Player Tracking



Project Pipeline

Phase 0

Data Collection

 
Selected plays are captured from NFL Madden 08 PC version.

Improperly collected videos are discarded.
 

Phase 1

Object Detection

 

Background Substraction

Background is removed using Gaussian mixture models.

Image Smoothening

2-D Gaussian filter is applied to to reduce noise.

Player Position Estimation

Local extema is clustered to approximate player location.

Optimize Estimated Location

Use prior domain knowledge to reject false positives early.
 

Phase 2

Object Tracking

 

Track Estimation

Use Kalman Filtering to estimate tracks as a function of time.

Track Optimization

Use iterative Heurestic search to find best fit tracks.

Compute Track Data

Calcuate distance travelled, Average velocity and Instantaneous Velocity.
 

Phase 3

Data Analysis

 

Estimate Net-Tracks

A bit map is created from the detected tracks.

Create Player heatmap

A normalized heatmap of the player locations is created.

Extract HOG features

HOG features are extracted from the tracks bitmap and player heatmap.

Classify Plays

Use 1-vs-All SVM to classify plays based on features extracted.