Multi-Circular Shape Approximations

• Multi-circular shape approximations are methods that represent and analyze complex shapes using collections of circular primitives, ensuring continuity, compactness, and computational efficiency.
• They are applied in CNC toolpath synthesis, probabilistic collision evaluation, and 3D reconstruction, providing measurable error bounds and robust performance under uncertainty.
• Advanced techniques leveraging chord error minimization and constrained optimization yield manufacturable arc and polynomial approximations that meet precise engineering requirements.

Multi-circular shape approximations are a family of geometric techniques in which complex shapes are represented, analyzed, or manipulated via collections of circular primitives—either as multiple arcs in the plane or as stacked/overlapping circles in 2D/3D. These approximations arise in diverse domains: from generating manufacturable CNC toolpaths under curvature and continuity constraints, to uncertainty-bounded collision probability estimation in robotics and autonomous driving, to 3D reconstruction of objects exhibiting circular symmetry. The efficacy of such approaches derives from the algebraic tractability, geometric regularity, and direct computational representation of circles and arcs.

1. Fundamentals and Problem Setting

Multi-circular shape approximation leverages the expressiveness and compactness of circular primitives to meet application-specific requirements regarding continuity, minimality, or computational efficiency. Key classes include:

• Planar arc/segment decomposition: Approximating object contours using a minimal sequence of straight segments and circular arcs or biarcs, typically with explicit bounds on chord error or angular deviation (Scarparo, 2013).
• Multi-circle covering and containment: Decomposing arbitrary shapes, such as vehicle or obstacle footprints, into unions of overlapping equal-radius circles, facilitating analytic computation for stochastic geometric queries (Tolksdorf et al., 2024).
• Axisymmetric 3D structure estimation: Fitting a vertical stack of circular cross-sections to infer the dimensions of objects presumed to be circularly symmetric under perspective views (Xompero et al., 2019).
• Polynomial arc approximation: Constructing parametric Bézier or spline representations that optimally approximate circular arcs over prescribed intervals and continuity, generalizing single-arc approaches and supporting composite shapes (Vavpetič et al., 2017).

Central to all these formulations are notions of error (radial, Hausdorff, coverage), explicit construction algorithms, and—where required—optimization of the number or placement of primitives.

2. Planar Arc-Chain Approximation and CNC Toolpath Generation

A prominent use case is CNC-compatible toolpath synthesis from discrete planar boundary data, where only straight lines (G1) and circular arcs (G2/G3) are natively supported by machine controllers. The method of Scarparo (Scarparo, 2013) proceeds in distinct algorithmic stages:

1. Corner detection: Compute localized geometric curvature changes via tangent vectors at each sampled point $P_i$; points exhibiting abrupt angular deviation (i.e., $\cos\alpha_i > -\epsilon_{\mathrm{angle}}$) are selected as "anchors." Subclusters are fitted by best-fit circles (Taubin's method); true corners are recovered as intersection points of adjacent fits.
2. Constrained arc fitting: For each smooth subchain $P_0,\ldots,P_{N+1}$, a single arc passing exactly through the endpoints is fitted by minimizing the sum of squared deviations in radius from the candidate center $C(t)$, parameterized as a one-dimensional unconstrained optimization:

$f(t) = \sum_{k=1}^N [R(t) - \|P_k - C(t)\|]^2$

yielding $C^*,R^*$.

1. Primitive selection and continuity: Each segment is tested for admissibility as a straight line, single arc, or requires splitting. Where consecutive arcs meet with insufficient geometric continuity (joint angle exceeding tolerance), a biarc is inserted to enforce $G^1$ smoothness, using explicit construction formulas.
2. Longest-arc sweep: The point chain is partitioned into maximal admissible arcs, minimizing the total primitive count.

This approach ensures a near-minimal, $G^0$- or $G^1$-continuous path whose elements correspond exactly to machine-instruction primitives, with explicit user control of chord error and angular tolerances. No higher-order splines are involved (Scarparo, 2013).

3. Multi-Circular Approximations for Probabilistic Collision Evaluation

In automated driving and robotics, fast yet accurate collision probability estimation under uncertainty is essential. The multi-circular shape approximation technique (Tolksdorf et al., 2024) is designed to replace Monte Carlo sampling by analytic evaluation, supporting real-time performance.

• For a compact planar region $S \subset \mathbb{R}^2$ (e.g., a vehicle), $N_c$ identical-radius circles are placed (typically along the principal axis), with centers $q_i$, $i=1,\ldots,N_c$. The minimal covering radius and inter-circle distance are derived from the geometry:

$$r_e = \sqrt{ \left(\frac{\ell}{2N_c}\right)^2 + \left(\frac{w}{2}\right)^2 },\quad d_c = 2\sqrt{ r_e^2 - \left(\frac{w}{2}\right)^2 }$$

• Under Gaussian position uncertainty for the object ($q_o \sim \mathcal{N}(\mu, \Sigma)$) and fixed ego pose, the collision event is

$$C = \bigcup_{i=1}^{N_c} \{ \|q_o - q_i\| \leq r_e + r_o \}$$

where $r_o$ is the radius of the object approximation. The collision probability $P(C)$ is computed via analytic integrals (single- or double-dimensional) for each disk, using inclusion–exclusion to correct for overlaps.

• Upper and lower bounds on $P(C)$ are found by switching between covering and inscribed circle configurations, providing guaranteed corridors for the true probability.

The approach achieves orders-of-magnitude speedup versus standard Monte Carlo sampling (e.g., 0.17 ms per evaluation vs. 56 ms for 10,000-sample MC with similar accuracy), and provides explicit error bounds via the corridor width $\Delta = P_{\text{up}} - P_{\text{low}}$ (Tolksdorf et al., 2024).

4. 3D Multi-Circular Shape Estimation from Multi-View Data

For reconstructing objects with circular symmetry from sparse perspective images, the LoDE pipeline (Xompero et al., 2019) samples and verifies stacks of horizontal circumferences:

• Initialization: For estimated object centroid $\tilde{\mathbf{X}}$, $L$ circumferences are placed at vertical positions $z_l$, each parameterized by center $(\tilde x, \tilde y, z_l)$ and decreasing candidate radii $r_{l}^{(i)}$.
• Generative sampling: At each radius, $N$ points are placed uniformly along the circle.
• Projection and mask-testing: Each 3D point $\mathbf{Q}_{n,l}^{(i)}$ is reprojected into both camera views; if all projected points coincide with 2D segmentation masks, the candidate circumference is accepted.
• Output extraction: Object width and height are read from the maximal converged radius and vertical span among accepted cross-sections.

This discrete, sampling-based approach allows robust recovery of size and shape, even under challenging imaging conditions and without depth sensing. Reported median errors are $\sim$10 mm (width) and $\sim$15 mm (height), with a localization success ratio (LSR) of 86.96% (Xompero et al., 2019).

5. Polynomial and Bézier Arc Approximations for Composite Shapes

For composite shapes comprising multiple contiguous arc segments, optimal polynomial curve approximation techniques offer a compact, systematically optimal representation (Vavpetič et al., 2017):

• Approximation metric: Reduce the geometric arc-approximation problem to minimizing the radial or squared-radius deviation between the true arc $c(s)$ over $s \in [−\varphi, \varphi]$ and its Bézier counterpart $p_n(t)$; the relevant error is $|\|p_n(t)\|_2 - 1|$.
• Constrained Chebyshev formulation: The squared-radius error is forced to vanish with prescribed multiplicity at the endpoints, yielding a polynomial minimax problem equivalent to constrained uniform approximation.
• Nonlinear system and explicit solutions: For prescribed order $(n,k)$, the system reduces to $n-k-1$ nonlinear equations for the interior control points, often admitting closed-form solutions in low-degree cases (e.g., quadratic or cubic, $G^0$ or $G^1$ contact).
• Extension to multi-arc shapes: Each segment is reparameterized and independently fit, with $G^1$ or $G^2$ continuity enforced at boundaries via shared control-point algebra—linear constraints suffice for matching derivatives. The block-diagonal structure facilitates both per-segment and coupled solves.
• Error guarantees: For degree $n$ and arc angle $\varphi$, the error decays rapidly with $n$ (e.g., $O(\varphi^{2n})$), yielding explicit and optimal (in the Chebyshev sense) polynomial approximants for use as global CAD or rendering primitives (Vavpetič et al., 2017).

6. Key Advantages, Limitations, and Extensions

The multi-circular paradigm provides several intrinsic benefits:

• Expressiveness: Complex curves and shapes are captured by flexible combinations of a small number of circles or arcs, with readily quantifiable approximation errors (Scarparo, 2013, Vavpetič et al., 2017).
• Algorithmic simplicity and speed: Analytic integrals and explicit constructions eliminate the need for sampling-based or iterative surface fitting in high-dimensional settings (Tolksdorf et al., 2024, Xompero et al., 2019).
• Manufacturability: The output primitives (segments, arcs) map directly to standard control instructions for numerical machinery (Scarparo, 2013).
• Robustness under geometric uncertainty and noise: Circular primitives admit analytic probability integrals and explicit continuity-enforcement, accommodating errors due to discretization, imaging, or registration (Tolksdorf et al., 2024, Xompero et al., 2019).

However, the approach assumes and relies upon inherent circularity or proximity to circular models; non-Gaussian or orientation-dependent uncertainties, non-axisymmetric 3D objects, or shapes with significant high-curvature or piecewise-linear features may require additional model adaptation or extended primitives. Possible extensions involve use of adaptive radius selection, advanced circle-placement (e.g., $k$-center optimization), and coupling to higher-order moment-based uncertainty representations (Tolksdorf et al., 2024).

7. Comparative Summary of Methodologies

Application Domain	Multi-Circular Methodology	Performance/Guarantees
CNC Toolpath Generation	Longest-arc/biarc fit (Scarparo, 2013)	Minimal tool moves, $G^0/G^1$ continuity, native G-code output
Probabilistic Collision	Circle union/inclusion-exclusion integrals (Tolksdorf et al., 2024)	Three-digit precision, two orders magnitude faster than Monte Carlo, error corridors
3D Object Reconstruction	Iterative circumference sampling/reprojection (Xompero et al., 2019)	Sub-centimeter geometry, robust to transparency, high LSR
Polynomial Arc Approximation	(Constrained) Chebyshev Bézier fits (Vavpetič et al., 2017)	Explicit error bounds, $G^k$ continuity, optimal minimax error for arc splines

The multi-circular shape approximation framework thus unifies a broad class of geometric, computational, and control applications where precision, tractability, and manufacturability are paramount. Each methodology aligns algorithmic construction with domain-specific constraints and performance guarantees, exploiting the unique advantages of circular primitives for deep integration into sensing, planning, and manufacturing pipelines.