Giving

Document Sample
Giving Powered By Docstoc
					Giving off the Right Signals!
            Simon Fothergill
            jsf29@cam.ac.uk




       Third Year Ph.D. Student
        Research Talk
                                        Jesus W1,
               28/04/2008         Head of the River
                                  May Bumps 2007
                                 Outline

• Part I
      Is anyone free to coach an outing at 0530
      tomorrow morning?
      (15 minutes, In preparation for Jesus Graduate Conference, 1700 Friday May 2nd)



• Part II
      The bigger picture & The smaller picture
  Automated coaching of technique
Why?
   –   Improve performance
   –   Avoid injury

   –   Can substitute a coach when not available
         •   Train in squads / boats of 8 rowers
         •   Coaches are busy people (2 weeks here are there)
         •   Expensive (amateur population is large)


   –   Even coaches are fallible!
         •   Subjective
         •   Get blinded
         •   Get tired
         •   Only have one pair of eyes


   –   Not a replacement!
         •   Imitating humans is hard
         •   A coach provides more than a assessment of technique
         •   We still use pencil and paper
         •   A coach is still needed to teach the machine
  Automated coaching of technique
What?

   1.       Provide a commentary on what the athlete is doing
   2.       Judge the quality of the performance
        •       Overall technique
        •       Individual Aspects of technique
        •       Description of what is right and wrong
   3.       Choice and Explanation of how to improve what


   –        Needs to happen retrospectively and during the performance, until muscle memory
            established correct technique.

   –        Correction and Assurance

   –        Precision of quality
        •       2 categories (“Its either right or wrong, now!”) Good or Bad
        •       4 categories (It is a practical scale) Good, Ok, Poor, Bad
           Ubiquitous computing

•   Electronic / Electrical / Mechanical devices
•   Miniature
•   Low powered
•   Wireless communications
•   Processing power

• Sensors
• Wearable
                           Reference: Computer Laboratory, University of Cambridge, SeSaMe Project (EPSRC)
                                         Hello Signals!
•   The World contains signals. What can you do with them?

•   Measure real world phenomena

•   Model the real world using the signals
     –   Content-based Information Retrieval
     –   Automatic itemised power consumption


•   Human body movement can be sensed to give motion data

•   Applications
     –   Medical
     –   Performing arts
     –   Monitoring and rehabilitation
     –   Body language
     –   Sports technique


•   Rowing
     –   Cyclical
     –   Highly technical
     –   Small movements
                                                   Laziness!
•   Modelling sports technique

     –   Traditionally done using biomechanics
           •    Take loads of accurate measurements
           •    Formulate rules concerning kinematics of movement
           •    Work out how fast a boat should be moving


     –   This is not how coaches do it (“That looks right!”)

     –   Why?
           •    Variation
                    –   Human
                    –   Marker placement
                    –   Sensor noise
           •    Amount of biomechanical data
           •    Rules don’t exist or unknown (for some aspects / sensors) (“relaxed”)
           •    Rules are fuzzy (“too”, “sufficient”)
           •    Rules are different for everyone
           •    Rules require formulation


     –   Supervised Machine learning
           •    Rough marker placement
           •    Automatic learning of the quality of a certain technique from labelled examples.
           •    Much easier, if it works!!
Data capture
Data capture
Data capture
    Experiments

  Domain      Sport, Cyclical, Rowing, Indoor rowing


Environment                 An office


Equipment     Concept II Model D Ergometer with PM3


 Markers              Erg frame, seat, handle
                                            Experiments
Experiment 1a : Assurance of new technique
                                            Assurance of new
          Description of performances                aspect of
                                                    technique.

                   Person ID                       8

                 Level of fatigue                Fresh

                      Style                     Natural

                      Rate                      Natural

                     Aspect                    Overreach

                 Score precision                   2

                   Score Min                       0

                   Score Max                       1

                  Score Mean                        -

                 Score Variance                     -

         Score quality relative densities     24x0 : 32x1

                     Scorers                    Simon F

                                            ~90 seconds x 2
        Population size (number strokes)
                                                  performances

         Score per stroke or performance    Per performance

             Leave-1-out correlation              0.98
                                                                 Mean handle trajectory for original performance
               Number false +/-ves                 0
                                                                 Mean handle trajectory when stopped overreaching
    Table : Definition of population of strokes
                             Experiments
Results




                        Mean handle trajectory for original performance
                        Mean handle trajectory when stopped overreaching

Conclusion

        The two consecutive stages in the training sequence of improving technique are distinguishable.
                   Experiments
Experiment 1b : Assurance of new technique
                                                             Assurance of new
                           Description of performances                aspect of
                                                                     technique.

                                    Person ID                       5

                                 Level of fatigue                 Fresh

                                       Style                     Natural

                                       Rate                      Natural

                                      Aspect                    Sewuence

                                  Score precision                   2

                                    Score Min                       0

                                    Score Max                       1

                                   Score Mean                        -

                                  Score Variance                     -

                          Score quality relative densities     24x0 : 32x1

                                      Scorers                    Simon F

                                                             ~90 seconds x 2
                         Population size (number strokes)
                                                                   performances

                         Score per stroke or performance     Per performance

                             Leave-1-out correlation               0.96

                               Number false +/-ves                  0

                   Table : Definition of population of strokes
                             Experiments
Results




                        Mean handle trajectory for original performance
                        Mean handle trajectory when stopped getting the sequence wrong

Conclusion

        The two consecutive stages in the training sequence of improving technique are distinguishable.
                               Bigger picture
•   The “right” signals
     – Correct change in sensors’ environment (correct technique)
     – Suitable sensors whose signals are sufficient to allow a change (correct or otherwise) to be detected

     Part 1
         How to get a model; (Algorithms, 3D motion trajectories, human body motion, phenomena from
         rowing technique ontology)

     Part 2
         Using the model; An attempt to pose and answer questions about the properties or theory of the
         inference procedure.

                        Relationship between fidelity of sensors and fidelity of phenomena at different
                        levels of semantic sophistication

                        Can properties be found to easily check whether some phenonema are possible to infer or
                        not, given the dataset.

                        Optimal sensor placement : Entropy map for the body

                        Predication (What is the perfect rowing technique?)
                Smaller picture
•   Data set
•   Pre-processing
•   Feature Extraction
•   Learning algorithms
                                          Data set
The population over which the algorithms are effective must be as wide as possible.
Population defined using these variables whose values will affect the final trajectories,
but do not describe it.
                   Sport, Cyclical, Rowing, Indoor                                 Normal, {list of aspects}
   Domain                                                  Natural & normal /
                                   rowing
                                                              Exaggerate faults
 Environment                 An office                                            Fresh, Tired (distance, rate)
                                                            Level of fatigue
                 Concept II Model D Ergometer with
  Equipment
                                  PM3                                                   10..40 / natural
                                                         Rate (Min/Max/Mean)
   Markers             Erg frame, seat, handle

                                 Performer, Distribution of score
 The handle trajectory for a stroke need not alter in ways only to do with 1 aspect of the
 technique that happens to be of interest. It is not possible to test all combinations, so a
 representative population is used by taking each stroke as a random sample of that
 persons normal technique at that time.
             Data processing

• Linear interpolation

• Transformation to erg co-ordinate system using PCA
                                                +Y

                                           +X
                                                     +Z


• Segmentation using
  sliding window over
  minima/maxima                    Fixed
                                   Moves
               Feature extraction
                • Ratio of drive time to recovery time
                • Angles between x-axis and principle
Rowing
                components of drive and recovery shapes
               • Wobble (lateral variance across z-axis)
                                                                Invariants
                                                                   – Speed
Cyclical
 Movement       •    Smoothness (of shape and speed)               – Not scale
  Quality

                          • Trajectory distance
                           • Trajectory length
                           • Trajectory height
Abstract
            • Five 1 st and 2nd order moments of the shape in

               the x-y plane (weighted uniformly and with
                          the instantaneous speed)
            Learning algorithms

• Normalised feature vector
• Perceptron
   – Gradient descent
      • Error function: Sum of the square of the differences
• Leave 1 out test                                                     Bias


• Sensitivity analysis                   Feature 0     Weight 0           Bias weight




                                                                  Linear
                                                                  combination
                                                                                        Composite
                                                                                        representation
                                        Feature N    Weight N                           of motion
                        Further Work

• Obtain professional coaches’ commentaries

• Continue to define experiments possible on data currently
  collected.
       • For individuals
           –   Novices: Assurance tests using coached aspects
           –   Novices: Cross-Normal using coached aspects
           –   Experts: Fatigued, At different rates, exaggerating
           –   Novice & Experts, use commentary
       • Cross person
           – Novices have similar faults
           – Use commentary

• Improve algorithms using sectioning over time and domain
                  Questions?

                       Thank you!


• Acknowledgements

  – The Rainbow group, Computer Laboratory, University of
    Cambridge, for the use of the VICON system.

  – Members of the DTG, Computer Laboratory, University of
    Cambridge, for willingly rowing!

				
DOCUMENT INFO