Repository files navigation

A Typescript implementation of the iterative closest point algorithm using both the point-to-point and point-to-plane variants, used for point cloud registration. An example of usage is shown in the provided React Three Fiber demo; you can visit the demohere.

• Simply install the npm package with the command below

$ npm install icpts

• If you'd like to build from source, pull the repository and navigate to theicpts directory. Runnpm i to install the dependencies, followed bynpm run build to build the package. The build artifacts will be placed in thedist directory, and the project can be used as a localnode module.

First import the package.

importicptsfrom"icpts"

We expose both ICP strategies using seperate functions under the namesicpts.pointToPlane andicpts.pointToPoint. Both strategies have identical interfaces. We expect the points from two point clouds, a source and a reference, to be provided using flat arrays ([x, y, z][]). Additional options are exposed to provide an error tolerance for early stopping, a maximum iteration count and an initial source pose transform.

Each strategy returns the optimal transform from the source cloud to the reference, stored in a flat array using column major ordering. The final error is also returned. Assumingsource andreference are defined:

importicptsfrom"icpts"constoptions={initialPose:IDENTITY,// [1, 0, 0, 0, 0, 1, ...]tolerance:1e-10,maxIterations:50};const{ transform, error}=icpts.pointToPoint(source,reference,options);// or icpts.pointToPlane

You may refer to more detailed example usage inicpts-demo or in theicpts tests, specificallyicpts.test.ts;

Pull requests, and general improvements / feedback are welcome. To run the project locally, follow the steps below:

• Pull the repository and navigate to theicpts directory.
• Runnpm i to install the dependencies.
• That's pretty much it. To test that everything is working, you can run the primary test withts-node ./test/icp.test.ts (yes, a test framework probably should've been added but we also don't have that many tests yet).

To run the demo site, navigate to the root of the repo and runpnpm install to install the dependencies. The site can be launched locally by then runningpnpm dev and navigating tolocalhost:3000.

Good question, it probably shouldn't (and I wouldn't recommend using it for anything half serious). To answer the question though, no one was brave enough to publish an ICP library using JavaScript/TypeScript (for good reason) so we decided why not? We also tried to make itsomewhat readable.

• Please note that there are limitations on the point cloud sizes due to the usage of wasm (witheigen-js) and the associated limitation on memory.
• Also due to the eigen dependency, the package is not fully supported in browser environments (and it certainly won't work in a web worker, so you wouldn't want to use it in a browser anyway). You can, however, easily use it on any node-based server, including Next JS server-side APIs.
• We are not using any robust variants of ICP, so successful point cloud registration requires decent initialization with some overlap between the clouds. Nevertheless, it seems like point-to-point seems to be more robust to the initial pose but converges more slowly than point-to-plane.

• Generalized ICP
• Consider adding more tests.

• [1]Point to Plane
• [2]Point to Point