whigg/picp

pICP: plain Iterative Closest Point, a no-frills implementation in C++ of a simple rigid 3D registration algorithm. Easy to integrate, uses minimal dependencies.

pICP

This repo provides a clean and fast implementation of the ICP method by Besl and McKay in C++ with minimal dependencies (in fact, only OpenCV is strictly required for data matrix manipulation and SVD computation). The algorithm is one of the baseline methods on rigid alignment of point clouds, and this implementation attempts to provide an entry point for learning it with no practical hassles. At the same time, it works fast and robustly enough to be used out-of-the box on any of your projects.

How it works

The problem

Given two point sets:

find a rotation R and translation T that minimizes the error:

SVD based transform estimation

Given a matrix W made by a set of correspondent N points, centered to its mean

obtain the Singular Value Decomposition of W = U * S * Vt.

Theorem without proof says that if rank( W ) = 3, the optimal solution of E( R, t ) is unique and given by:

This process can be run iteratively, selecting correspondent points by different criteria (neighbourhood, random search...), until convergence.

Dependencies

Only OpenCV is required if you are just using the ICP class on your project.

A real test executable is provided, allowing to register any desired pair of point clouds (only .obj format supported at the moment). The PointCloud class uses TinyObjLoader (header included) for parsing OBJ files, and the PlainICP executable requires Boost for parsing external arguments.

Usage

Just include the ICP class header on your project.

Check the provided cmake project for compiling an executable which shows how to use the class:

mkdir ./build
cd ./build
cmake ..
make

Sample call with provided sample OBJ files:

./PlainICP  --source-obj-file ../samples/bunny_head.obj --target-obj-file ../samples/bunny.obj --transformed-obj-file ../samples/aligned_bunny_head.obj --tolerance 0.0005 --verbose 1

License

This software release is primarily MIT licensed. Some files contain third-party code under other licens