Post on 13-Jan-2016
transcript
April Status
Provide Rich Data Set
• Consistent Imagery
• Play Metadata
• Ground Truth Player Positions
April Status
Consistent Imagery
• Coaching Video of 2007 GT football Season– The Sideline View was chosen because it showed all players on the field
April Status
Consistent Imagery
• Zooms and Pans to keep all players in view– Tracking and Stabilization Issues
• Cuts to scoreboard after each play– Enables Play Detection– Enables Situation Awareness
April Status
Consistent Imagery
• Hi Resolution Video
• Image Artifacts do exist
April Status
Consistent Imagery
• Distribution– Send a common set of files, using indexing to identify all 886 individual plays
April Status
Play Metadata
• Metadata retrieved from proprietary system– Export metadata for each game as .txt file– Recombine as an MS Excel Spreadsheet
April Status
Play Metadata
• 19 Attributes for 886 Plays
• Appended video frame information to metadata– Create MATLAB functions to recognize scoreboard frames
FORMATIONPERSONNEL
GROUPMOTIONS PLAY CODE
PLAY DESCRIPTION
PASS/RUN RESULT PLAY RESULT GAIN WHO
DEFENSE GAME PLAY# DOWN DISTANCE FIELD POSITION
HASH DRIVE # DRIVE PLAY# DRIVE RESULT
+ = AVI FILE START_FRAME END_FRAME
April Status
Play Metadata
• Understanding Metadata– Playbook– Coaching Assistants
April Status
Play Metadata
• Culled 886 to 189 – MS Excel Pivot Table
• Static Formations and Standard Personnel• Selected the top 40 play descriptions and their top 40 formations
April Status
Play Metadata
• Play labels were too specific (too few instances)
• Created Taxonomy to facilitate play recognition based on categories – Run Plays
• Wide Left• Middle Left• Middle Right• Wide Right• NOTA
– Pass Plays• Roll Out• Drop Back
– Short– Combo
» Smash» Y Curl» Option» CMBK» NOTA Combo
– Deep
• Screen• None Of The Above (NOTA)
April Status
Play Metadata
• Combine Views into an Online Application– Central data location for play review and analysis
April Status
Ground Truth Player Positions
– Built using MATLAB– Trained 7 students for
~35 clicks per frame– Average 109 Frames per
play– 7-15 Reference points
(yard lines, sidelines and hash marks)
– 22 Players plus 2 Officials (clicked on hip)
– 1 Ball (hard to detect)– Run plays tracked until ball
reaches LOS– Pass plays tracked until
receiver determined– 572,250 clicks to date
150 plays109 frames per play 35 clicks per frame
- As of 5/21 83 plays clicked and distributed70 plays clicked awaiting audit36 assigned for clicking
• Manual Tracking Application
April Status
Ground Truth Player Positions
• Ground Truth Data Files
April Status
Ground Truth Player Positions
– Built using MATLAB
– Labels reference points, players, officials and ball
– Check plays for proper labeling and major tracking inconsistencies
– Annotate frames with ball actions
– Audited data files distributed to all members of the CARVE team
• Manual Auditing Application
April Status
Ground Truth Player Positions
• Annotations with player locations in image coordinates
April Status
Ground Truth Player Positions
• Transfer Data to Ortho-Rectified Field– Use reference points to calculate an homography matrix (H) for each frame– Use H to rectify points onto a scaled field diagram
+H+
April Status
Ground Truth Player Positions
• Use Rectified Data as Input for Feature Recognition