






G1-C2
Tracking of football players
http://vismod.media.mit.edu/vismod/demos/football/tracking.htm
The Effect
Football players are tracked and boxes are placed around them. This data is then used to create trajectories for all players. Data from actual teams, much like the data being used in class, is the input.
How We Think it Was Done
We think the effect was done by running blob detection and field/background subtraction on each frame. Then, each blob could be tracked from frame to frame using its characteristics, like average color, velocity, direction, and position, to determine the path of the blob. These paths are then used to show what each player does during a play.
How It Was Actually Done
First, a model of the field is created. Then, frame by frame, the input video is transformed to match to the overhead view of the model. This is done by first using Canny edge detection and a line finding algorithm and then matching the line/hashmark features to the model through a perspective 4 point transform.
Next, each object is made to correspond to a "closed-world." These closed worlds are manually initialized to contain a defining feature or features, such as a logo, hash mark, or yard line, as well as one to several players. This will reduce the complexity of the problem because there is less data to analyze and less interaction needed between players. The closed world regions are then isolated frame by frame using player motion blobs. Once each closed world is determined, field items, such as numbers, grass, lines, etc. are removed to isolate the player. This data is then used to track the players path.
These paths are then used to create 3D cylinder men on a plane. Their movements are fed through a bayesian network to determine their actions throughout the play.
Alternate Ways of Doing the Effect
- Tracking may be more accurate with a high definition camera.
- Tracking would be improved by using a stationary camera attached to the roof of the stadium so that no transformations to the model of the field have to be made. Then it would be easier to do a simple background subtraction. As stated by the creators of the project, lens distortion created by zooming cameras prevented them from using background subtraction as an option
- Manual tracking frame by frame.
How This Applies to Our Project
This applies to our project because we will also be using a sideline view and a 2D model of a basketball court. We are then using this to find the 2D position of a player when he/she takes a shot, much like how the above video shows when the QB throws a pass. The positions of the shooter will be marked on the virtual court after a shot is taken. We are also doing tracking of players and placing indicators below them, however it will be using a true stationary camera.
Link to this Page
- DVFX 2007 Groups last edited on 2 May 2007 at 12:51 am by coppola-win.cc.gatech.edu