Automated Assessment of Kinaesthetic Performance in Rowing Simon Fothergill Ph.D. student Digital Technology Group, Computer Laboratory, University of Cambridge SeSAME Plenary Meeting, 2nd September 2010, Cardiff
Can assessment of kinaesthetic performance be automated? Feedback is fundamental pedagogical mechanism is sport Sense and Optimise Automate to supplement.
Rowing is a novel domain for well known algorithms Capturing Kinetics Collection of Corpora Stroke similarity Identifying Improvements Useful feedback
Synchronised capture of multiple forms of kinetics
Simple, real-time feedback helps fatigued athletes
Post-workout feedback
Rich, flexible source of data Dataset Real and uncontrived Large Representative of the performance High fidelity Synchronised Segmented Data capture system Compatible Equipment augmentation Annotation Security Portable Cheap Physically robust Extensible platform
Reliable, real-world deployment for over 1 year
Stroke Similarity is an important form of feedback Basic and sophisticated forms of feedback Questionnaire (GB Rowing news feed), observations of deployment and coaching sessions, coaches comments Analysis of kinematic trajectories impacts many areas Movement variability profiles as diagnostic tool, could suggest fatigue, higher variability can reduce injury Training may become inefficient if consistency drops off, abnormal behaviour can be detected, similarity to ideal (coach defined) targets can be measured, consistency is a good coarse grain performance metric (for novices). A definition of is arbitrary and subjective Characteristics of motion trajectories Overall or individual aspect Different populati ons of strokes, such as inter and intra athlete
Collection of Corpora is logistically challenging! The number of unsupervised, unselfconscious, and curious athletes with range of skills is limited An online system was used to collect performance annotations from national coaches due to their availability. Judgements for overall performances and the handle trajectories were collected A B relative comparison considered better than scale Video quality considered acceptable given comments Overlay considered better side by side 1000's of strokes were captured 20 expert coaches (national and international GB Rowing and CU(L)(W)BC) each gave from about 30 minutes to 3 hours.
Capturing expert opinions on forms of similarity
Evaluate known trajectory and shape similarity metrics Classes of algorithms: Difference in distance Difference in duration Difference in moments Difference in outline distance Accumulative error Euclidean distance (binary chop) Hausdorff distance Hausdorff with temporal constraints LCSS DTW (2D, shape matching, truncated) d2 d1 E.g. d1 < d2
Evaluation with limited, subjective annotations
Evaluate known trajectory and shape similarity metrics d2 d1 E.g. d1 < d2 Algorithm weighting += (0.8 * c)
Results : Overall performance, inter-athlete WeightingAlgorithm DurationDifferenceMetric AccumulationOfError NoEndShapeMatching2DTWMetric2D_2.0_5_ WearingOutDTWMetric_2.0_ WearingOutDTWMetric_2.0_ DTWMetric_ / ShapeMatching2DTWMetric2D_2.0_ NoEndShapeMatching2DTWMetric2D_2.0_5_ ShapeMatchingDTWMetric2D_2.0_ NoEndShapeMatching2DTWMetric2D_2.0_5_ ShapeMatching2DTWMetric2D_2.0_ EuclideanDistanceMetric LCSSMetric_1.0_2.0 Percentage agreement with trusted consensus of best algorithm: 76%
Results : Overall performance, intra-athlete WeightingAlgorithm 12.39EuclideanDistanceMetric 11.83LCSSMetric_1.0_ DurationDifferenceMetric 8.24 AccumulationOfError 5.15 Hausdorff2Metric 5.15 Hausdorff1Metric 2.22 NoEndShapeMatching2DTWMetric2D_2.0_5_ ShapeMatching2DTWMetric2D_2.0_ NoEndShapeMatching2DTWMetric2D_2.0_5_ DistanceDifferenceMetric 1.44 SpeedInvariantEuclideanDistanceMetric 1.26 ShapeMatching2DTWMetric2D_2.0_ NoEndShapeMatching2DTWMetric2D_2.0_5_10 Percentage agreement with trusted consensus of best algorithm: 82%
Results : Handle trajectory, inter-athlete WeightingAlgorithm 38.73NoEndShapeMatching2DTWMetric2D_2.0_5_ NoEndShapeMatching2DTWMetric2D_2.0_5_ ShapeMatchingDTWMetric2D_2.0_ ShapeMatchingDTWMetric2D_2.0_ NoEndShapeMatching2DTWMetric2D_2.0_5_ ShapeMatching2DTWMetric2D_2.0_ DTWMetric2D_ DTWMetric_ Hausdorff2Metric Hausdorff1Metric ShapeMatching2DTWMetric2D_2.0_ AccumulativeErrorMetric LCSSMetric_1.0_2.0 Percentage agreement with trusted consensus of best algorithm: 77%
Results : Handle trajectory, intra-athlete WeightingAlgorithm ShapeMatching2DTWMetric2D_2.0_ DurationDifferenceMetric DTWMetric_ NoEndShapeMatching2DTWMetric2D_2.0_5_ LCSSMetric_1.0_ NoEndShapeMatching2DTWMetric2D_2.0_5_ Hausdorff1Metric 9.24 DTWMetric2D_ MomentsDifferenceMetric 8.71 NoEndShapeMatching2DTWMetric2D_2.0_5_ Hausdorff2Metric 7.82 ShapeMatchingDTWMetric2D_2.0_ ShapeMatchingDTWMetric2D_2.0_ EuclideanDistanceMetric Percentage agreement with trusted consensus of best algorithm: 57% (Duration difference = 59%)
Summary & Discussion Overall Performance similarity Inter-athlete: DurationDifferenceMetric (76%) Intra-athlete: EuclideanDistanceMetric (82%) Handle trajectory similarity Inter-athlete: DTW (NoEndShapeMatching2DTWMetric2D) (77%) Intra-athlete: DTW (ShapeMatching2DTWMetric2D) (57%) Rate is an important aspect of the overall technique Explain no reduction for overall intra-athlete case Euclidean distance – spatio-temporal, (bias towards time) DTW – spatio-temporal, 2D, bias towards shape (sections)
Conclusions The length of the warping path between two handle trajectories from the Discrete Time Warping algorithm is the best of the algorithms investigated to approximate expert coaches judgements of similarity of technique between the corresponding rowing strokes with a reliability of ~60/70%. The overall, summary measures of similarity between whole performances can be told from video recordings can be approximated with reliability of ~70/80%. Sensor systems ; devil in the detail! Collection of large corpora and expert annotation is fraught! Basic and sophisticated forms of feedback have started to be provided using pervading sensors.
Other Work Stroke similarity: More careful consideration of the influence of the trajectory characteristics on similarity to further refine algorithms. Use of more than 3D motion trajectories. Identifying Improvements: Evaluate algorithms based on HMMs using annotations provided using a 4 value Lickert scale of importance an individual aspect of technique is addressed, where the consensus is modelled as a Normal distribution with high disagreement. “Importance addressed” and the aspects of technique were carefully chosen using free, natural english comments on performances provided by expert coaches.
Acknowledgements GB Rowing CUWBC Jesus College Boatclub Jesus College BoatClub Trust Cantabs Boatclub ISEA DTG Rainbow group SeSAME Computer Laboratory Jesus College Andy Hopper Sean Holden George Coulouris Rob Harle Andy Riice Brian Jones Marcelo Pias Salman Taherian Richard Gibbens Andrei Breve Alan Blackwell Joe Newman Andrew Lewis
Relevant Calls for papers Mobisys 2011 CHI 2011 Data Mining Journal ICVNZ 2011 (27 th Sept 2010) Pattern Recognition (Interdisciplinary research struggles against too generic or too broad calls?)
Questions Thank you for your attention. Comments and questions, please!