics of a sequence – Such as constant speed, acceleration of the camera etc

Tracking methods Kalman filtering suitable for systems with Gaussian noise Condensation(or called particle filter) suitable for systems with Non-Gaussian noise SFMKalman vga

SFM Kalman V9a 3 Tracking methods • Kalman filtering, suitable for systems with Gaussian noise • Condensation (or called particle filter), suitable for systems with Non-Gaussian noise

Part O Basic concept of Kalman filter SFMKalman vga

SFM Kalman V9a 4 Part 0 Basic concept of Kalman filter

Introduction A system(e.g. radar tracking a plane)can be modelled by a system transition dynamic function using the Newtons' law (linear) Measurement may contain noise(assume Ga aussian Kalman filter predict and update the system to reduce the effect of noise SFMKalman vga

Introduction • A system (e.g. radar tracking a plane) can be modelled by a system transition dynamic function using the Newtons’ law (linear). • Measurement may contain noise (assume Gaussian) • Kalman filter predict and update the system to reduce the effect of noise. SFM Kalman V9a 5

System state and x=u, dynamic position, velocity/ Newtons law:lk=l4-1+l·△t xk is the state at time k Soxx= Axk_+o, where A k-l hence a is the motion model Q is system noise(by wind) K Li]1o 1t Z is measurement on the Ax+O du =-= velocit radar screen d t R is measurement noise Assume u>>h.soZ≈l Rad SFMKalman Vga scre

System state and dynamic • xk is the state at time k – xk=A*xk-1 – A is the motion model – Q is system noise (by wind) • Z is measurement on the radar screen, – R is measurement noise – Assume u>>h, so Zu SFM Kalman V9a 6 Ax Q Q u t u u u x t A x Ax Q where u u u t [position, velocity] u u dt du x u k k k k k k k k k k T T T = + +             =      =        = = + = +  = =       = − − − − − 1 1 1 1 1 0 1 1 ,hence 0 1 1 So , Newtons' law : , [ , ]     u = position z Radar screen velocity dt du u = = h

Can add acceleration if we want x=lu,u, u position, velocity, acceleration/ Newtons' lay:l=l421+i△+(△) SO x= Ax,+o where △M0.5(△7)2 A=01△t hence △t0.5(△1)ak-1 -1 Ax +o SFMKalman vga

Can add acceleration if we want • SFM Kalman V9a 7 ( ) ( ) ( ) Ax Q u u u t t t u u u x t t t A x Ax Q where u u u t u t [position, velocity acceleration] x u u u k k k k k k k k k k k k T T = +                        =           =              = = + = +  +  = = − − − − − − 1 1 1 1 2 2 1 2 1 0 0 1 0 1 1 0.5 ,hence 0 0 1 0 1 1 0.5 So , 2 1 Newtons' law : , [ , , ]        

Kalman filter always predict and update to find the state of the plane x Kalman filter offers optimum prediction by considering the system and measurement noise eviiserror kerro下 ev = error △R tk(predicted at time k-1) At time k+1 At time k xk l(actual state at time k- SFM Kalman voa

Kalman filter SFM Kalman V9a 8 xk-1 (actual state at time k-1) xk At time k xk+1 At time k+1 ek-1=error • Always predict and update to find the state of the plane x • Kalman filter offers optimum prediction by considering the system and measurement noise ek=error ek+1=error ˆ (predicted at time 1) 1 x k- k− k x ˆ 1 ˆ k+ x

Part 1 Introduction to Kalman filter(KF)and Extended Kalman filters(EKF SFMKalman vga

SFM Kalman V9a 9 Part 1 Introduction to Kalman filter (KF) and Extended Kalman filters (EKF)

Kalman filter introduction B ased on An introduction to the kalman Filter Source Technical Report: TR95-041Year of Publication: 1995 Authors Greg Welch gary Bishop publisher university of north carolina at Chapel Hill Chapel Hill, NC, US (http://www.cs.unc.edu/welch/media/pdf/kalma n intro. pdf) SFM Kalman vga

SFM Kalman V9a 10 Kalman filter introduction • Based on • An Introduction to the Kalman FilterSourceTechnical Report: TR95-041 Year of Publication: 1995 Authors Greg Welch Gary Bishop Publisher University of North Carolina at Chapel Hill Chapel Hill, NC, US (http://www.cs.unc.edu/~welch/media/pdf/kalma n_intro.pdf)