{"id":2102,"date":"2016-08-08T10:55:13","date_gmt":"2016-08-08T14:55:13","guid":{"rendered":"https:\/\/blogs.mathworks.com\/steve\/?p=2102"},"modified":"2019-11-01T16:46:02","modified_gmt":"2019-11-01T20:46:02","slug":"pokemon-go-meets-matlab","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/steve\/2016\/08\/08\/pokemon-go-meets-matlab\/","title":{"rendered":"Pok\u00e9mon Go meets MATLAB"},"content":{"rendered":"

Until about two weeks ago, I had given absolutely no thought to the possibility of a relationship between MATLAB and Pokémon Go, this summer's worldwide mobile gaming phenomenon. But then I heard about someone using some basic image processing techniques to hack the game by automating the search for PokéStops.<\/p>

Well, I don't play Pokémon Go and I don't know what a PokéStop is, but I understood the basic idea behind the search automation.<\/p>

Here is a screen shot from the game.<\/p>

url = 'https:\/\/blogs.mathworks.com\/steve\/files\/pokemon-go-screen.jpg'<\/span>;\r\nrgb = imread(url);\r\nimshow(rgb)\r\n<\/pre>\"\" 
(Thanks very much to the people who sent me screen shots, especially Haripriya and Chris.)<\/p>
That geometric pattern of nested blue circles is a PokéStop. Our task is to find those. I'm going to show the basic outline of a solution with these steps:<\/p>
1. Segment the image by color.<\/p>
2. Clean up the segmentation using a morphological closing.<\/p>
3. Compute the areas and centroids of the connected components in the segmentation.<\/p>
4. Choose the largest object.<\/p>
To get a quick idea of how to segment the image by color, I like to use the Color Thresholder, usually with the Lab color space. Here's a screen shot that shows how I have adjusted the a* and b* sliders to pick out the range of colors associated with the PokéStop.<\/p>
\"\" <\/p>
Here's the resulting mask that I saved from the Color Thresholder<\/a>.<\/p>
imshow(BW)\r\n<\/pre>\"\" 
Now let's use morphological closing to turn the PokéStop pattern into a single, connected blob.<\/p>
BW2 = imclose(BW,strel('disk'<\/span>,20));\r\nimshow(BW2)\r\n<\/pre>\"\" 
Next, use regionprops<\/tt><\/a> to find all the connected components and compute their areas and centroids. In recent versions of the Image Processing Toolbox, you can tell regionprops<\/tt> to return the result as a table, which makes the results easier to read.<\/p>
t = regionprops('table'<\/span>,BW2,'area'<\/span>,'centroid'<\/span>)\r\n<\/pre>
\r\nt = \r\n\r\n    Area         Centroid    \r\n    _____    ________________\r\n\r\n     1157     37.43    665.14\r\n    13259    260.56    313.99\r\n        2       217      49.5\r\n        6       259      31.5\r\n        1       265        82\r\n        6     356.5    31.667\r\n       12     387.5      31.5\r\n\r\n<\/pre>
Finally, identify the table row with the biggest area, and get the corresponding centroid.<\/p>
[~,j] = max(t.Area);\r\nlocation = t.Centroid(j,:)\r\n<\/pre>
\r\nlocation =\r\n\r\n  260.5581  313.9893\r\n\r\n<\/pre>
Let's see how well that worked by superimposing the location on the original image.<\/p>
imshow(rgb)\r\nhold on<\/span>\r\nplot(location(1),location(2),'dy'<\/span>,'MarkerSize'<\/span>,10,'MarkerFaceColor'<\/span>,'y'<\/span>)\r\nhold off<\/span>\r\n<\/pre>\"\" 
There you go.<\/p>
Catch 'em all!<\/p>