Fusion.cpp revision a01b4e237d57b74689576a3d486a2b2b903e74f4
1/* 2 * Copyright (C) 2011 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#include <stdio.h> 18 19#include <utils/Log.h> 20 21#include "Fusion.h" 22 23namespace android { 24 25// ----------------------------------------------------------------------- 26 27/* 28 * gyroVAR gives the measured variance of the gyro's output per 29 * Hz (or variance at 1 Hz). This is an "intrinsic" parameter of the gyro, 30 * which is independent of the sampling frequency. 31 * 32 * The variance of gyro's output at a given sampling period can be 33 * calculated as: 34 * variance(T) = gyroVAR / T 35 * 36 * The variance of the INTEGRATED OUTPUT at a given sampling period can be 37 * calculated as: 38 * variance_integrate_output(T) = gyroVAR * T 39 * 40 */ 41static const float gyroVAR = 1e-7; // (rad/s)^2 / Hz 42static const float biasVAR = 1e-8; // (rad/s)^2 / s (guessed) 43 44/* 45 * Standard deviations of accelerometer and magnetometer 46 */ 47static const float accSTDEV = 0.05f; // m/s^2 (measured 0.08 / CDD 0.05) 48static const float magSTDEV = 0.5f; // uT (measured 0.7 / CDD 0.5) 49 50static const float FREE_FALL_THRESHOLD = 0.981f; 51static const float SYMMETRY_TOLERANCE = 1e-10f; 52 53// ----------------------------------------------------------------------- 54 55template <typename TYPE, size_t C, size_t R> 56static mat<TYPE, R, R> scaleCovariance( 57 const mat<TYPE, C, R>& A, 58 const mat<TYPE, C, C>& P) { 59 // A*P*transpose(A); 60 mat<TYPE, R, R> APAt; 61 for (size_t r=0 ; r<R ; r++) { 62 for (size_t j=r ; j<R ; j++) { 63 double apat(0); 64 for (size_t c=0 ; c<C ; c++) { 65 double v(A[c][r]*P[c][c]*0.5); 66 for (size_t k=c+1 ; k<C ; k++) 67 v += A[k][r] * P[c][k]; 68 apat += 2 * v * A[c][j]; 69 } 70 APAt[j][r] = apat; 71 APAt[r][j] = apat; 72 } 73 } 74 return APAt; 75} 76 77template <typename TYPE, typename OTHER_TYPE> 78static mat<TYPE, 3, 3> crossMatrix(const vec<TYPE, 3>& p, OTHER_TYPE diag) { 79 mat<TYPE, 3, 3> r; 80 r[0][0] = diag; 81 r[1][1] = diag; 82 r[2][2] = diag; 83 r[0][1] = p.z; 84 r[1][0] =-p.z; 85 r[0][2] =-p.y; 86 r[2][0] = p.y; 87 r[1][2] = p.x; 88 r[2][1] =-p.x; 89 return r; 90} 91 92 93template<typename TYPE, size_t SIZE> 94class Covariance { 95 mat<TYPE, SIZE, SIZE> mSumXX; 96 vec<TYPE, SIZE> mSumX; 97 size_t mN; 98public: 99 Covariance() : mSumXX(0.0f), mSumX(0.0f), mN(0) { } 100 void update(const vec<TYPE, SIZE>& x) { 101 mSumXX += x*transpose(x); 102 mSumX += x; 103 mN++; 104 } 105 mat<TYPE, SIZE, SIZE> operator()() const { 106 const float N = 1.0f / mN; 107 return mSumXX*N - (mSumX*transpose(mSumX))*(N*N); 108 } 109 void reset() { 110 mN = 0; 111 mSumXX = 0; 112 mSumX = 0; 113 } 114 size_t getCount() const { 115 return mN; 116 } 117}; 118 119// ----------------------------------------------------------------------- 120 121Fusion::Fusion() { 122 Phi[0][1] = 0; 123 Phi[1][1] = 1; 124 125 Ba.x = 0; 126 Ba.y = 0; 127 Ba.z = 1; 128 129 Bm.x = 0; 130 Bm.y = 1; 131 Bm.z = 0; 132 133 init(); 134} 135 136void Fusion::init() { 137 mInitState = 0; 138 139 mGyroRate = 0; 140 141 mCount[0] = 0; 142 mCount[1] = 0; 143 mCount[2] = 0; 144 145 mData = 0; 146} 147 148void Fusion::initFusion(const vec4_t& q, float dT) 149{ 150 // initial estimate: E{ x(t0) } 151 x0 = q; 152 x1 = 0; 153 154 // process noise covariance matrix: G.Q.Gt, with 155 // 156 // G = | -1 0 | Q = | q00 q10 | 157 // | 0 1 | | q01 q11 | 158 // 159 // q00 = sv^2.dt + 1/3.su^2.dt^3 160 // q10 = q01 = 1/2.su^2.dt^2 161 // q11 = su^2.dt 162 // 163 164 // variance of integrated output at 1/dT Hz 165 // (random drift) 166 const float q00 = gyroVAR * dT; 167 168 // variance of drift rate ramp 169 const float q11 = biasVAR * dT; 170 171 const float u = q11 / dT; 172 const float q10 = 0.5f*u*dT*dT; 173 const float q01 = q10; 174 175 GQGt[0][0] = q00; // rad^2 176 GQGt[1][0] = -q10; 177 GQGt[0][1] = -q01; 178 GQGt[1][1] = q11; // (rad/s)^2 179 180 // initial covariance: Var{ x(t0) } 181 // TODO: initialize P correctly 182 P = 0; 183} 184 185bool Fusion::hasEstimate() const { 186 return (mInitState == (MAG|ACC|GYRO)); 187} 188 189bool Fusion::checkInitComplete(int what, const vec3_t& d, float dT) { 190 if (hasEstimate()) 191 return true; 192 193 if (what == ACC) { 194 mData[0] += d * (1/length(d)); 195 mCount[0]++; 196 mInitState |= ACC; 197 } else if (what == MAG) { 198 mData[1] += d * (1/length(d)); 199 mCount[1]++; 200 mInitState |= MAG; 201 } else if (what == GYRO) { 202 mGyroRate = dT; 203 mData[2] += d*dT; 204 mCount[2]++; 205 if (mCount[2] == 64) { 206 // 64 samples is good enough to estimate the gyro drift and 207 // doesn't take too much time. 208 mInitState |= GYRO; 209 } 210 } 211 212 if (mInitState == (MAG|ACC|GYRO)) { 213 // Average all the values we collected so far 214 mData[0] *= 1.0f/mCount[0]; 215 mData[1] *= 1.0f/mCount[1]; 216 mData[2] *= 1.0f/mCount[2]; 217 218 // calculate the MRPs from the data collection, this gives us 219 // a rough estimate of our initial state 220 mat33_t R; 221 vec3_t up(mData[0]); 222 vec3_t east(cross_product(mData[1], up)); 223 east *= 1/length(east); 224 vec3_t north(cross_product(up, east)); 225 R << east << north << up; 226 const vec4_t q = matrixToQuat(R); 227 228 initFusion(q, mGyroRate); 229 } 230 231 return false; 232} 233 234void Fusion::handleGyro(const vec3_t& w, float dT) { 235 if (!checkInitComplete(GYRO, w, dT)) 236 return; 237 238 predict(w, dT); 239} 240 241status_t Fusion::handleAcc(const vec3_t& a) { 242 // ignore acceleration data if we're close to free-fall 243 if (length(a) < FREE_FALL_THRESHOLD) 244 return BAD_VALUE; 245 246 if (!checkInitComplete(ACC, a)) 247 return BAD_VALUE; 248 249 const float l = 1/length(a); 250 update(a*l, Ba, accSTDEV*l); 251 return NO_ERROR; 252} 253 254status_t Fusion::handleMag(const vec3_t& m) { 255 // the geomagnetic-field should be between 30uT and 60uT 256 // reject obviously wrong magnetic-fields 257 if (length(m) > 100) 258 return BAD_VALUE; 259 260 if (!checkInitComplete(MAG, m)) 261 return BAD_VALUE; 262 263 const vec3_t up( getRotationMatrix() * Ba ); 264 const vec3_t east( cross_product(m, up) ); 265 vec3_t north( cross_product(up, east) ); 266 267 const float l = 1 / length(north); 268 north *= l; 269 270 update(north, Bm, magSTDEV*l); 271 return NO_ERROR; 272} 273 274void Fusion::checkState() { 275 // P needs to stay positive semidefinite or the fusion diverges. When we 276 // detect divergence, we reset the fusion. 277 // TODO(braun): Instead, find the reason for the divergence and fix it. 278 279 if (!isPositiveSemidefinite(P[0][0], SYMMETRY_TOLERANCE) || 280 !isPositiveSemidefinite(P[1][1], SYMMETRY_TOLERANCE)) { 281 LOGW("Sensor fusion diverged; resetting state."); 282 P = 0; 283 } 284} 285 286vec4_t Fusion::getAttitude() const { 287 return x0; 288} 289 290vec3_t Fusion::getBias() const { 291 return x1; 292} 293 294mat33_t Fusion::getRotationMatrix() const { 295 return quatToMatrix(x0); 296} 297 298mat34_t Fusion::getF(const vec4_t& q) { 299 mat34_t F; 300 F[0].x = q.w; F[1].x =-q.z; F[2].x = q.y; 301 F[0].y = q.z; F[1].y = q.w; F[2].y =-q.x; 302 F[0].z =-q.y; F[1].z = q.x; F[2].z = q.w; 303 F[0].w =-q.x; F[1].w =-q.y; F[2].w =-q.z; 304 return F; 305} 306 307void Fusion::predict(const vec3_t& w, float dT) { 308 const vec4_t q = x0; 309 const vec3_t b = x1; 310 const vec3_t we = w - b; 311 const vec4_t dq = getF(q)*((0.5f*dT)*we); 312 x0 = normalize_quat(q + dq); 313 314 // P(k+1) = F*P(k)*Ft + G*Q*Gt 315 316 // Phi = | Phi00 Phi10 | 317 // | 0 1 | 318 const mat33_t I33(1); 319 const mat33_t I33dT(dT); 320 const mat33_t wx(crossMatrix(we, 0)); 321 const mat33_t wx2(wx*wx); 322 const float lwedT = length(we)*dT; 323 const float ilwe = 1/length(we); 324 const float k0 = (1-cosf(lwedT))*(ilwe*ilwe); 325 const float k1 = sinf(lwedT); 326 327 Phi[0][0] = I33 - wx*(k1*ilwe) + wx2*k0; 328 Phi[1][0] = wx*k0 - I33dT - wx2*(ilwe*ilwe*ilwe)*(lwedT-k1); 329 330 P = Phi*P*transpose(Phi) + GQGt; 331 332 checkState(); 333} 334 335void Fusion::update(const vec3_t& z, const vec3_t& Bi, float sigma) { 336 vec4_t q(x0); 337 // measured vector in body space: h(p) = A(p)*Bi 338 const mat33_t A(quatToMatrix(q)); 339 const vec3_t Bb(A*Bi); 340 341 // Sensitivity matrix H = dh(p)/dp 342 // H = [ L 0 ] 343 const mat33_t L(crossMatrix(Bb, 0)); 344 345 // gain... 346 // K = P*Ht / [H*P*Ht + R] 347 vec<mat33_t, 2> K; 348 const mat33_t R(sigma*sigma); 349 const mat33_t S(scaleCovariance(L, P[0][0]) + R); 350 const mat33_t Si(invert(S)); 351 const mat33_t LtSi(transpose(L)*Si); 352 K[0] = P[0][0] * LtSi; 353 K[1] = transpose(P[1][0])*LtSi; 354 355 // update... 356 // P -= K*H*P; 357 const mat33_t K0L(K[0] * L); 358 const mat33_t K1L(K[1] * L); 359 P[0][0] -= K0L*P[0][0]; 360 P[1][1] -= K1L*P[1][0]; 361 P[1][0] -= K0L*P[1][0]; 362 P[0][1] = transpose(P[1][0]); 363 364 const vec3_t e(z - Bb); 365 const vec3_t dq(K[0]*e); 366 const vec3_t db(K[1]*e); 367 368 q += getF(q)*(0.5f*dq); 369 x0 = normalize_quat(q); 370 x1 += db; 371 372 checkState(); 373} 374 375// ----------------------------------------------------------------------- 376 377}; // namespace android 378 379