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
23 namespace 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 */
41 static const float gyroVAR = 1e-7; // (rad/s)^2 / Hz
42 static const float biasVAR = 1e-8; // (rad/s)^2 / s (guessed)
43
44 /*
45 * Standard deviations of accelerometer and magnetometer
46 */
47 static const float accSTDEV = 0.05f; // m/s^2 (measured 0.08 / CDD 0.05)
48 static const float magSTDEV = 0.5f; // uT (measured 0.7 / CDD 0.5)
49
50 static const float SYMMETRY_TOLERANCE = 1e-10f;
51
52 /*
53 * Accelerometer updates will not be performed near free fall to avoid
54 * ill-conditioning and div by zeros.
55 * Threshhold: 10% of g, in m/s^2
56 */
57 static const float FREE_FALL_THRESHOLD = 0.981f;
58 static const float FREE_FALL_THRESHOLD_SQ =
59 FREE_FALL_THRESHOLD*FREE_FALL_THRESHOLD;
60
61 /*
62 * The geomagnetic-field should be between 30uT and 60uT.
63 * Fields strengths greater than this likely indicate a local magnetic
64 * disturbance which we do not want to update into the fused frame.
65 */
66 static const float MAX_VALID_MAGNETIC_FIELD = 100; // uT
67 static const float MAX_VALID_MAGNETIC_FIELD_SQ =
68 MAX_VALID_MAGNETIC_FIELD*MAX_VALID_MAGNETIC_FIELD;
69
70 /*
71 * Values of the field smaller than this should be ignored in fusion to avoid
72 * ill-conditioning. This state can happen with anomalous local magnetic
73 * disturbances canceling the Earth field.
74 */
75 static const float MIN_VALID_MAGNETIC_FIELD = 10; // uT
76 static const float MIN_VALID_MAGNETIC_FIELD_SQ =
77 MIN_VALID_MAGNETIC_FIELD*MIN_VALID_MAGNETIC_FIELD;
78
79 /*
80 * If the cross product of two vectors has magnitude squared less than this,
81 * we reject it as invalid due to alignment of the vectors.
82 * This threshold is used to check for the case where the magnetic field sample
83 * is parallel to the gravity field, which can happen in certain places due
84 * to magnetic field disturbances.
85 */
86 static const float MIN_VALID_CROSS_PRODUCT_MAG = 1.0e-3;
87 static const float MIN_VALID_CROSS_PRODUCT_MAG_SQ =
88 MIN_VALID_CROSS_PRODUCT_MAG*MIN_VALID_CROSS_PRODUCT_MAG;
89
90 // -----------------------------------------------------------------------
91
92 template <typename TYPE, size_t C, size_t R>
scaleCovariance(const mat<TYPE,C,R> & A,const mat<TYPE,C,C> & P)93 static mat<TYPE, R, R> scaleCovariance(
94 const mat<TYPE, C, R>& A,
95 const mat<TYPE, C, C>& P) {
96 // A*P*transpose(A);
97 mat<TYPE, R, R> APAt;
98 for (size_t r=0 ; r<R ; r++) {
99 for (size_t j=r ; j<R ; j++) {
100 double apat(0);
101 for (size_t c=0 ; c<C ; c++) {
102 double v(A[c][r]*P[c][c]*0.5);
103 for (size_t k=c+1 ; k<C ; k++)
104 v += A[k][r] * P[c][k];
105 apat += 2 * v * A[c][j];
106 }
107 APAt[j][r] = apat;
108 APAt[r][j] = apat;
109 }
110 }
111 return APAt;
112 }
113
114 template <typename TYPE, typename OTHER_TYPE>
crossMatrix(const vec<TYPE,3> & p,OTHER_TYPE diag)115 static mat<TYPE, 3, 3> crossMatrix(const vec<TYPE, 3>& p, OTHER_TYPE diag) {
116 mat<TYPE, 3, 3> r;
117 r[0][0] = diag;
118 r[1][1] = diag;
119 r[2][2] = diag;
120 r[0][1] = p.z;
121 r[1][0] =-p.z;
122 r[0][2] =-p.y;
123 r[2][0] = p.y;
124 r[1][2] = p.x;
125 r[2][1] =-p.x;
126 return r;
127 }
128
129
130 template<typename TYPE, size_t SIZE>
131 class Covariance {
132 mat<TYPE, SIZE, SIZE> mSumXX;
133 vec<TYPE, SIZE> mSumX;
134 size_t mN;
135 public:
Covariance()136 Covariance() : mSumXX(0.0f), mSumX(0.0f), mN(0) { }
update(const vec<TYPE,SIZE> & x)137 void update(const vec<TYPE, SIZE>& x) {
138 mSumXX += x*transpose(x);
139 mSumX += x;
140 mN++;
141 }
operator ()() const142 mat<TYPE, SIZE, SIZE> operator()() const {
143 const float N = 1.0f / mN;
144 return mSumXX*N - (mSumX*transpose(mSumX))*(N*N);
145 }
reset()146 void reset() {
147 mN = 0;
148 mSumXX = 0;
149 mSumX = 0;
150 }
getCount() const151 size_t getCount() const {
152 return mN;
153 }
154 };
155
156 // -----------------------------------------------------------------------
157
Fusion()158 Fusion::Fusion() {
159 Phi[0][1] = 0;
160 Phi[1][1] = 1;
161
162 Ba.x = 0;
163 Ba.y = 0;
164 Ba.z = 1;
165
166 Bm.x = 0;
167 Bm.y = 1;
168 Bm.z = 0;
169
170 x0 = 0;
171 x1 = 0;
172
173 init();
174 }
175
init()176 void Fusion::init() {
177 mInitState = 0;
178
179 mGyroRate = 0;
180
181 mCount[0] = 0;
182 mCount[1] = 0;
183 mCount[2] = 0;
184
185 mData = 0;
186 }
187
initFusion(const vec4_t & q,float dT)188 void Fusion::initFusion(const vec4_t& q, float dT)
189 {
190 // initial estimate: E{ x(t0) }
191 x0 = q;
192 x1 = 0;
193
194 // process noise covariance matrix: G.Q.Gt, with
195 //
196 // G = | -1 0 | Q = | q00 q10 |
197 // | 0 1 | | q01 q11 |
198 //
199 // q00 = sv^2.dt + 1/3.su^2.dt^3
200 // q10 = q01 = 1/2.su^2.dt^2
201 // q11 = su^2.dt
202 //
203
204 const float dT2 = dT*dT;
205 const float dT3 = dT2*dT;
206
207 // variance of integrated output at 1/dT Hz (random drift)
208 const float q00 = gyroVAR * dT + 0.33333f * biasVAR * dT3;
209
210 // variance of drift rate ramp
211 const float q11 = biasVAR * dT;
212 const float q10 = 0.5f * biasVAR * dT2;
213 const float q01 = q10;
214
215 GQGt[0][0] = q00; // rad^2
216 GQGt[1][0] = -q10;
217 GQGt[0][1] = -q01;
218 GQGt[1][1] = q11; // (rad/s)^2
219
220 // initial covariance: Var{ x(t0) }
221 // TODO: initialize P correctly
222 P = 0;
223 }
224
hasEstimate() const225 bool Fusion::hasEstimate() const {
226 return (mInitState == (MAG|ACC|GYRO));
227 }
228
checkInitComplete(int what,const vec3_t & d,float dT)229 bool Fusion::checkInitComplete(int what, const vec3_t& d, float dT) {
230 if (hasEstimate())
231 return true;
232
233 if (what == ACC) {
234 mData[0] += d * (1/length(d));
235 mCount[0]++;
236 mInitState |= ACC;
237 } else if (what == MAG) {
238 mData[1] += d * (1/length(d));
239 mCount[1]++;
240 mInitState |= MAG;
241 } else if (what == GYRO) {
242 mGyroRate = dT;
243 mData[2] += d*dT;
244 mCount[2]++;
245 if (mCount[2] == 64) {
246 // 64 samples is good enough to estimate the gyro drift and
247 // doesn't take too much time.
248 mInitState |= GYRO;
249 }
250 }
251
252 if (mInitState == (MAG|ACC|GYRO)) {
253 // Average all the values we collected so far
254 mData[0] *= 1.0f/mCount[0];
255 mData[1] *= 1.0f/mCount[1];
256 mData[2] *= 1.0f/mCount[2];
257
258 // calculate the MRPs from the data collection, this gives us
259 // a rough estimate of our initial state
260 mat33_t R;
261 vec3_t up(mData[0]);
262 vec3_t east(cross_product(mData[1], up));
263 east *= 1/length(east);
264 vec3_t north(cross_product(up, east));
265 R << east << north << up;
266 const vec4_t q = matrixToQuat(R);
267
268 initFusion(q, mGyroRate);
269 }
270
271 return false;
272 }
273
handleGyro(const vec3_t & w,float dT)274 void Fusion::handleGyro(const vec3_t& w, float dT) {
275 if (!checkInitComplete(GYRO, w, dT))
276 return;
277
278 predict(w, dT);
279 }
280
handleAcc(const vec3_t & a)281 status_t Fusion::handleAcc(const vec3_t& a) {
282 // ignore acceleration data if we're close to free-fall
283 if (length_squared(a) < FREE_FALL_THRESHOLD_SQ) {
284 return BAD_VALUE;
285 }
286
287 if (!checkInitComplete(ACC, a))
288 return BAD_VALUE;
289
290 const float l = 1/length(a);
291 update(a*l, Ba, accSTDEV*l);
292 return NO_ERROR;
293 }
294
handleMag(const vec3_t & m)295 status_t Fusion::handleMag(const vec3_t& m) {
296 // the geomagnetic-field should be between 30uT and 60uT
297 // reject if too large to avoid spurious magnetic sources
298 const float magFieldSq = length_squared(m);
299 if (magFieldSq > MAX_VALID_MAGNETIC_FIELD_SQ) {
300 return BAD_VALUE;
301 } else if (magFieldSq < MIN_VALID_MAGNETIC_FIELD_SQ) {
302 // Also reject if too small since we will get ill-defined (zero mag)
303 // cross-products below
304 return BAD_VALUE;
305 }
306
307 if (!checkInitComplete(MAG, m))
308 return BAD_VALUE;
309
310 // Orthogonalize the magnetic field to the gravity field, mapping it into
311 // tangent to Earth.
312 const vec3_t up( getRotationMatrix() * Ba );
313 const vec3_t east( cross_product(m, up) );
314
315 // If the m and up vectors align, the cross product magnitude will
316 // approach 0.
317 // Reject this case as well to avoid div by zero problems and
318 // ill-conditioning below.
319 if (length_squared(east) < MIN_VALID_CROSS_PRODUCT_MAG_SQ) {
320 return BAD_VALUE;
321 }
322
323 // If we have created an orthogonal magnetic field successfully,
324 // then pass it in as the update.
325 vec3_t north( cross_product(up, east) );
326
327 const float l = 1 / length(north);
328 north *= l;
329
330 update(north, Bm, magSTDEV*l);
331 return NO_ERROR;
332 }
333
checkState()334 void Fusion::checkState() {
335 // P needs to stay positive semidefinite or the fusion diverges. When we
336 // detect divergence, we reset the fusion.
337 // TODO(braun): Instead, find the reason for the divergence and fix it.
338
339 if (!isPositiveSemidefinite(P[0][0], SYMMETRY_TOLERANCE) ||
340 !isPositiveSemidefinite(P[1][1], SYMMETRY_TOLERANCE)) {
341 ALOGW("Sensor fusion diverged; resetting state.");
342 P = 0;
343 }
344 }
345
getAttitude() const346 vec4_t Fusion::getAttitude() const {
347 return x0;
348 }
349
getBias() const350 vec3_t Fusion::getBias() const {
351 return x1;
352 }
353
getRotationMatrix() const354 mat33_t Fusion::getRotationMatrix() const {
355 return quatToMatrix(x0);
356 }
357
getF(const vec4_t & q)358 mat34_t Fusion::getF(const vec4_t& q) {
359 mat34_t F;
360
361 // This is used to compute the derivative of q
362 // F = | [q.xyz]x |
363 // | -q.xyz |
364
365 F[0].x = q.w; F[1].x =-q.z; F[2].x = q.y;
366 F[0].y = q.z; F[1].y = q.w; F[2].y =-q.x;
367 F[0].z =-q.y; F[1].z = q.x; F[2].z = q.w;
368 F[0].w =-q.x; F[1].w =-q.y; F[2].w =-q.z;
369 return F;
370 }
371
predict(const vec3_t & w,float dT)372 void Fusion::predict(const vec3_t& w, float dT) {
373 const vec4_t q = x0;
374 const vec3_t b = x1;
375 const vec3_t we = w - b;
376
377 // q(k+1) = O(we)*q(k)
378 // --------------------
379 //
380 // O(w) = | cos(0.5*||w||*dT)*I33 - [psi]x psi |
381 // | -psi' cos(0.5*||w||*dT) |
382 //
383 // psi = sin(0.5*||w||*dT)*w / ||w||
384 //
385 //
386 // P(k+1) = Phi(k)*P(k)*Phi(k)' + G*Q(k)*G'
387 // ----------------------------------------
388 //
389 // G = | -I33 0 |
390 // | 0 I33 |
391 //
392 // Phi = | Phi00 Phi10 |
393 // | 0 1 |
394 //
395 // Phi00 = I33
396 // - [w]x * sin(||w||*dt)/||w||
397 // + [w]x^2 * (1-cos(||w||*dT))/||w||^2
398 //
399 // Phi10 = [w]x * (1 - cos(||w||*dt))/||w||^2
400 // - [w]x^2 * (||w||*dT - sin(||w||*dt))/||w||^3
401 // - I33*dT
402
403 const mat33_t I33(1);
404 const mat33_t I33dT(dT);
405 const mat33_t wx(crossMatrix(we, 0));
406 const mat33_t wx2(wx*wx);
407 const float lwedT = length(we)*dT;
408 const float hlwedT = 0.5f*lwedT;
409 const float ilwe = 1/length(we);
410 const float k0 = (1-cosf(lwedT))*(ilwe*ilwe);
411 const float k1 = sinf(lwedT);
412 const float k2 = cosf(hlwedT);
413 const vec3_t psi(sinf(hlwedT)*ilwe*we);
414 const mat33_t O33(crossMatrix(-psi, k2));
415 mat44_t O;
416 O[0].xyz = O33[0]; O[0].w = -psi.x;
417 O[1].xyz = O33[1]; O[1].w = -psi.y;
418 O[2].xyz = O33[2]; O[2].w = -psi.z;
419 O[3].xyz = psi; O[3].w = k2;
420
421 Phi[0][0] = I33 - wx*(k1*ilwe) + wx2*k0;
422 Phi[1][0] = wx*k0 - I33dT - wx2*(ilwe*ilwe*ilwe)*(lwedT-k1);
423
424 x0 = O*q;
425 if (x0.w < 0)
426 x0 = -x0;
427
428 P = Phi*P*transpose(Phi) + GQGt;
429
430 checkState();
431 }
432
update(const vec3_t & z,const vec3_t & Bi,float sigma)433 void Fusion::update(const vec3_t& z, const vec3_t& Bi, float sigma) {
434 vec4_t q(x0);
435 // measured vector in body space: h(p) = A(p)*Bi
436 const mat33_t A(quatToMatrix(q));
437 const vec3_t Bb(A*Bi);
438
439 // Sensitivity matrix H = dh(p)/dp
440 // H = [ L 0 ]
441 const mat33_t L(crossMatrix(Bb, 0));
442
443 // gain...
444 // K = P*Ht / [H*P*Ht + R]
445 vec<mat33_t, 2> K;
446 const mat33_t R(sigma*sigma);
447 const mat33_t S(scaleCovariance(L, P[0][0]) + R);
448 const mat33_t Si(invert(S));
449 const mat33_t LtSi(transpose(L)*Si);
450 K[0] = P[0][0] * LtSi;
451 K[1] = transpose(P[1][0])*LtSi;
452
453 // update...
454 // P = (I-K*H) * P
455 // P -= K*H*P
456 // | K0 | * | L 0 | * P = | K0*L 0 | * | P00 P10 | = | K0*L*P00 K0*L*P10 |
457 // | K1 | | K1*L 0 | | P01 P11 | | K1*L*P00 K1*L*P10 |
458 // Note: the Joseph form is numerically more stable and given by:
459 // P = (I-KH) * P * (I-KH)' + K*R*R'
460 const mat33_t K0L(K[0] * L);
461 const mat33_t K1L(K[1] * L);
462 P[0][0] -= K0L*P[0][0];
463 P[1][1] -= K1L*P[1][0];
464 P[1][0] -= K0L*P[1][0];
465 P[0][1] = transpose(P[1][0]);
466
467 const vec3_t e(z - Bb);
468 const vec3_t dq(K[0]*e);
469 const vec3_t db(K[1]*e);
470
471 q += getF(q)*(0.5f*dq);
472 x0 = normalize_quat(q);
473 x1 += db;
474
475 checkState();
476 }
477
478 // -----------------------------------------------------------------------
479
480 }; // namespace android
481
482