Cricket Activity Detection

Ashok Kumar, Javesh Garg
Guide: Prof. Amitabha Mukerjee
{ashokkrm, javeshg, amit}@iitk.ac.in

Abstract: The use of Optical flow analysis along with the shot boundary detection can greatly help in the analysis of broadcasted sports' videos. In this project we classify different types of cricket strokes played by a batsman during the match. The agent first splits the cricket match video into shots using supervised learning, and finds out when the batsman plays the stroke.Agent then classifies the stroke type using optical flow technique.

Why so much fuss about this?

In the field of computer vision, analysis of sports' videos is one of recently researched topic. In cricket, an important task in activity detection is to classify batsman strokeplay during a match. In our paper, we classify the various cricket strokes played by a batsman into four different directions.
The complexity of the problem is due to the fact that videos are at 25fps and also the broadcasted videos use several cameras with multiple transition effects and action-replays. The completely dynamic environment also adds to the problem and finding any meaningful pattern in the video sequence can be quite cumbersome. Intermittent crowd view, scoreboards etc. also adds to complexity the probelem. The main motivation for choosing this work is that this can be a major step in further developing an automatic commentary system which would be a huge contribution to the field of sports. Some work like ball-start detection and cricket highlights retrieval have already been done in this field. Our approach although presently tested on complete videos, can be easily used with streaming videos also.

Short journey through working:

This requires solving problem of shot boundary detection in a video. We split the video into shots by supervised learning approach using colour histograms. The shot boundaries can be any one of 'cut' or a 'dissolve/fade'. We then classify the video frames into four classes namely, ground, fielder, pitch and other using multi-class SVM. Using this classification, we can find out the video segments in which cricket stroke is played by a batsman. Now we have a basic entity of a few frames where the batsman hits the ball. We do optical flow analysis on this part of the video to determine the direction of the stroke played by the batsman. This takes into account the camera pan and zoom also.

Resourses: