Chess Game Recognition

:triangular_flag_on_post: The goal of this project is to be able to do a live demonstration where we regocnise the pieces on a chess board and their location using your phone camera. This can be devided into three sub goals:

At least chess game reconstruction with a single camera from a fixed position.
Chess game reconstruction from multiple angels.
Free-hand video chess game reconstruction. (achieved)

Approach

The program takes the follwing steps to convert the video frame to the chess board representation:

Take 1 out of every n frames (n = frames_to_skip)
Size frame down to 640x480
Use retrained yolov8n model to detect corners
Use retrained yolov8s model to detect pieces
Convert image to topdown view of chess board only
Devide the topdown image in 8 equaly sized rows and columns
Use transformation matrix from step 5 to get predictions locations from step 4 in the topdown view
Use grid from step 6 to get the locations (a1, a2,... h7, h8) of the chess pieces
Create FEN string
Use FEN string to detect legal moves (stops wrong predictions from impacting performance)
Update current chess board accordingly

Requirements

All the requirements for this project are listed in the requirements.txt file. These packages can be intalled using pip install requirements.txt.

Live Camera feed

In order to use the camera feed from your phone you have to follow the following steps:

Download an app such as "IP Webcam" for Android or "iVCam" for iOS and Windows
Start a stream in the "IP Webcam app"
Change the location in the helper.get_ip() fuction (make sure the location is added in the helper.py file, or change the url directly)

import helper

video_feed_url = helper.get_ip('home')
cap = cv2.VideoCapture(video_feed_url)

import helper

video_feed_url = "https:// ... /video" # add your ip adress
cap = cv2.VideoCapture(video_feed_url)

Video input

You can also use video input to run/test the program. This can be done by using the following code:


video_path = 'test_videos/VIDM4.mp4'  
cap = cv2.VideoCapture(video_path)

Code will performe well on test_videos/VID1.mp4, test_videos/VID2.mp4 and test_videos/VIDM4.mp4. The other test videos cause trouble for step 3 and 4 of the program.

Training Process

The training set consists of images from multiple angles in multiple lighting conditions. The chessboard used is: https://www.bol.com/nl/nl/p/sincer-magnetisch-schaakset/9300000169267633/?s2a=&bltgh=gP9quxkqp-zszO3Jx3dGSg.2_35_36.37.FeatureOptionButton#productTitle.

Corner dataset

The annotated data set for the corners can be accesed via the following link: https://universe.roboflow.com/uni-d5gj3/chess-corner-points

Pieces dataset

The annotated data set for the chess pieces can be accesed via the following link: https://universe.roboflow.com/uni-d5gj3/chess-piece-normal

Run Instructions

Set up a chess board with 1 King, 1 Queen, 1 Rook and 2 pawns for both black and white. Once video feed is established run main.py. In the standard configuration, as seen in commit named "Demo Version Full", the program should show the topdown view with grid and red dots indicating a predicted piece once we have detected four corners of a chessboard. It shows 2 chessboards ones a valid initial state is found. The window called "Detected Board" shows the predictions of the current frame. This window will show some flikkering as predictions are not fully accurate. The window called "Board" will shows the current accepted board with highlights for the last move.

Inspiration technical approach

James Gallagher. (Mar 10, 2023). Represent Chess Boards Digitally with Computer Vision. Roboflow Blog: https://blog.roboflow.com/chess-boards/

Chess piece normal

Chess piece normal Computer Vision Project