By now, the curvature estimation uses line fitting on left and right laser ring neighbors of each point.
The angle between these lines and the depth gap between the studied point and these neighborhoods are analyzed.
With this MR, the curvature is estimated as in LOAM algorithm with a convolution. A kernel (1 1 ... -N 1 1 ) of size N+1 is applied on points' neighborhood.
This solution permits to achieve comparable results on common datasets relatively to previous version. However, results on forest-walking datasets are not as good as with previous version (see picture below, left : LOAM based, right : master result).
Some results on keypoints extraction are saved here for more info