librobotcontrol-sys 0.4.0

/**
 * <rc/math/kalman.h>
 *
 * @brief      Kalman filter implementation
 *
 * This may be used to implement a discrete time linear or extended kalman
 * filter.
 *
 * For the linear case, initialize the filter with rc_kalman_alloc_lin() which
 * takes in the linear state matrices. These are then saved and used by
 * rc_kalman_update_lin to calculate the predicted state x_pre and predicted
 * sensor measurements h internally each step.
 *
 *
 * Basic loop structure for Linear case:
 *
 * ```C
 * rc_kalman_t kf = rc_kalman_empty();
 * rc_kalman_alloc_lin(&kf,F,G,H,Q,R,Pi);
 * while(running){
 *      measure sensors, calculate y;
 *      rc_kalman_update_lin(&kf, u, y);
 *      calculate new actuator control output u;
 *      save u for next loop;
 * }
 * rc_kalman_free(&kf);
 * return;
 * ```
 *
 * For the non-linear EKF case, the user should initialize the filter with
 * rc_kalman_alloc_ekf() which does NOT take in the state transition matrices
 * F,G,H. Instead it's up to the user to calculate the new predicted state x_pre
 * and predicted sensor measurement h for their own non-linear system model each
 * step. Those user-calculated predictions are then given to the non-linear
 * rc_kalman_update_ekf() function.
 *
 *
 * Basic loop structure for non-linear EKF case:
 *
 * ```C
 * rc_kalman_t kf = rc_kalman_empty();
 * rc_kalman_alloc_ekf(&kf,Q,R,Pi);
 * while(running){
 *      measure sensors, calculate y
 *      calculate new Jacobian F given x_est from previous loop;
 *      calculate new predicted x_pre using x_est from previous
 *      loop;
 *      calulate new predicted sensor readings h from x_pre above;
 *      calculate new Jacobian H given x_pre;
 *      rc_kalman_update_ekf(&kf, F, x_pre, H, y, h);
 *      calculate new actuator control output u;
 *      save u for next loop;
 * }
 * rc_kalman_free(&kf);
 * return;
 * ```
 *
 * @date       April 2018
 * @author     Eric Nauli Sihite & James Strawson
 *
 * @addtogroup Kalman
 * @ingroup    Math
 * @{
 */


#ifndef RC_KALMAN_H
#define RC_KALMAN_H

#ifdef __cplusplus
extern "C" {
#endif

#include <stdint.h>
#include <rc/math/vector.h>
#include <rc/math/matrix.h>


/*
 * @brief      Struct to contain full state of kalman filter
 */
typedef struct rc_kalman_t {
	/** @name State Transition Matrices for linear filter only */
	///@{
	rc_matrix_t F;		///< undriven state-transition model
	rc_matrix_t G;		///< control input model
	rc_matrix_t H;		///< observation-model
	///@}

	/** @name Covariance Matrices */
	///@{
	rc_matrix_t Q;		///< Process noise covariance set by user
	rc_matrix_t R;		///< Measurement noise covariance set by user
	rc_matrix_t P;		///< Predicted state error covariance calculated by the update functions
	rc_matrix_t Pi;		///< Initial P matrix set by user
	///@}

	/** @name State estimates */
	///@{
	rc_vector_t x_est;	///< Estimated state x[k|k]   = x[k|k-1],y[k])
	rc_vector_t x_pre;	///< Predicted state x[k|k-1] = f(x[k-1],u[k])
	///@}

	/** @name other */
	///@{
	int initialized;	///< set to 1 once initialized with rc_kalman_alloc
	uint64_t step;		///< counts times rc_kalman_measurement_update has been called
	///@}
} rc_kalman_t;

#define RC_KALMAN_INITIALIZER {\
	.F = RC_MATRIX_INITIALIZER,\
	.G = RC_MATRIX_INITIALIZER,\
	.H = RC_MATRIX_INITIALIZER,\
	.Q = RC_MATRIX_INITIALIZER,\
	.R = RC_MATRIX_INITIALIZER,\
	.P = RC_MATRIX_INITIALIZER,\
	.Pi = RC_MATRIX_INITIALIZER,\
	.x_est = RC_VECTOR_INITIALIZER,\
	.x_pre = RC_VECTOR_INITIALIZER,\
	.initialized = 0,\
	.step = 0}

/**
 * @brief      Critical function for initializing rc_kalman_t structs
 *
 * This is a very important function. If your rc_kalman_t struct is not a global
 * variable, then its initial contents cannot be guaranteed to be anything in
 * particular. Therefore it could contain problematic contents which could
 * interfere with functions in this library. Therefore, you should always
 * initialize your filters with rc_kalman_empty() before using with any other
 * function in this library such as rc_kalman_alloc. This serves the same
 * purpose as rc_matrix_empty, rc_vector_empty, rc_filter_empty, and
 * rc_ringbuf_empty.
 *
 * @return     Empty zero-filled rc_kalman_t struct
 */
rc_kalman_t rc_kalman_empty(void);

/**
 * @brief      Allocates memory for a Kalman filter of given dimensions
 *
 * @param      kf    pointer to struct to be allocated
 * @param[in]  F     undriven state-transition model
 * @param[in]  G     control input model
 * @param[in]  H     observation model
 * @param[in]  Q     Process noise covariance, can be updated later
 * @param[in]  R     Measurement noise covariance, can be updated later
 * @param[in]  Pi    Initial P matrix
 *
 * @return     0 on success, -1 on failure
 */
int rc_kalman_alloc_lin(rc_kalman_t* kf, rc_matrix_t F, rc_matrix_t G, rc_matrix_t H, rc_matrix_t Q, rc_matrix_t R, rc_matrix_t Pi);

/**
 * @brief      Allocates memory for a Kalman filter of given dimensions
 *
 * @param      kf    pointer to struct to be allocated
 * @param[in]  Q     Process noise covariance, can be updated later
 * @param[in]  R     Measurement noise covariance, can be updated later
 * @param[in]  Pi    Initial P matrix
 *
 * @return     0 on success, -1 on failure
 */
int rc_kalman_alloc_ekf(rc_kalman_t* kf, rc_matrix_t Q, rc_matrix_t R, rc_matrix_t Pi);


/**
 * @brief      Frees the memory allocated by a kalman filter's matrices and
 *             vectors. Also resets all values to 0 like rc_kalman_empty().
 *
 * @param      kf    pointer to user's rc_kalman_t struct
 *
 * @return     0 on success or -1 on failure.
 */
int rc_kalman_free(rc_kalman_t* kf);


/**
 * @brief      reset state and dynamics of the filter to 0
 *
 * Q and R are constant, and therefore not reset. P is set to identity*P_init.
 *
 * @param      kf    pointer to struct to be reset
 *
 * @return     0 on success, -1 on failure
 */
int rc_kalman_reset(rc_kalman_t* kf);


/**
 * @brief      Kalman Filter state prediction step based on physical model.
 *
 * Uses the state estimate and control input from the previous timestep to
 * produce an estimate of the state at the current timestep. This step pdates P
 * and the estimated state x. Assume that you have calculated f(x[k|k],u[k]) and
 * F(x[k|k],u[k]) before calling this function.
 *
 * - Kalman linear state prediction
 *   - x_pre[k|k-1] = F*x[k-1|k-1] +  G*u[k-1]
 *   - P[k|k-1] = F*P[k-1|k-1]*F^T + Q
 * - Kalman measurement Update:
 *   - h[k] = H * x_pre[k]
 *   - S = H*P*H^T + R
 *   - L = P*(H^T)*(S^-1)
 *   - x_est[k|k] = x[k|k-1] + L*(y[k]-h[k])
 *   - P[k|k] = (I - L*H)*P[k|k-1]
 *
 * @param      kf    pointer to struct to be updated
 * @param      u     control input
 * @param[in]  y     sensor measurement
 *
 * @return     0 on success, -1 on failure
 */
int rc_kalman_update_lin(rc_kalman_t* kf, rc_vector_t u, rc_vector_t y);


/**
 * @brief      Kalman Filter measurement update step.
 *
 * Updates L, P, & x_est. Assumes that you have done the non-linear prediction
 * step in your own function which should calculate the Jacobians F(x[k|k-1]) &
 * H(x[k|k-1]), the predicted sensor value h(x[k|k-1]), and of course the
 * predicted state x_pre[k|k-1]
 *
 * -Kalman measurement Update:
 * - P[k|k-1] = F*P[k-1|k-1]*F^T + Q
 * - S = H*P*H^T + R
 * - L = P*(H^T)*(S^-1)
 * - x[k|k] = x[k|k-1] + L*y
 * - P[k|k] = (I - L*H)*P
 *
 * Also updates the step counter in the rc_kalman_t struct
 *
 * @param      kf     pointer to struct to be updated
 * @param[in]  F      Jacobian of state transition matrix linearized at x_pre
 * @param[in]  H      Jacobian of observation matrix linearized at x_pre
 * @param[in]  x_pre  predicted state
 * @param[in]  y      new sensor data
 * @param[in]  h      Ideal estimate of y, usually h=H*x_pre.
 *
 * @return     0 on success, -1 on failure
 */
int rc_kalman_update_ekf(rc_kalman_t* kf, rc_matrix_t F, rc_matrix_t H, rc_vector_t x_pre,  rc_vector_t y, rc_vector_t h);


#ifdef __cplusplus
}
#endif

#endif // RC_KALMAN_H

/** @}  end group math*/