#include "calibration/common/sphere_fit_calibration.h" #include #include #include #include #include "calibration/util/cal_log.h" #include "common/math/mat.h" #include "common/math/vec.h" // FORWARD DECLARATIONS /////////////////////////////////////////////////////////////////////////////// // Utility for converting solver state to a calibration data structure. static void convertStateToCalStruct(const float x[SF_STATE_DIM], struct ThreeAxisCalData *calstruct); static bool runCalibration(struct SphereFitCal *sphere_cal, const struct SphereFitData *data, uint64_t timestamp_nanos); #define MIN_VALID_DATA_NORM (1e-4) // FUNCTION IMPLEMENTATIONS ////////////////////////////////////////////////////////////////////////////// void sphereFitInit(struct SphereFitCal *sphere_cal, const struct LmParams *lm_params, const size_t min_num_points_for_cal) { ASSERT_NOT_NULL(sphere_cal); ASSERT_NOT_NULL(lm_params); // Initialize LM solver. lmSolverInit(&sphere_cal->lm_solver, lm_params, &sphereFitResidAndJacobianFunc); // Reset other parameters. sphereFitReset(sphere_cal); // Set num points for calibration, checking that it is above min. if (min_num_points_for_cal < MIN_NUM_SPHERE_FIT_POINTS) { sphere_cal->min_points_for_cal = MIN_NUM_SPHERE_FIT_POINTS; } else { sphere_cal->min_points_for_cal = min_num_points_for_cal; } } void sphereFitReset(struct SphereFitCal *sphere_cal) { ASSERT_NOT_NULL(sphere_cal); // Set state to default (diagonal scale matrix and zero offset). memset(&sphere_cal->x0[0], 0, sizeof(float) * SF_STATE_DIM); sphere_cal->x0[eParamScaleMatrix11] = 1.f; sphere_cal->x0[eParamScaleMatrix22] = 1.f; sphere_cal->x0[eParamScaleMatrix33] = 1.f; memcpy(sphere_cal->x, sphere_cal->x0, sizeof(sphere_cal->x)); // Reset time. sphere_cal->estimate_time_nanos = 0; } void sphereFitSetSolverData(struct SphereFitCal *sphere_cal, struct LmData *lm_data) { ASSERT_NOT_NULL(sphere_cal); ASSERT_NOT_NULL(lm_data); // Set solver data. lmSolverSetData(&sphere_cal->lm_solver, lm_data); } bool sphereFitRunCal(struct SphereFitCal *sphere_cal, const struct SphereFitData *data, uint64_t timestamp_nanos) { ASSERT_NOT_NULL(sphere_cal); ASSERT_NOT_NULL(data); // Run calibration if have enough points. if (data->num_fit_points >= sphere_cal->min_points_for_cal) { return runCalibration(sphere_cal, data, timestamp_nanos); } return false; } void sphereFitSetInitialBias(struct SphereFitCal *sphere_cal, const float initial_bias[THREE_AXIS_DIM]) { ASSERT_NOT_NULL(sphere_cal); sphere_cal->x0[eParamOffset1] = initial_bias[0]; sphere_cal->x0[eParamOffset2] = initial_bias[1]; sphere_cal->x0[eParamOffset3] = initial_bias[2]; } void sphereFitGetLatestCal(const struct SphereFitCal *sphere_cal, struct ThreeAxisCalData *cal_data) { ASSERT_NOT_NULL(sphere_cal); ASSERT_NOT_NULL(cal_data); convertStateToCalStruct(sphere_cal->x, cal_data); cal_data->calibration_time_nanos = sphere_cal->estimate_time_nanos; } void sphereFitResidAndJacobianFunc(const float *state, const void *f_data, float *residual, float *jacobian) { ASSERT_NOT_NULL(state); ASSERT_NOT_NULL(f_data); ASSERT_NOT_NULL(residual); const struct SphereFitData *data = (const struct SphereFitData*)f_data; // Verify that expected norm is non-zero, else use default of 1.0. float expected_norm = 1.0; ASSERT(data->expected_norm > MIN_VALID_DATA_NORM); if (data->expected_norm > MIN_VALID_DATA_NORM) { expected_norm = data->expected_norm; } // Convert parameters to calibration structure. struct ThreeAxisCalData calstruct; convertStateToCalStruct(state, &calstruct); // Compute Jacobian helper matrix if Jacobian requested. // // J = d(||M(x_data - bias)|| - expected_norm)/dstate // = (M(x_data - bias) / ||M(x_data - bias)||) * d(M(x_data - bias))/dstate // = x_corr / ||x_corr|| * A // A = d(M(x_data - bias))/dstate // = [dy/dM11, dy/dM21, dy/dM22, dy/dM31, dy/dM32, dy/dM3,... // dy/db1, dy/db2, dy/db3]' // where: // dy/dM11 = [x_data[0] - bias[0], 0, 0] // dy/dM21 = [0, x_data[0] - bias[0], 0] // dy/dM22 = [0, x_data[1] - bias[1], 0] // dy/dM31 = [0, 0, x_data[0] - bias[0]] // dy/dM32 = [0, 0, x_data[1] - bias[1]] // dy/dM33 = [0, 0, x_data[2] - bias[2]] // dy/db1 = [-scale_factor_x, 0, 0] // dy/db2 = [0, -scale_factor_y, 0] // dy/db3 = [0, 0, -scale_factor_z] float A[SF_STATE_DIM * THREE_AXIS_DIM]; if (jacobian) { memset(jacobian, 0, sizeof(float) * SF_STATE_DIM * data->num_fit_points); memset(A, 0, sizeof(A)); A[0 * SF_STATE_DIM + eParamOffset1] = -calstruct.scale_factor_x; A[1 * SF_STATE_DIM + eParamOffset2] = -calstruct.scale_factor_y; A[2 * SF_STATE_DIM + eParamOffset3] = -calstruct.scale_factor_z; } // Loop over all data points to compute residual and Jacobian. // TODO(dvitus): Use fit_data_std when available to weight residuals. float x_corr[THREE_AXIS_DIM]; float x_bias_corr[THREE_AXIS_DIM]; size_t i; for (i = 0; i < data->num_fit_points; ++i) { const float *x_data = &data->fit_data[i * THREE_AXIS_DIM]; // Compute corrected sensor data calDataCorrectData(&calstruct, x_data, x_corr); // Compute norm of x_corr. const float norm = vecNorm(x_corr, THREE_AXIS_DIM); // Compute residual error: f_x = norm - exp_norm residual[i] = norm - data->expected_norm; // Compute Jacobian if valid pointer. if (jacobian) { if (norm < MIN_VALID_DATA_NORM) { return; } const float scale = 1.f / norm; // Compute bias corrected data. vecSub(x_bias_corr, x_data, calstruct.bias, THREE_AXIS_DIM); // Populate non-bias terms for A A[0 + eParamScaleMatrix11] = x_bias_corr[0]; A[1 * SF_STATE_DIM + eParamScaleMatrix21] = x_bias_corr[0]; A[1 * SF_STATE_DIM + eParamScaleMatrix22] = x_bias_corr[1]; A[2 * SF_STATE_DIM + eParamScaleMatrix31] = x_bias_corr[0]; A[2 * SF_STATE_DIM + eParamScaleMatrix32] = x_bias_corr[1]; A[2 * SF_STATE_DIM + eParamScaleMatrix33] = x_bias_corr[2]; // Compute J = x_corr / ||x_corr|| * A matTransposeMultiplyVec(&jacobian[i * SF_STATE_DIM], A, x_corr, THREE_AXIS_DIM, SF_STATE_DIM); vecScalarMulInPlace(&jacobian[i * SF_STATE_DIM], scale, SF_STATE_DIM); } } } void convertStateToCalStruct(const float x[SF_STATE_DIM], struct ThreeAxisCalData *calstruct) { memcpy(&calstruct->bias[0], &x[eParamOffset1], sizeof(float) * THREE_AXIS_DIM); calstruct->scale_factor_x = x[eParamScaleMatrix11]; calstruct->skew_yx = x[eParamScaleMatrix21]; calstruct->scale_factor_y = x[eParamScaleMatrix22]; calstruct->skew_zx = x[eParamScaleMatrix31]; calstruct->skew_zy = x[eParamScaleMatrix32]; calstruct->scale_factor_z = x[eParamScaleMatrix33]; } bool runCalibration(struct SphereFitCal *sphere_cal, const struct SphereFitData *data, uint64_t timestamp_nanos) { float x_sol[SF_STATE_DIM]; // Run calibration const enum LmStatus status = lmSolverSolve(&sphere_cal->lm_solver, sphere_cal->x0, (void *)data, SF_STATE_DIM, data->num_fit_points, x_sol); // Check if solver was successful if (status == RELATIVE_STEP_SUFFICIENTLY_SMALL || status == GRADIENT_SUFFICIENTLY_SMALL) { // TODO(dvitus): Check quality of new fit before using. // Store new fit. #ifdef SPHERE_FIT_DBG_ENABLED CAL_DEBUG_LOG( "[SPHERE CAL]", "Solution found in %d iterations with status %d.\n", sphere_cal->lm_solver.num_iter, status); CAL_DEBUG_LOG( "[SPHERE CAL]", "Solution:\n {%s%d.%06d [M(1,1)], %s%d.%06d [M(2,1)], " "%s%d.%06d [M(2,2)], \n" "%s%d.%06d [M(3,1)], %s%d.%06d [M(3,2)], %s%d.%06d [M(3,3)], \n" "%s%d.%06d [b(1)], %s%d.%06d [b(2)], %s%d.%06d [b(3)]}.", CAL_ENCODE_FLOAT(x_sol[0], 6), CAL_ENCODE_FLOAT(x_sol[1], 6), CAL_ENCODE_FLOAT(x_sol[2], 6), CAL_ENCODE_FLOAT(x_sol[3], 6), CAL_ENCODE_FLOAT(x_sol[4], 6), CAL_ENCODE_FLOAT(x_sol[5], 6), CAL_ENCODE_FLOAT(x_sol[6], 6), CAL_ENCODE_FLOAT(x_sol[7], 6), CAL_ENCODE_FLOAT(x_sol[8], 6)); #endif memcpy(sphere_cal->x, x_sol, sizeof(x_sol)); sphere_cal->estimate_time_nanos = timestamp_nanos; return true; } else { #ifdef SPHERE_FIT_DBG_ENABLED CAL_DEBUG_LOG( "[SPHERE CAL]", "Solution failed in %d iterations with status %d.\n", sphere_cal->lm_solver.num_iter, status); #endif } return false; }