This project aims to create a 3D visualisation of the trajectory of a table tennis ball that is captured by three synchronized cameras from different angles.
Run ball_tracking.m to get the results for ball tracking.
It processes the video files in the Videos/CAM1, Videos/CAM2 and Videos/CAM3 directories.
The ball tracking results obtained for a video are saved in the same directory as the video file. The saved files are:
- background of the image (png file)
- coordinates of the ball for different frame (csv file)
The csv filenames are just the video filenames suffixed with .csv. Eg. Result of video.mp4 is saved as video.mp4.csv.
The png filenames are just the video filenames suffixed with background.png.
E.g. Result of video.mp4 is saved as video.mp4background.png.
Execute remove_tracking_results.py to remove the generated result files.
Execute tracking_error.py to see the accuracies of the generated x and y positions for the table tennis ball in the videos. The resultant csv files from Videos/CAMx/ are compared to the truth files of the corresponding videos in Actual/.
The original annotation .csv file for each camera has 2D (x, y) points corresponding to each frame in the video taken by that camera. All three cameras are not guaranteed to be able to view the ball in each frame, so we need to make sure we triangulate the positions of the ball in a frame only when it can be viewed from all three cameras.
The clean_annotation_files.py script will create CAMx-seqNum_cleaned.csv files in CleanData/. These cleaned files will have only those 2D points corresponding to a frame that has (x, y) coordinates in all the three original .csv files.
Run triangulate_positions.m to get the triangulated 3D positions using the following data:
- camera calibration matrix of each camera
- 2D points of ball on image planes of each camera (from
CleanData/CAMx-seqNum_cleaned.csv)
3D positions of the ball are calculated using Perspective Projection equations. The .csv files containing the 3D (x, y, z) coordinates for each sequence will be written to results/3dpts_seqNum.csv.
Based on the calculated 3D positions, the trajectory of the table tennis ball is displayed, along with the location and orientation of each camera and the table. This will be saved to results/traj_seqNum.jpg for each sequence.
For error analysis, the 3D position data is smoothed using the Savitzky-Golay filter and the standard deviation between the original 3D positions and smoothed 3D positions is calculated and displayed. The results will be saved to results/error_visualization_seqNum.jpg for each sequence.