Robust Kernels Library¶

TUM Autonomous Motorsport package for robust kernels for scan registration

Robust kernels down-weight correspondences with large residuals so that outliers (dynamic objects, wrong matches) do not dominate the registration. The kernel is a compile-time property of the config (TConfig::KERNEL) and returns a scalar weight w ∈ [0, 1] for a residual, given a kernel_scale. Implementations are based on small_gicp and KISS-ICP.

Available Kernels¶

Kernel	Weight `w(r)`	Behavior
`NONE`	`1`	No down-weighting; plain least squares.
`HUBER`	`1` if `‖r‖² < s`, else `s / ‖r‖²`	Quadratic for inliers, linear for outliers; mild down-weighting.
`CAUCHY`	`1 / (1 + ‖r/s‖²)`	Smooth, heavier down-weighting than Huber; never reaches zero.
`GEMANMCCLURE`	`s² / (s + ‖r‖²)²`	Strong outlier rejection, weight → 0 for large residuals; KISS-ICP default.
`WELSCH`	`exp(-‖r/s‖²)`	Exponential cut-off; the most aggressive, effectively hard-rejects far residuals.

Here r is the correspondence residual and s the kernel_scale.