Miscellaneous. Regarding reading materials. Again, ask questions (if you have) and ask them earlier

Transcription

1 Miscellaneous Regarding reading materials Reading materials will be provided as needed If no assigned reading, it means I think the material from class is sufficient Should be enough for you to do your homework But, this is 4xx, so not directly applying formulas careful reasoning required It may require you dig out your old textbooks (e.g. calculus, etc.) Again, ask questions (if you have) and ask them earlier

2 CS 460/560 Introduction to Computational Robotics Fall 2017, Rutgers University Lecture 06 Kalman Filter Intro Instructor: Jingjin Yu

3 Outline Uncertainty Model of dynamical systems Bayesian filtering: the concept An illustrative example Applications of Kalman filters Derivation of Kalman Filter A 1D example

4 Uncertainty An every day experience Where exactly are we? E.g., in a classroom We do not know! Position is estimated We can get philosophical A more accurate model A dynamical system with state x (a function of time) Positions (i.e., x) are estimates Associate each location x with some probability This gives us a probability distribution P(x) For a robot, P(x) changes as the robot moves around Kalman filter (and other Bayesian filters) tracks P x as x changes over time

5 Modeling Dynamical Systems A dynamical system (e.g., a car) is often modeled as x ሶ = f(x, u) x: the state of the system, E.g., for a car, x = (x 1, x 2, θ) x ሶ = dx dt For a car, ሶ is the time derivative, i.e., the velocity of the system x = ( x 1 ሶ, x 2 ሶ, θ) ሶ u: the control input E.g., for a real car, u = (θ, v) θ is the front wheel bearing v is the forward speed (for a 2-wheel drive, assuming no slippery) u may be speed, acceleration, and so on f: system evolution function How do x, u determine xሶ In discrete settings, often written as x t = f(x t1, u t1 ) May view this as integration of the continuous model: x t = x t1 + t1 Often written as x k = f(x k1, u k1 ) t xdt ሶ

6 Modeling Dynamical Systems, Continued Examples A car going at fixed speed along x 1 -axis: x 1 ሶ = 1 In this case, f x, u = 1 is a constant An accelerating car along x 1 -axis with acceleration a: ሶ u = a, the acceleration, f(x, u) = at, does not depend on x x 1 = at A car going clockwise along the unit circle around the origin at unit speed (x 1, x 2 ) x ሶ = ( xሶ 1, xሶ 2, θ) ሶ = (x 2, x 1, 1) Initial condition: x 1 = 1, x 2 = 0 The car will keep circling the unit circle at unit speed So it takes 2π time to go one round Linear and non-linear systems Linear systems: f is a linear function, e.g., x ሶ = Ax + Bu Non-linear systems: f is non-linear What to grasp from the last two slides? Dynamical systems may be modeled as we have described In particular, given x k1, u k1, and f(x, u), we can predict x k (x 2, x 1 )

7 Kalman Filter as a Bayesian Filter Kalman filter is a type of Bayesian filters over a Hidden Markov model x k1 u k1 x k x k1 u k1 z k1 z k x k x i s are hidden (actual) system states They cannot be known exactly We can only observe x i using sensors to get z i The (discrete) process is modeled as a two-step iterative one Noisy state change: x k = f x k1, u k1 + ω k1 Noisy measurement after state change: z k = h x k + ν k More details coming up The data that we get are u 0, z 1, u 1, z 2, u 2, z 3, We want to provide x k Ƹ as an accurate estimate of x k Yields Kalman filters, particle filters, and so on white noises

8 An Example A hypothetical measurement of a variable x Mean: 0.5 Variance: sequential measurements Note: only the mean is shown in the second figure, not the variance Raw measurement Filtered result Important: Kalman filter is not simple averaging! It has (limited) predictive power

9 Applications Numerous applications GPS Minimizes error in tracking the position in altitude, latitude, longitude Reduces error from ~30 meters to less than 5 meters Aircraft autopilot Internal inertial guidance system generates errors over time Minimize with a 6D Kalman filter: yaw, pitch, roll, altitude, latitude, longitude Many, many other similar applications Missile guidance Radar Economic signals, e.g., stock time series Weather forecasting

10 Kalman Filter in More Detail Kalman filter is a minimum mean square estimator (MMSE) for estimating the state x R n of a discrete-time controlled process with a linear system equation and a linear observer under white noise. Linear stochastic system x k = Ax k1 + Bu k1 + ω k1, ω k1 N 0, Q (1) With a linear observer (sensor) z k = Hx k + ν k, ν k ~N 0, R (2) ω and ν are unknown but independent (i.e., Q and R are unknown) Kalman filter tries to provide estimate of true x k through the minimization of the estimation error based on (1) and 2 It does the minimization using the MMSE

11 Input and Output of a Kalman Filter A Kalman filter provides an estimate of x k under uncertainty This is an iterative process In each iteration, the input: x k1, P k1, u k1, z k,a, B, H x k1, P k1 : estimated system state/variance at time k 1; x 0, P 0 are guessed u k1 : system input at time k 1, e.g., how hard the gas pedal is pressed f(x k1, u k1 ) = Ax k1 + Bu k1, A and B are known z k =Hx k : the observation of x k, H is known The output: x k, P k x k, P k : estimated system state/variance at time k An illustration x k1 x k u k1 x k1 x k u k1 x k x k1 The Kalman filter computes a distribution The estimate is not a single value! For 1D, two values: mean + variance For dimension n, an n-vector and an n n covariance matrix Same applies to other variables z k

12 Deriving the MMSE (I) x k1 u k1 x k The goal is to find true state x k But recall this is not possible because x k is a hidden state A trick: x, P, A, B, u, z are high dimensional, but can treat as 1D In each iteration, a Kalman filter does two updates Time update: x k = A x k1 + Bu k1 Error of this step e k x k x k (a priori) Variance is P k = E[e k e k T ] Measurement update: x k = x k + K k z k H x k The term z k H x k is called the measurement innovation It gives us the new information from z k that is not already in x k Error of this step e k x k x k (a posteriori) Variance is P k = E[e k e T k ] Kalman filter seeks the best K k to minimize E[ e k 2 ] This is the same as minimizing the trace of P k e k x k x k = x k x k K k z k H x k z k1 z k

13 Deriving the MMSE (II) e k x k x k = x k x k K k z k H x k = x k x k K k z k Hx k + Hx k H x k = x k x k K k H x k x k K z k Hx k = I K k H x k x k K k z k Hx k = I K k H e k K k ν k x k1 z k1 u k1 x k z k P k = E e k e k T = I K k H E e k e k T I K k H T + E[K k ν k ν k T K k T ] = I K k H P K I K k H T + K k RK k T To minimize the trace of P k, take tr(p k) K k = 0 Yields K k = P k H T HP k H T +R, the optimal Kalman gain

14 Interpreting the Kalman Gain x k1 u k1 x k Recall the Kalman gain K k = P k H T z k1 HP k H T +R is used for computing x k z k x k = x k + K k z k H x k As R 0, measurement becomes more accurate, K k H 1 x k = x k + K k z k H x k H 1 z k As P k 0, state propagation becomes more accurate, K k 0 x k = x k + K k z k H x k x k

15 Deriving The Iterative Update Algorithm Recall - in each step, two updates Time update: x k = A x k1 + Bu k1 Error of this step e k x k x k (a priori) Variance is P k = E[e k e k T ] Measurement update: x k = x k + K k z k H x k Error of this step e k x k x k (a posteriori) Variance is P k = E[e k e k T ] For time update x k = A x k1 + Bu k1 e k x k x k = Ax k1 + Bu k1 + ω k1 x k = A x k1 x k1 + ω k1 = Ae k1 + ω k1 P k = E e k e k T T = AE e k1 e k1 = AP k1 A T + Q A T + Q For measurement update x k = x k + K k z k H x k Already computed P k H T K k = HP k H T +R P k = I K k H P K I K k H T + K k RK T k = I K k H P k

16 Tuning Parameters and Running the Filter We have the iterative update algorithm Time update x k = A x k1 + Bu k1 P k = AP k1 A T + Q To run the algorithm First estimate Q and R offline Starting from some estimate and then tuning This is known as system identification Then start filter with some initial x 0 and P 0 Usually P k and K k will quickly converge Measurement update x k = x k + K k z k H x k P k H T K k = HP k H T + R P k = I K k H P k

17 Example: Estimating a Random Constant Suppose we are measuring a random constant, e.g., temperature of a light bulb System: x k = x k1 + ω k1, x k, ω k R Observation: z k = x k + ν k, z k, ν k R Filter update equations Time update x k = x k1 P k = P k1 + Q Measurement update x k = x k + K k z k H x k K k = P k P k + R 1 P k = 1 K k P k Running the example If both (real, not our estimated) Q and R are large, it s hopeless If P 0 is small, it trusts x_0 a lot If x 0 is bad, then it takes long time to converge with small P 0