Source: full paper text, not abstract only. This is the conference-length version of the 2019 journal paper; the numbers are essentially identical.
Research theme
Planar (2D) simulation models simplify the body by assuming the left and right hip share one joint centre in the plane of the model, and likewise the left and right shoulders. This paper asks, for the first time, how much accuracy that assumption actually costs in a movement that is deliberately “side-on” — where the hips and shoulders are close to the sagittal plane and their projections approach maximum separation. Cricket fast bowling, javelin throwing and overhead racket shots are named as the movements most at risk.
Method: the same 16-segment angle-driven model of the front foot contact phase, customised to one elite fast bowler, run in three variants and each fitted with its own best-possible viscoelastic parameters (via genetic algorithm) so that the comparison is fair:
- PM — plain planar: hips coincident, shoulders coincident
- HM — non-coincident hip joint centres only
- TM — non-coincident hip and shoulder joint centres, plus a variable-length trunk+head segment to represent side flexion
What they measured
Same four-component score as the 2016 paper:
- Ground reaction force at the front foot (% of peak vertical force)
- Centre of mass velocity at release (%)
- Trunk orientation angle (degrees)
- Ball release velocity (%)
- Combined into one overall RMS difference (%)
Findings
The plain planar model could not match the recordings. Overall RMS difference 8.9%, with the error concentrated in force and ball speed.
Full comparison (Table 1 of the paper):
Component PM (plain planar) HM (hips free) TM (hips + shoulders free) Force (%) 18 12 11 COM velocity (%) 0.2 0.1 0.1 Orientation angle (°) 0.9 1.2 0.9 Ball velocity (%) 3.8 3.2 1.7 RMS difference (%) 8.9 6.4 5.7 Freeing the hips alone fixed most of the force problem (18% → 12%) but barely touched ball speed (3.8% → 3.2%).
Freeing the shoulders as well halved the ball-speed error (3.2% → 1.7%).
Per-direction force errors (Figure 2): horizontal 11.4% → 10.5% → 8.6%; vertical 23.0% → 13.7% → 13.6%. So the vertical force is helped enormously by freeing the hips, and essentially not at all by freeing the shoulders.
Mechanism for the force error: in the plain planar model the arms and legs both hang off an averaged point on the torso, so they sit closer to the trunk than they really are. That misplaces the whole-body centre of mass, which shortens the moment arm between the centre of mass and the centre of pressure. To still match the observed COM velocity and trunk rotation, the model has to invent a different ground reaction force.
Mechanism for the ball-speed error: the massless segments add degrees of freedom to the linkage, giving a better representation of the real chain from ground to hand, and hence a better end-point (hand/ball) velocity.
This is a methodological/causal result about models, established by controlled comparison — not a correlational finding about bowlers.
What this means for video and motion analysis
No direct coaching cue. But this is one of the most directly relevant papers in the whole catalogue for anyone doing or trusting video analysis:
- The number to remember: pretending a fast bowler is flat costs about 18% on force and about 3.8% on ball speed. Those are not rounding errors. An 18% error on peak ground reaction force is the difference between a load number that informs injury conversations and one that doesn’t.
- Side-on video is a projection, and the projection is lossy in a specific, identifiable way — the separation of left/right hips and left/right shoulders in that view. In fast bowling that separation is near its maximum, precisely because the action is side-on. So the assumption is worst exactly where a coach is most likely to be filming.
- Ball-speed estimates degrade less than force estimates. If your pipeline reports ball release speed from side-on footage, the planarity penalty is a few percent. If it reports forces or loads, the penalty is around a fifth. Scale your confidence accordingly.
- The lesson is not “use 3D”. The lesson is “keep the out-of-plane geometry as an explicit input”. Felton kept 2D dynamics but drove the pelvis and shoulder separations from measured data. A practical analogue for a video system: even one extra front-on or behind-the-arm view, used only to recover hip and shoulder separation, recovers most of the accuracy without a full 3D solve.
- Vertical force stays stubbornly wrong (~13.6%) no matter what. The authors are explicit that this residual is not caused by the joint-centre assumption. It is a pin-joint compliance problem. So no amount of better geometry will fix vertical loading estimates from a rigid-linkage model.
Caveats and limits
- n = 1. One elite male fast bowler.
- Front foot contact phase only.
- Angle-driven. The model is fed the recorded joint angles; the comparison is about representational adequacy, not predictive power.
- A 2-page conference abstract. It reports means with no dispersion; the 2019 journal version adds standard deviations and a proper held-out evaluation trial. Cite 2019 for anything load-bearing.
- Both HM and TM still carry ~11–12% force error, so “fixed” means “less bad”, not “good”.
Relationship to other Felton work
- Directly extends Felton & King (2016) by adding the controlled PM/HM/TM comparison that 2016 lacked.
- Superseded by Felton, Yeadon & King (2019), J Appl Biomech 35, 157–163, which is the same study with full statistics, an independent fourth evaluation trial, and the published viscoelastic parameter tables. Where the two differ in presentation, prefer 2019.
- TENSION with 2016: the 2016 paper concluded the massless-segment method “was suitable to reproduce predominately planar movements” without reporting the plain-planar baseline. This paper supplies that baseline and shows the plain planar model was clearly inadequate (8.9% vs 5.7%). The claim in 2016 is not wrong, but it is incomplete in a way that flatters planar modelling.
- Cited as reference [118] in McErlain-Naylor, King & Felton (2021), where it is presented as the current best answer to the planarity problem in torque-driven modelling.