研究旋转矩阵,旋转向量,单位四元数,都是为了表达机器人的姿态。欧拉角在SLAM的应用中不多,就不涉及了。
正交矩阵的行列式为正负1
所谓奇异性,我理解的就是做不到一一对应,在这里是一个姿态可以对应多个旋转向量
在上图中,向量 v \mathbf v v沿着旋转轴 n \mathbf n n旋转 θ \mathit \theta θ角,变为 v ′ \mathbf v' v′。现在我们要将 v ′ \mathbf v' v′分解成向量 v \mathbf v v,旋转轴 n \mathbf n n,旋转角 θ \mathit \theta θ组成的量
v ′ = v k + v p ′ = v k + a + b = ( v ⋅ n ) n + ∣ v i ∣ sin θ j + v i cos θ = ( v ⋅ n ) n + ∣ ( n × v ) × n ∣ sin θ n × v ∣ n × v ∣ + ( ( n × v ) × n ) cos θ = ( v ⋅ n ) n + ∣ n ∣ ∣ v ∣ sin α ∣ n ∣ sin π 2 sin θ n × v ∣ n ∣ ∣ v ∣ sin α + ( ( n × v ) × n ) cos θ = ( v ⋅ n ) n + ( n × v ) sin θ + ( v − ( v ⋅ n ) n ) cos θ = ( 1 − cos θ ) ( v ⋅ n ) n + n ∧ v sin θ + v cos θ = ( 1 − cos θ ) n n T v + sin θ n ∧ v + cos θ v = [ cos θ I + ( 1 − cos θ ) n n T + sin θ n ∧ ] v \begin{split} \mathbf v' & = \mathbf v_k + \mathbf v'_p \\ & = \mathbf v_k + \mathbf a + \mathbf b \\ & = (\mathbf v \cdot \mathbf n) \mathbf n + | \mathbf v_i|\sin\theta \mathbf j + \mathbf v_i \cos \theta \\ & = (\mathbf v \cdot \mathbf n) \mathbf n + |(\mathbf n \times \mathbf v) \times \mathbf n | \sin \theta \frac {\mathbf n \times \mathbf v}{|\mathbf n \times \mathbf v|} + ((\mathbf n \times \mathbf v) \times \mathbf n)\cos \theta \\ & = (\mathbf v \cdot \mathbf n) \mathbf n + |\mathbf n| |\mathbf v|\sin \alpha |\mathbf n| \sin \frac{\pi}{2} \sin\theta \frac{\mathbf n \times \mathbf v}{|\mathbf n| |\mathbf v| \sin \alpha} + ((\mathbf n \times \mathbf v) \times \mathbf n)\cos \theta \\ & = (\mathbf v \cdot \mathbf n) \mathbf n + (\mathbf n \times \mathbf v) \sin \theta + (\mathbf v - (\mathbf v \cdot \mathbf n ) \mathbf n) \cos \theta \\ & = (1 - \cos \theta) (\mathbf v \cdot \mathbf n) \mathbf n + \mathbf n^{\land} \mathbf v \sin \theta + \mathbf v \cos \theta \\ & = (1 - \cos \theta) \mathbf n \mathbf n^{\mathrm T} \mathbf v + \sin \theta \mathbf n^{\land} \mathbf v + \cos \theta \mathbf v \\ & = [\cos \theta \mathbf I + (1 - \cos \theta) \mathbf n \mathbf n^{\mathrm T} + \sin \theta \mathbf n^{\land}] \mathbf v \end{split} v′=vk+vp′=vk+a+b=(v⋅n)n+∣vi∣sinθj+vicosθ=(v⋅n)n+∣(n×v)×n∣sinθ∣n×v∣n×v+((n×v)×n)cosθ=(v⋅n)n+∣n∣∣v∣sinα∣n∣sin2πsinθ∣n∣∣v∣sinαn×v+((n×v)×n)cosθ=(v⋅n)n+(n×v)sinθ+(v−(v⋅n)n)cosθ=(1−cosθ)(v⋅n)n+n∧vsinθ+vcosθ=(1−cosθ)nnTv+sinθn∧v+cosθv=[cosθI+(1−cosθ)nnT+sinθn∧]v
要注意点乘得到的是标量;叉乘得到的是向量,且要用右手定则分析其方向。
( v ⋅ n ) n = n n T v (\mathbf v \cdot \mathbf n) \mathbf n = \mathbf n \mathbf n^{\mathrm T} \mathbf v (v⋅n)n=nnTv可以从微观层面去证明,很简单。
上面的计算有很多细节值得思考,比如第三行的 ∣ v i ∣ sin θ j | \mathbf v_i|\sin\theta \mathbf j ∣vi∣sinθj为什么要乘单位向量 j \mathbf j j? v i cos θ \mathbf v_i \cos \theta vicosθ为什么又不乘?
上式中中括号里的内容是一个矩阵,就是旋转矩阵 R \mathbf R R:
R = cos θ I + ( 1 − cos θ ) n n T + sin θ n ∧ \begin{equation} \mathbf R = \cos \theta \mathbf I + (1 - \cos \theta) \mathbf n \mathbf n^{\mathrm T} + \sin \theta \mathbf n^{\land} \end{equation} R=cosθI+(1−cosθ)nnT+sinθn∧
将(1)式两边取迹:
t r ( R ) = 3 cos θ + ( 1 − cos θ ) n T n + sin θ ⋅ 0 = 3 cos θ + 1 − cos θ = 2 cos θ + 1 \begin{align} \mathrm tr(\mathbf R) & = 3 \cos \theta + (1 - \cos \theta) \mathbf n^{\mathrm T} \mathbf n + \sin \theta \cdot 0 \notag \\ & = 3 \cos \theta + 1 - \cos \theta \notag \\ & = 2 \cos \theta + 1 \end{align} tr(R)=3cosθ+(1−cosθ)nTn+sinθ⋅0=3cosθ+1−cosθ=2cosθ+1
从旋转矩阵中得到旋转角:
θ = arccos t r ( R ) − 1 2 \theta = \arccos \frac {\mathrm tr(\mathbf R) - 1} {2} θ=arccos2tr(R)−1
至于转轴 n \mathbf n n,把它看作普通的向量,用它对应的旋转矩阵 R \mathbf R R去左乘旋转它,它不会发生任何变化,于是有下式:
R n = n \mathbf R \mathbf n =\mathbf n Rn=n
因此,转轴 n \mathbf n n是矩阵 R \mathbf R R的特征值1对应的特征向量,用求特征向量的方法去求即可,再归一化可以得到转轴 n \mathbf n n:
R n = n s.t. n T n = 1 \mathbf R \mathbf n = \mathbf n \qquad \textit{s.t.} \qquad \mathbf n^{\mathrm T} \mathbf n = 1 Rn=ns.t.nTn=1
用四元数 q \mathbf q q对纯虚四元数数表示的空间点 p = [ 0 , v p T ] T \mathbf p= [0,\mathbf v_p^{\mathrm T}]^{\mathrm T} p=[0,vpT]T进行旋转有:
p ′ = q p q − 1 = q + p + q − 1 = q + ( q − 1 ) ⊕ p = [ s − v T v s I + v ∧ ] [ s v T − v s I + v ∧ ] [ 0 v p ] = [ s 2 + v T v s v T − s v T − v T v ∧ s v − s v − v ∧ v v v T + s 2 I + 2 s v ∧ + ( v ∧ ) 2 ] [ 0 v p ] = [ 1 ( v ∧ v ) T 0 v v T + s 2 I + 2 s v ∧ + ( v ∧ ) 2 ] [ 0 v p ] = [ 1 0 T 0 v v T + s 2 I + 2 s v ∧ + ( v ∧ ) 2 ] [ 0 v p ] = [ 0 ( v v T + s 2 I + 2 s v ∧ + ( v ∧ ) 2 ) v p ] \begin{split} \mathbf p^{'} & = \mathbf q \mathbf p \mathbf q^{-1} \\ & = \mathbf q^{+} \mathbf p^{+} \mathbf q^{-1} \\ & = \mathbf q^{+} (\mathbf q^{-1})^{\oplus} \mathbf p \\ & = \begin{bmatrix} s & -\mathbf v^{\mathrm T} \\ \mathbf v & s \mathbf I + \mathbf v^{\land} \end{bmatrix} \begin{bmatrix} s & \mathbf v^{\mathrm T} \\ -\mathbf v & s \mathbf I + \mathbf v^{\land} \end{bmatrix} \begin{bmatrix} 0 \\ \mathbf v_{p} \end{bmatrix} \\ & = \begin{bmatrix} s^2 + \mathbf v^{\mathrm T} \mathbf v & s\mathbf v^{\mathrm T} -s\mathbf v^{\mathrm T} - \mathbf v^{\mathrm T} \mathbf v^{\land} \\ s\mathbf v - s\mathbf v - \mathbf v^{\land} \mathbf v& \mathbf v \mathbf v^{\mathrm T} + s^2 \mathbf I + 2s \mathbf v^{\land} + (\mathbf v^{\land})^2 \end{bmatrix} \begin{bmatrix} 0 \\ \mathbf v_{p} \end{bmatrix} \\ & = \begin{bmatrix} 1 & (\mathbf v^{\land} \mathbf v)^{\mathrm T} \\ \mathbf 0 & \mathbf v \mathbf v^{\mathrm T} + s^2 \mathbf I + 2s \mathbf v^{\land} + (\mathbf v^{\land})^2 \end{bmatrix} \begin{bmatrix} 0 \\ \mathbf v_{p} \end{bmatrix} \\ & = \begin{bmatrix} 1 & \mathbf 0^{\mathrm T} \\ \mathbf 0 & \mathbf v \mathbf v^{\mathrm T} + s^2 \mathbf I + 2s \mathbf v^{\land} + (\mathbf v^{\land})^2 \end{bmatrix} \begin{bmatrix} 0 \\ \mathbf v_{p} \end{bmatrix} \\ & = \begin{bmatrix} 0 \\ (\mathbf v \mathbf v^{\mathrm T} + s^2 \mathbf I + 2s \mathbf v^{\land} + (\mathbf v^{\land})^2) \mathbf v_{p}\end{bmatrix} \end{split} p′=qpq−1=q+p+q−1=q+(q−1)⊕p=[sv−vTsI+v∧][s−vvTsI+v∧][0vp]=[s2+vTvsv−sv−v∧vsvT−svT−vTv∧vvT+s2I+2sv∧+(v∧)2][0vp]=[10(v∧v)TvvT+s2I+2sv∧+(v∧)2][0vp]=[100TvvT+s2I+2sv∧+(v∧)2][0vp]=[0(vvT+s2I+2sv∧+(v∧)2)vp]
上面可以看出被单位四元数旋转后的坐标点(纯虚四元数)依然是一个纯虚四元数
由此可得单位四元数到旋转矩阵的变换:
R = v v T + s 2 I + 2 s v ∧ + ( v ∧ ) 2 \begin{equation} \mathbf R = \mathbf v \mathbf v^{\mathrm T} + s^2 \mathbf I + 2s \mathbf v^{\land} + (\mathbf v^{\land})^2 \end{equation} R=vvT+s2I+2sv∧+(v∧)2
将(3)式两边取迹:
t r ( R ) = v T v + 3 s 2 + 0 + t r ( ( v ∧ ) 2 ) \mathrm tr(\mathbf R) = \mathbf v^{\mathrm T} \mathbf v + 3s^2 + 0 + \mathrm tr((\mathbf v^{\land})^2) tr(R)=vTv+3s2+0+tr((v∧)2)
其中: ( v ∧ ) 2 = [ 0 − v 3 v 2 v 3 0 − v 1 − v 2 v 1 0 ] [ 0 − v 3 v 2 v 3 0 − v 1 − v 2 v 1 0 ] = [ − v 2 2 − v 3 2 ⋯ ⋯ ⋯ − v 1 2 − v 3 2 ⋯ ⋯ ⋯ − v 1 2 − v 2 2 ] \begin{split} (\mathbf v^{\land})^2 & = \begin{bmatrix} 0 & -v_3 & v_2 \\ v_3 & 0 & -v_1 \\ -v_2 & v_1 & 0 \end{bmatrix} \begin{bmatrix} 0 & -v_3 & v_2 \\ v_3 & 0 & -v_1 \\ -v_2 & v_1 & 0 \end{bmatrix} \\ & = \begin{bmatrix} -v_2^2 - v_3^2 & \cdots & \cdots \\ \cdots & -v_1^2-v_3^2 & \cdots \\ \cdots & \cdots & -v_1^2-v_2^2 \end{bmatrix} \end{split} (v∧)2=⎣ ⎡0v3−v2−v30v1v2−v10⎦ ⎤⎣ ⎡0v3−v2−v30v1v2−v10⎦ ⎤=⎣ ⎡−v22−v32⋯⋯⋯−v12−v32⋯⋯⋯−v12−v22⎦ ⎤
t r ( R ) = v 1 2 + v 2 2 + v 3 2 + 3 s 2 − 2 ( v 1 2 + v 2 2 + v 3 2 ) = 3 s 2 − ( v 1 2 + v 2 2 + v 3 2 ) = 3 s 2 − ( 1 − s 2 ) = 4 s 2 − 1 \begin{align} \mathrm tr(\mathbf R) & = v_1^2 + v_2^2 + v_3^2 + 3s^2 - 2(v_1^2 + v_2^2 + v_3^2) \notag \\ & = 3s^2 - (v_1^2 + v_2^2 + v_3^2) \notag \\ & = 3s^2 - (1 - s^2) \notag \\ & = 4s^2 - 1 \end{align} tr(R)=v12+v22+v32+3s2−2(v12+v22+v32)=3s2−(v12+v22+v32)=3s2−(1−s2)=4s2−1
稍作变换(4)式可以从旋转矩阵 R \mathbf R R中求得单位四元数的实部:
s = t r ( R ) + 1 4 s = \sqrt {\frac {\mathrm tr(\mathbf R) +1}{4}} s=4tr(R)+1
下面开始求虚部,将虚部当作待旋转的坐标点带入(3)式有:
R v = ( v v T + s 2 I + 2 s v ∧ + ( v ∧ ) 2 ) v = v v T v + s 2 v + 0 + 0 = ( 1 − s 2 ) v + s 2 v = v \begin{split} \mathbf R \mathbf v & = (\mathbf v \mathbf v^{\mathrm T} + s^2 \mathbf I + 2s \mathbf v^{\land} + (\mathbf v^{\land})^2) \mathbf v \\ & = \mathbf v \mathbf v^{\mathrm T} \mathbf v + s^2 \mathbf v + 0 + 0 \\ & = (1 - s^2) \mathbf v + s^2 \mathbf v \\ & = \mathbf v \end{split} Rv=(vvT+s2I+2sv∧+(v∧)2)v=vvTv+s2v+0+0=(1−s2)v+s2v=v
由此可见虚部也是矩阵 R \mathbf R R的特征值1对应的特征向量,即与旋转向量 θ n \theta \mathbf n θn在 R 3 \mathbb R^3 R3中指向同一个方向,但是他们之间的模不一样,旋转向量的模是量转角 θ \theta θ,这里虚部的模由 s 2 + v T v = 1 s^2 + \mathbf v^{\mathrm T} \mathbf v = 1 s2+vTv=1确定。故由旋转矩阵求虚部为
R v = v s.t. s 2 + v T v = 1 \mathbf R \mathbf v = \mathbf v \qquad \textit{s.t.} \qquad s^2 + \mathbf v^{\mathrm T} \mathbf v = 1 Rv=vs.t.s2+vTv=1
由式(2)与式(4)可得:
t r ( R ) = 2 cos θ + 1 = 4 s 2 − 1 \mathrm tr(\mathbf R) = 2 \cos \theta +1 = 4s^2 -1 tr(R)=2cosθ+1=4s2−1
即
cos θ = 2 s 2 − 1 \cos \theta = 2s^2 -1 cosθ=2s2−1
又有三角函数关系
cos θ = 2 cos 2 θ 2 − 1 \cos \theta = 2\cos ^2 \frac{\theta}{2} -1 cosθ=2cos22θ−1
于是有实部与转角的相互转换:
s = cos θ 2 或 θ = 2 arccos s s = \cos \frac{\theta}{2} \quad \text{或} \quad\theta = 2 \arccos s s=cos2θ或θ=2arccoss
刚刚已经分析,单位四元数的虚部 v \mathbf v v与旋转向量的转轴 n \mathbf n n在 R 3 \mathbb R^3 R3中指向同一个方向,区别是他们的模服从不同的约束。单位四元数转换为转轴 n \mathbf n n:
n = v ∣ v ∣ = v 1 − s 2 = v 1 − cos 2 θ 2 = v sin θ 2 \mathbf n = \frac {\mathbf v}{|\mathbf v|} = \frac {\mathbf v}{\sqrt{1 - s^2}} = \frac {\mathbf v}{\sqrt{1 - \cos^2 \frac {\theta}{2}}} = \frac {\mathbf v}{\sin \frac {\theta}{2}} n=∣v∣v=1−s2v=1−cos22θv=sin2θv
旋转向量转换为虚部 v \mathbf v v:
v = sin θ 2 n = 1 − s 2 n \mathbf v = \sin \frac {\theta}{2} \mathbf n = \sqrt{1 - s^2} \mathbf n v=sin2θn=1−s2n
旋转矩阵 R \mathbf R R | 旋转向量 θ n \theta \mathbf n θn | 单位四元数 q = [ s , v ] \mathbf q = [s, \mathbf v] q=[s,v] | |
---|---|---|---|
旋转矩阵 R \mathbf R R | R = cos θ I + ( 1 − cos θ ) n n T + sin θ n ∧ \mathbf R = \cos \theta \mathbf I + (1 - \cos \theta) \mathbf n \mathbf n^{\mathrm T} + \sin \theta \mathbf n^{\land} R=cosθI+(1−cosθ)nnT+sinθn∧ | R = v v T + s 2 I + 2 s v ∧ + ( v ∧ ) 2 \mathbf R = \mathbf v \mathbf v^{\mathrm T} + s^2 \mathbf I + 2s \mathbf v^{\land} + (\mathbf v^{\land})^2 R=vvT+s2I+2sv∧+(v∧)2 | |
旋转向量 θ n \theta \mathbf n θn | θ = arccos t r ( R ) − 1 2 R n = n s.t. n T n = 1 \theta = \arccos \frac {\mathrm tr(\mathbf R) - 1} {2} \\ \mathbf R \mathbf n = \mathbf n \quad \textit{s.t.} \quad \mathbf n^{\mathrm T} \mathbf n = 1 θ=arccos2tr(R)−1Rn=ns.t.nTn=1 | θ = 2 arccos s n = v ∣ v ∣ = v 1 − s 2 = v sin θ 2 \quad\theta = 2 \arccos s \\ \mathbf n = \frac {\mathbf v}{\lvert \mathbf v \rvert} = \frac {\mathbf v}{\sqrt{1 - s^2}} = \frac {\mathbf v}{\sin \frac {\theta}{2}} θ=2arccossn=∣v∣v=1−s2v=sin2θv | |
单位四元数 q = [ s , v ] \mathbf q = [s, \mathbf v] q=[s,v] | s = t r ( R ) + 1 4 R v = v s.t. s 2 + v T v = 1 s = \sqrt {\frac {\mathrm tr(\mathbf R) +1}{4}} \\ \mathbf R \mathbf v = \mathbf v \quad \textit{s.t.} \quad s^2 + \mathbf v^{\mathrm T} \mathbf v = 1 s=4tr(R)+1Rv=vs.t.s2+vTv=1 | s = cos θ 2 v = sin θ 2 n = 1 − s 2 n s = \cos \frac{\theta}{2} \\ \mathbf v = \sin \frac {\theta}{2} \mathbf n = \sqrt{1 - s^2} \mathbf n s=cos2θv=sin2θn=1−s2n |