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