Canonical forms and parameter identification problems in perspective systems (Q1611929)

A motivation of this paper comes from machine vision problems of identification of motion parameters from observation data of a projection of feature points gathered over a time interval. Mathematically, the following problem is considered. Given a linear system \[ \left\{ \begin{matrix} \dot{\mathcal{X}}={\mathcal{A}}^T{\mathcal{X}}, \quad {\mathcal{X}}(0)={\mathcal{X}}_0\\ {\mathcal{Y}}={\mathcal{C}}{\mathcal{X}}, \end{matrix} \right. \tag{1} \] \({\mathcal{X}}\), \({\mathcal{Y}}\) belonging, respectively, to projective spaces \(RP^{n-1}\) and \(RP^{m-1}\), and asuming that \({\mathcal{Y}}(t)\), \(t\in [0,T)\), is known, compute the entries of the matrices \({\mathcal{A}}\), \({\mathcal{C}}\), and the initial condition \({\mathcal{X}}_0\). The system (1) represents in the paper either the affine dynamics of a rigid body or the Riccati dynamics in \(\mathbb{R}^3\) [cf. \textit{B. K. Ghosh}, \textit{H. Inaba} and \textit{S. Takahashi}, ``Identification of Riccati dynamics under perspective and orthographic observations'', IEEE Trans. Autom. Control, 45, No. 7, 1267--1278 (2000; Zbl 0988.93016)]. The output map of (1) takes the form of either perspective or orthographic projection. It is observed that the identification is possible only modulo orbits of the so-called perspective group \({\mathcal{P}}\cong GL(n)\times \mathbb{R}\times \mathbb{R}^+\) acting on (1) in the following way: \[ ((P,\lambda,\mu),{\mathcal{A}}^T,{\mathcal{C}})\mapsto (\lambda I+P{\mathcal{A}}^TP^{-1} \mu{\mathcal{C}}P^{-1}). \] Two cases are studied: both \({\mathcal{A}}^T\) and \({\mathcal{C}}\) are unknown, and unknown \({\mathcal{A}}^T\), but known \({\mathcal{C}}\), referred to as the unknown and the known camera calibration case. The main results of the paper consist in providing canonical forms of the system (1) acted on by the perspective group. Parameters of those canonical forms have been identified from noisy observations using the extended Kalman filter technique. Computer simulations serve as an illustration of the theory.