Modeling the Model Athlete

1
Modeling the Model Athlete

• Automatic Coaching of Rowing Technique

Simon Fothergill, Fourth Year Ph.D. student,
Digital Technology Group, Computer Laboratory,
University of Cambridge
DTG Monday Meeting, 10th November 2008
Based on paper Modelling the Model Athlete
Automatic Coaching of Rowing Technique Simon
Fothergill, Rob Harle, Sean Holden SSSPR08
Orlando, Florida, USA, December 2008
2
Supplementary Sports coaching

• Feedback is vital
• Rowing technique is complex, precise and easy to
  capture
• Good coaches arent enough
• Sensor signals need interpreting
• Biomechanical rules are complex and require
  specific sensors, if they exist at all

3
Pattern Recognition

• Statistical
• Arbitrary features that summerise the data in
  some way. E.g. RGB values, number of X
• Structural
• Consider constituent parts and how they are
  related. E.g. contains, above, more red
• Combination
• Distance
• Shape moments / smoothness

4
System overview
Individual aspect of technique
Good

• Population of
• strokes

Stroke quality classifier
stroke
Bad
Motion capture system Lightweight markers
Preprocessing of motion data
Feature extraction
Classification
5
Motion capture

• Bat system
• Inertial sensors
• Optical motion capture
• VICON
• Nintendo Wii controllers

6
Preprocessing

• Compensate for occlusions
• Transform to the erg co-ordinate system defined
  by seat
• Segment performance into strokes using handle
  trajectory extremities

7
Feature extraction

• Art
• Modified various algorithms until found a good
  one for a set of strokes where each stroke is
  obviously different in over-all quality.

8
Abstract features

• Length
• Height
• Distance
• Shape moments (?11, ?12, ?21, ?02, ?20)
• Speed moments (µ 11, µ 12, µ 21, µ 02, µ20)

?(s)
9
Physical Performance features

• Wobble (lateral variance)
• Speed smoothness
• µ-subtract,
• LPF (3Hz)
• dS/dt/dt,
• ?
• Shape smoothness
• LPF (6Hz),
• dS/dt/dt,
• gt threshold (0.4ms-2)

10
Domain features

• Ratio (drive time recovery time)
• Drive and recovery angles

11
System overview
Individual aspect of technique
Good

• Population of
• strokes

Stroke quality classifier
stroke
Bad
Motion capture system Lightweight markers
Preprocessing of motion data
Feature extraction
Classification
12
Machine learning

• Normalisation and Negation
• Each features values are normalised to roughly
  between 0 and 1
• Highly negatively correlated features are negated
• Good strokes are scored as 1
• Bad strokes are scored as 0

13
Machine learning

• Classification

Method 1 Moore-Penrose F w s (F-1
Moore-Penrose pseudo-inverse of feature
matrix) Method 2 Gradient descent Error
function Sum of the square of the
differences weights initialised to 0 750
iterations 0.001 learning rate
14
Machine learning

• Validation of models
• Training repeated using populations formed by
  leaving out different sets of strokes
• Unseen strokes are then classified
• Each stroke left out exactly once
• Multiple performers (each performer left out)
• Sensitivity analysis
• Threshold computed to minimise misclassification
• Features
• Iterations

15
Empirical Validation

• Population
• Six novice, male rowers in their mid-twenties
• 60kg and 90kg
• Very little or no rowing experience.
• Not initially fatigued, comfortable rate,
  uncontrived manner.
• Scoring
• Single expert (coach)
• Score whole performances (95 representative)
• Bad Expert considers a significant floor in
  technique
• Good Expert considers a noticeable improvement
• Experimental method
• Basic explanation
• Give performance (30 strokes)
• Repeat to fatigue
• Identify fault
• Teach correction
• Give performance (30 strokes) whilst coach helps
  to maintain improved technique (for accumulating
  aspects)

16
Empirical Validation

• For an Individual and specific aspect
• Training just that single aspects
• Recognition of that single aspects with realistic
  combinations of different qualities for different
  aspects

17
Empirical Validation

• Across Individuals

18
Discussion and Conclusions

• Useful features
• ?02, ?20 µ02 and µ20 used in at least 90 of the
  final feature sets for both algorithms.
• Comparison of techniques
• For single athletes, gradient descent not as
  fast
• For multiple athletes, gradient descent more
  reliable
• Encouragingly low misclassification
• Suggets inter-variation from different athletes
  gt athletes intra-variation

19
Further Work

• Characterisation of the process
• Population
• Domain
• Algorithms
• Reversing the models to allow prediction of
  optimal individual aspects of technique that can
  be merged to an optimal technique for an
  individual

20
References

• Modelling the Model Athlete Automatic Coaching
  of Rowing Technique Simon Fothergill, Rob Harle,
  Sean Holden SSSPR08 Orlando, Florida, USA,
  December 2008
• Ilg, Mezger Giese. Estimation of Skill Levels
  in Sports Based on Hierarchical Spatio-Temporal
  Correspondences. DAGM 2003, LNCS 2781, pp.
  523-531, 2003.
• Murphy, Vignes, Yuh, Okamura. Automatic Motion
  Recognition and Skill Evaluation for Dynamic
  Tasks. EuroHaptics 2003, 2003.
• Gordon. Automated Video Assessment of Human
  Performance. J. Greer (ed) Proceedings of AI-ED
  95. pp. 541-546, 1995.
• Rosen, Solazzo, Hannaford Sinanan. Objective
  Laparoscopic Skills Assessments of Surgical
  Residents Using Hidden Markov Models Based on
  Haptic Information and Tool/Tissue Interactions.
  The Ninth Conference on Medicine Meets Virtual
  Reality, 2001.
• Joint IAPR International Workshops on Structural
  and Syntactic Pattern Recognition and Statistical
  Techniques in Pattern Recognition (SSSPR 2008)
  Orlando, Florida, USA, December 4-6, 2008
  (http//ml.eecs.ucf.edu/ssspr/index.php)
• 19th International Conference of Pattern
  Recognition, ICPR 2008 (http//www.icpr2008.org/)
• Computer Laboratory, University of Cambridge
  (www.cl.cam.ac.uk)

21
Acknowledgements

• Professor Andy Hopper
• Dr Sean Holden
• Dr Rob Harle
• Dr Joseph Newman
• Brian Jones
• Dr Mbou Eyole-Monono
• The Digital Technology Group, Computer Laboratory
• The Rainbow Group, Computer Laboratory

22
Thank you!

• Questions?