Delaunay Triangulation applied in real-time on a WebCam stream

10 May 2013 by David Corvoysier

I already detailed in a previous post how typical image processing algorithms could be applied in real-time on a video stream using the HTML5 canvas to produce video effects.

In this article, I will explain how the same kind of effects can be applied on a WebCam stream thanks to the getUserMedia API.

Capturing WebCam frames in the Canvas

It is not possible (yet ?) to grab directly a WebCam frame to inject it in the HTML5 canvas: you have to go through an intermediary step, capturing first the stream in a video element, then transferring it frame by frame in the canvas.

As displayed below, we will use the getUserMedia API to grab the video stream from a WebCam and inject it in a video element:

navigator.getUserMedia({video: true}, 
	function(stream) {
		video.src = URL.createObjectURL(stream);
		// Process frames here
	function (error) {
		// Error handling

When the getUserMedia API is called, the user will be prompted to give access to its WebCam:

  • acceptance will trigger the first function that creates an URL object from the stream and pass it to the video element,
  • refusal will trigger the error function.

Once the stream has been successfully redirected to the video element, we will start capturing frames in the canvas at regular intervals timed by requestAnimationFrame. To make sure that there is actually something to be captured, we test against the HAVE_ENOUGH_DATA state for the video element before grabbing a frame.

function tick() {

	if (video.readyState === video.HAVE_ENOUGH_DATA) {
		ctx.drawImage(video, 0, 0, canvasWidth, canvasHeight);
	var imageData = ctx.getImageData(0, 0, cwidth, cheight);

Once a frame has been transferred to the canvas, it is extracted as a byte array to apply our image processing algorithms.

Detecting the singularity points

We will now use a computer vision algorithm called FASTto detect remarkable points (“corners”) in the image. To ease corner detection, the frame has first to be converted to a grayscale image.

Several Open Source implementations of the corresponding algorithms exist on the Web: we will use the JSFeat library that provides a neat wrapper around optimized implementations of the most typical ones, including grayscale and FAST.

jsfeat.fast_corners.detect(img_u8, corners, 5);

The corners detected in the image are stored in an array of points: {x,y}.

Applying Delaunay Triangulation

We then apply a Delaunay Triangulation alogithm to the set of points to identify triangles covering the image.

var triangles = triangulate(vertices);

We will use at this stage a fast Open Source implementation of the Delaunay Triangulation algorithm developed by ironwallaby.

Rendering back to Canvas

The final step is to render back the result of the Delaunay Triangulation to the canvas by painting each triangle with a color representing its content in the original image.

For the sake of simplicity, we will just pick the color of a point that we know to be inside the triangle, but a more complex process could be used to select a real average of the triangle colors.

var getTriangleColor = function (img,triangle) {
	var getColor = function (point) {
		var offset = (point.x+point.y*cwidth)*4;
		return {[offset],[offset+1],[offset+2]  };
	var midPoint = function (point1,point2) {
		return {x:(point1.x+point2.x)/2,
	// Pick a point inside the triangle
	var point1 = midPoint(triangle.a,triangle.b);
	var point = midPoint(point1,triangle.c);
	return getColor({x:Math.floor(point.x),y:Math.floor(point.y)});

Each triangle is painted using simple canvas drawing primitives:

for(var i=0;i<triangles.length;i++) {
	var color = triangles[i].color = getTriangleColor(imageData,triangles[i]);
	gridCtx.fillStyle = 'rgb('+


Click on the image below to see how it works when all the pieces are put together:

The code for this demo is available on github.

comments powered by Disqus