A Portable, Inexpensive Point-Tracking System for Validation of

Wearable Biomechanics Sensors

G. P. Bailey and R. K. Harle

Computer Laboratory, University of Cambridge, William Gates Building, 15 JJ Thomson Avenue, Cambridge, U.K.

Keywords:

Stereo Vision, Ground Truth, Wearable Sensors, In-Field Assessment, Running, Gait, Foot Kinematics,

Continuous Sensing.

Abstract:

In-ﬁeld validation of the accuracy of wearable sensors is desirable since algorithms that perform well in a

laboratory setting may not perform as well in real-world use. However, the use cases can be challenging. For

example, a foot worn wearable designed to measure foot trajectory should expect to be used in a variety of sce-

narios ranging from straightforward (running track) to challenging (a woodland area with many undulations).

Typically the more challenging the scenario the more difﬁcult it is to get ground truth with conventional

systems. We describe a low-cost, highly-portable, point tracking system that can be used where space and

infrastructure is limited. The system is built around a pair of commodity video cameras in a stereo setup.

We demonstrate how to conﬁgure the cameras, a novel technique to approximate shutter synchronisation to

sub-frame interval, and we benchmark the system indoors against gold-standard motion capture systems. For

a runner 3 m from the cameras were able to recover their foot trajectory with a mean spatial deviation of

1.7±1.1 cm.

1 INTRODUCTION

Recent advances in wearable sensing have driven

an interest in in-situ measurement of athletic perfor-

mance. For the novice or amateur, wearable sen-

sors promise much, including automatic diagnosis of

faults in technique; motivation from quantitative per-

formance metrics available at every session; and sup-

port for rehabilitation when injured. There is value

even at the elite level—where detailed technique and

performance analysis is already commonplace—by

taking assessments out of the laboratory and into the

training or competition venue, moving them into the

background and making them more frequent.

A considerable challenge for the development of

these sensors is validation outside the laboratory. Our

motivation for this work is a foot-mounted inertial

sensor that is able to track the gait of runners to high

accuracy (Bailey and Harle, 2014a; Bailey and Harle,

2014b). The gold-standard measurement setups for

gait typically involve expensive 3D motion capture

systems such as Vicon. Unfortunately these systems

are bulky, hard to conﬁgure, only enable capture in a

small volume, and typically struggle outdoors, mak-

ing them laboratory-bound. Gait is therefore assessed

in the laboratory, with the athlete on a treadmill—

an approach we have previously used to evaluate our

wearable system (Bailey and Harle, 2014a; Bailey

and Harle, 2014b; Bailey and Harle, 2015).

However, the laboratory is often a poor simula-

tion of the real training environment. Treadmills limit

movement and force speeds, uneven terrain and gra-

dients are not considered, changes due to fatigue are

rarely captured due to short test durations, and the

athlete is acutely aware of being assessed, potentially

causing a change in behaviour. Because of these dif-

ferences, a successful evaluation of a system in the

laboratory does not necessarily imply the same—or

even similar—success will be achieved outdoors. Un-

fortunately it is all too easy for a system to produce an

erroneous but plausible result. This is not dissimilar

to the situation with general ﬁtness trackers that base

activity levels on step counting. In the laboratory the

step counters are very accurate because the tests are

typically constrained and contrived. In the real world,

step counting errors of 20% or more have been regu-

larly reported. However, without a ground truth, users

of such systems just trust the number they are given

because it has a believable order of magnitude.

In this paper we describe the development and

evaluation of a low-cost, highly portable point track-

ing system—i.e. a system capable of recovering the

Bailey, G. and Harle, R.

A Por table, Inexpensive Point-Tracking System for Validation of Wearable Biomechanics Sensors.

DOI: 10.5220/0006082601150123

In Proceedings of the 4th International Congress on Sport Sciences Research and Technology Support (icSPORTS 2016), pages 115-123

ISBN: 978-989-758-205-9

115

3D trajectory of a point within some deﬁned volume.

Our system functions outdoors and and serves as a

valuable in-ﬁeld ground truth for wearable sensors.

The system was designed to facilitate the evaluation

of our foot tracking inertial sensors, and we demon-

strate its value in this context.

2 BACKGROUND

Video systems are widely used in sports analysis as

they are easy to use and provide excellent qualitative

data. Quantitative data can be extracted from single

camera video systems by calibrating the video, how-

ever measurements must take place in a single plane

in which a calibration object is present. This can be

problematic for sports analysis as measurements may

not always occur within the correct plane—many ath-

lete motions are non-planar—leading to measurement

error. However, a calibrated stereo vision system is

able to estimate non-planar motions in 3D. The cam-

eras must be carefully calibrated to allow triangula-

tion of corresponding points in the two camera frames

(Zhang, 2002; Hartley and Zisserman, 2004; Heikkil

and Silv

en, 1997). This research area is very mature

and robust stereo vision software is easily obtained:

Matlab’s Computer Vision Toolkit and OpenCV are

two popular choices.

Stereo vision is not without its issues, however.

Calibration is labour intensive, the capture volume is

relatively small (wide angle lenses allow for greater

capture volumes, at the cost of accuracy), and the two

cameras must be synchronised. The level of synchro-

nisation required depends on the desired positioning

accuracy and the expected speed of the object being

tracked. High-end systems typically synchronise to

milliseconds or better using a wired synchronisation

signal.

Parallel work in wearable sensors has aimed to

keep the portability of video systems while mitigating

its limitations. For example, investigation of inertial

sensors for foot tracking has been investigated previ-

ously for walking (Mariani et al., 2010; Sabatini et al.,

2005) and running (Bailey and Harle, 2014a; Bailey

and Harle, 2014b). We have previously proposed that

such systems should constitute an always-on, in-situ,

wearable system for gait assessment in runners. How-

ever, these systems have typically been evaluated in a

lab-based environment, either on a treadmill, or in a

contrived overground environment since the methods

of assessment are not easy to use in more challeng-

ing environments. However, there is a question as to

how such systems would perform in real world envi-

ronments. A previous study attempted to assess iner-

Figure 1: GoPro Hero4 cameras and jig. The distance

between the cameras is 0.25 m.

tial foot mounted trajectory recovery in a track en-

vironment (Bichler et al., 2012). The results were

poor for some spatial metrics. The reason for this

is unclear, although the authors suggest the particular

video reference system used was not robust or accu-

rate enough.

3 A LOW-COST PORTABLE

STEREO VISION GROUND

TRUTH

We sought to develop a lab-validated video reference

system to make ground truth measurements in real-

life scenarios. The requirements for our stereo vision

motion capture system were:

• a fast shutter speed (to minimise motion blur);

• a high frame rate (≥100 frames per second (fps)

to capture enough detail);

• a large depth of ﬁeld (to prevent the need for con-

stant refocusing);

• a wide angle lens (to capture more of the run); and

• a small form factor that is robust and easily oper-

ated.

We use two GoPro Hero4 consumer cameras cap-

turing 1080p (1920×1080) video at 120 fps. We

mounted them to a custom rig as shown in Figure 1.

The mounting plates were ﬁxed to the wooden plat-

form, giving an inter-camera spacing of 0.25 m.

A sample pair of images is provided in Figure 2.

The conﬁguration allows capture of up to two typical

strides with the athlete approximately 3 m away. The

GoPro cameras use a wide angle lens to get this ﬁeld

of view, which introduces signiﬁcant radial distortion.

We account for this in the calibration step.

3.1 Camera Calibration

The system must be calibrated to establish both the

intrinsic camera parameters (e.g. focal length, distor-

tion, etc.) and the extrinsic parameters (camera posi-

tions and poses relative to each other). This is a key

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support

116

(a) Left (b) Right

Figure 2: A pair of sample frames from the stereo camera system.

(a) Original (b) Undistorted

Figure 3: An original frame with an undistorted frame.

problem in computer vision and we apply a standard

solution using a checkerboard pattern held at different

positions and orientations in the ﬁeld of view of the

cameras. Once a sufﬁcient number of images of the

checkerboard are captured, calibration algorithms can

be used to ﬁnd the necessary parameters. Since we did

not require instantaneous trajectory feedback, we ran

the calibration software (Matlab’s Computer Vision

Toolkit, http://uk.mathworks.com/products/computer-

vision/) post-hoc.

To facilitate fast in-ﬁeld calibration we recorded

continuously while the checkerboard was moved

around. We then extracted a series of stills to be used

for calibration. To avoid camera synchronisation er-

rors inﬂuencing the calibration (see section 3.3), the

checkerboard was held steady for a couple of seconds

at each position where a still was to be taken. Do-

ing so ensured that even a poor synchronisation (e.g.

out by multiple frames) will not adversely impact the

calibration.

Accurate estimation of the extrinsic parameters is

key to an accurate tracking result. However, the cam-

eras had to be removed from their housings to down-

load data. Therefore we took a new calibration se-

quence with each new measurement session.

Table 1 lists a typical calibration output and Fig-

ure 3 shows a captured image and its undistorted out-

put as an example. We note that the two cameras

feature similar intrinsic parameters, as expected for

mass-produced hardware. The estimated position of

camera 2 relative to camera 1 (247.6826±0.3813 mm)

closely matches our manually-measured 250 mm.

3.2 Determining Trajectory

Once the cameras are calibrated and synchronised

(see Section 3.3) images can be undistorted and

paired. The next step is to identify the 2D pixel co-

ordinates of the point(s) being tracked in each image

frame. If each point exhibits high contrast to its sur-

roundings (e.g. a brightly coloured sticker) this pro-

cess can be automated using standard computer vision

techniques. In this work we were tracking a single

point post-hoc so we preferred manual point identiﬁ-

cation to avoid any error that may be introduced by

an algorithm. The ﬁnal step is to triangulate the 3D

position of the point relative to (arbitrarily) camera 1.

We used the Matlab toolkit’s triangulate function

for this computation.

3.3 Camera Synchronisation

For successful output from a stereo vision system, it is

important to synchronise the two camera shutters, or

to know the period between the two shutters ﬁring—

the shutter offset. Full shutter synchronisation is hard

A Portable, Inexpensive Point-Tracking System for Validation of Wearable Biomechanics Sensors

117

Table 1: Example calibration results.

(a) Intrinsic Parameters

Parameter Camera 1 Camera 2

Focal length (pixels) 897.24±3.56, 898.41±3.52 894.80±3.51, 895.24±3.49

Principal point (pixels) 991.14±3.79, 552.04±3.08 968.52±

3.82, 539.68±2.96

Skew 2.81±0.61 2.74±0.57

Radial distortion -0.265±0.0020,

0.099±0.002, -0.020±0.001

-0.268±0.002,

0.103±0.002, -0.022±0.001

Tangential distortion -0.0000±0.0003, -

0.0005±0.0002

-0.0000±0.0003, -

0.0002±0.0002

(b) Extrinsic Parameters (Relative to Camera 1)

Camera 2 Parameter Value

Rotation 0.0234±0.0004, -

0.0026±0.0009, -

0.0067±0.0001

Translation (mm) 247.68±0.38, -4.31±0.26, -

6.45±1.08

to achieve without a physical wire between cameras,

which is not available on the Hero4 (or almost any

consumer-grade camera). Wireless synchronisation is

offered on the Hero4 via its WiFi radio, but this is not

intended for accurate synchronisation—we observed

offsets of multiple frames in our tests. This is ac-

ceptable for the intended use—creating stereoscopic

video—since our visual systems cannot perceive a lag

of a few frames at 120 fps. For motion tracking, how-

ever, tighter synchronisation is necessary, as we show

here.

At 120 fps, the inter-frame period is a little over

8 ms. At its fastest point, a typical jogger’s foot will

move approximately 7 cm in this time. For a camera–

runner distance of approximately 3 m and a camera

separation of 0.25 m as we envisage here, the ge-

ometry of the angulation is not forgiving of such an

error—see Figure 4. If the foot is at point A when

the ﬁrst (topmost) camera captures an image and at

point B (7 cm away from A) when the second camera

captures, the triangulation will predict a position at C.

This is 37 cm away from the true position in this ex-

ample. The observed difference will depend on where

in the frame the foot is, but errors of this magnitude

are not unexpected for the geometry we describe. If

we assume a maximum foot speed of 10 ms

−1

then

we need a synchronisation of

s to ensure the foot

does not move d m between frames. For d =5 mm

this requires synchronisation of 0.5 ms or better—

i.e. sub-frame synchronisation (or equivalently much

Plane

of foot

Figure 4: A scale drawing of the geometry for cameras

0.25 m apart and 3 m from the runner.

faster framerate) is necessary to give reliable depth

perception.

In order to estimate the shutter offset, we ﬁrst as-

sume it to be constant—i.e. that the shutter clocks

do not drift signiﬁcantly during an experiment. This

is a reasonable proposition for such short durations,

validated by the results we present in section 3.5.2.

Two challenges remain: to determine the shutter off-

set value; and to estimate the marker position at the

appropriate point between frames.

Addressing the latter ﬁrst, we assume that the

movement between two consecutive frames can be

approximated by a straight line and we linearly inter-

polate the position from the measured positions of the

marker in the two frames in camera 2 that are either

side of the shutter event in camera 1.

To estimate the shutter offset, we have already

established that the 120 Hz sampling of the video

is insufﬁcient. However, the 48 kHz audio streams

can be synchronised to tens of microseconds by

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support

118

cross-correlating the signal made by a synchronisa-

tion sound. Unfortunately, we found the synchro-

nisation between the video and audio streams is it-

self only approximate. This is common on mass-

market consumer cameras; the brain cannot perceive a

video/audio offset below approximately 20 ms. There

is thus little motivation to increase device complexity

and cost to achieve synchronisation better than 2–3

frames.

Our novel synchronisation technique is based on

the observation that a given erroneous shutter offset

will give larger depth estimates for faster moving ob-

jects for motion parallel to the camera sensor (‘hor-

izontal’ motion). i.e. the depth estimate is corre-

lated with the horizontal speed of the object when the

shutter offset is incorrect. When the shutter offset is

correct, we expect no correlation between horizontal

speed and depth estimate.

To exploit this we record an object accelerating

horizontally at approximately constant depth from the

camera. We use a simple pendulum, although any pla-

nar motion will do so long as it has sufﬁcient velocity

and acceleration range, ideally incorporating a period

of zero velocity. We then consider a sequence of off-

set values from -0.5 f to +0.5 f increasing in units of

0.01 f , where f is the frame interval. For each we

compute the correlation coefﬁcient between the speed

of the motion and the depth estimate. Note that the

true object speed is unobservable without the depth

information. However the ‘image speed‘ (in pixels

per second) is an acceptable surrogate since it is pro-

portional and thus exhibits the same correlation prop-

erties.

Figure 5(a) shows a typical progression of the cor-

relation coefﬁcient as the shutter offset is changed.

The true offset is associated with zero correlation. In

the example shown, this corresponds to 0.18 f . Fig-

ure 5(b) shows a top-down view of the pendulum tra-

jectory for different offset values. Assuming the pen-

dulum moved only in the vertical plane parallel to the

camera sensor, we expect to see a straight line when

the offset is correct—we see that an offset of 0.18 f

did indeed produce the expected result. Note that all

offsets agree on the depth for the extremes of the mo-

tion: this reinforces the observation that the depth of a

stationary object (which the pendulum is at either ex-

treme) is independent of the shutter offset accuracy.

The full details of the shutter offset determination is

given in Algorithm 1.

As an aside we note that the pendulum is, in prin-

ciple, redundant when the runner passes the cam-

era rig parallel to the camera sensors. In this case

limbs will typically exhibit the necessary acceleration

range. In practice, we found many amateur runners

Algorithm 1: extract shutter offset.

input : video1, video from camera 1

input : video2, video from camera 2

output: Shutter offset in (fractional) frames

← extract marker path(video1)

← extract marker path(video2)

f ←

find sync to nearest frame(video1,video2)

v ← differentiate(low pass filter(m

))

o f f set ← -0.5

while (o f f set < 0.5) do

rs ← new array

← interpolate image coords(m

offset)

t ← extract 3d trajectory(m

)

d ← low pass filter(depth(t))

r ← pearson coefficient(v,d)

rs.append(r)

offset ← offset + 0.01

return: f - 0.5 + 0.01×argmin(rs)

Table 2: Median trajectory errors for different step and shut-

ter offsets.

Step no.

Median error (cm)

Previous Nearest Interpolated Next

0 7.6 3.6 1.8 15.1

1 7.9 4.1 1.6 16.6

2 9.3 3.0 2.2 16.7

3 10.2 4.8 2.9 16.7

4 9.1 3.9 2.4 16.1

did not keep their limb motions planar and we had

more reliable results using an explicit synchronisation

process with the pendulum.

Figure 6 illustrates the importance of this syn-

chronisation scheme. It shows the raw trajectories

generated from the stereo vision system for a sample

step using nearest-frame synchronisation and interpo-

lated synchronisation. These steps were recorded in-

doors with Vicon ground truth (dashed red lines). We

see that the error was predominantly in the depth co-

ordinate. This is due to the camera being side-on to

the treadmill.

The interpolated result is also notably closer to

the ground truth. We quantitatively assess the error

by taking the median value of the distances between

corresponding points in the stereo vision and Vicon

trajectories. Table 2 shows the results for a series of

different steps, conﬁrming that the interpolated offset

is at least as good as taking the nearest frame, and

usually signiﬁcantly better.

A Portable, Inexpensive Point-Tracking System for Validation of Wearable Biomechanics Sensors

119

−0.6 −0.4 −0.2 0.0 0.2 0.4 0.6

Hypothetical Shutter Offset (Frames)

−1.0

−0.8

−0.6

−0.4

−0.2

0.0

0.2

0.4

0.6

0.8

Correlation Coefﬁcient

(a) Correlation coefﬁcient variation

0 200 400 600 800 1000 1200 1400 1600

x-coordinate (mm)

−100

−50

100

150

200

y-coordinate (mm)

-0.5

0.18

0.5

(b) Pendulum trajectory (top-down view)

Figure 5: Determining the shutter offset.

Sync point Top-down view Side-on view 3D view

Nearest frame

−600 −400 −200 0 200 400

X Coordinate (mm)

−400

−200

200

400

600

Y Coordinate (mm)

Stereo

Vicon

−600 −400 −200 0 200 400

X Coordinate (mm)

−400

−200

200

400

600

Y Coordinate (mm)

Stereo

Vicon

X Coordinate (mm)

−600

−400

−200

200

400

Y Coordinate (mm)

−400

−200

200

400

600

Z Coordinate (mm)

−400

−200

200

400

600

Stereo

Vicon

Interpolated

value

−600 −400 −200 0 200 400

X Coordinate (mm)

−400

−200

200

400

600

Y Coordinate (mm)

Stereo

Vicon

−600 −400 −200 0 200 400

X Coordinate (mm)

−400

−200

200

400

600

Y Coordinate (mm)

Stereo

Vicon

X Coordinate (mm)

−600

−400

−200

200

400

Y Coordinate (mm)

−600

−400

−200

200

400

Z Coordinate (mm)

−400

−200

200

400

600

Stereo

Vicon

Figure 6: Effect of synchronisation error for an example step (ground truth provided by Vicon).

3.4 Trajectory Smoothing

The trajectory results in Figure 6 demonstrate that

even the interpolated synchronisation trajectory has

noise of a few centimetres in the depth axis. We ex-

pect smooth foot trajectories (as conﬁrmed by the Vi-

con tracks) so we smooth the stereo vision trajectory.

We used a fourth order Butterworth low-pass ﬁlter

with a 10Hz cut-off frequency. This was found em-

pirically to give the best results (Figure 7).

3.5 Accuracy Evaluation

3.5.1 Method

To assess the accuracy of the stereo camera system,

we evaluated it in a lab-based environment. We used

a Vicon motion capture system to record the trajectory

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support

120

−800 −600 −400 −200 0 200 400

X Coordinate (mm)

−300

−200

−100

100

200

Y Coordinate (mm)

stereo

vicon

(a) Raw, unﬁltered

−800 −600 −400 −200 0 200 400

X Coordinate (mm)

−300

−200

−100

100

200

Y Coordinate (mm)

stereo

vicon

(b) Low-Pass Filtered

Figure 7: Raw and smoothed stereo vision trajectories vs Vicon ground truth (top-down view). (Depth (y − axis) exaggerated

for visibility).

of a marker attached to a treadmill runner’s shoe. We

selected a treadmill speed of (3.4 ms

−1

, a typical run-

ning pace). The stereo vision system was used to track

the same marker across ﬁve steps in each trial. As

before a treadmill–camera distance of approximately

3 m was used.

Three runners were recruited and each performed

three runs on the treadmill: one where they were in

the centre of the vision system’s frame; one to the left;

and one to the right (see Figure 8). This resulted in

three trials of ﬁve steps for three camera angles, and a

total of over 3800 3D points for evaluation. To com-

pare the Vicon and vision trajectories, we manually

aligned the two co-ordinate systems and looked at the

co-ordinate errors.

3.5.2 Accuracy Results and Discussion

Table 3 gives the mean and standard deviation of

the errors observed between the Vicon and vision

trajectories. We found a 3D euclidean error of ap-

proximately 2±1 cm. The errors at the edge of the

camera frame (where lens distortion is strongest) are

marginally greater but surprisingly similar.

These errors might be expected to improve out-

doors, where the higher ambient light level permits

a faster shutter speed (the GoPro cameras adapt the

shutter speed automatically). Faster shutter speeds

correspond to reduced motion blur and thus more ac-

curate pixel co-ordinates of markers. Nonetheless, ac-

curacies around 2 cm are of value to verify the op-

eration of in-situ sensors, especially given the higher

portability and signiﬁcantly reduced cost compared to

motion capture systems.

Table 3: Laboratory trajectory errors.

Aspect

Mean error ± Standard deviation (cm)

x-axis y-axis z-axis

3D error

magnitude

Left -0.0±0.8 -0.0±1.1 -0.1±1.6 1.8±1.1

Centre -0.4±0.7 -0.3±1.1 0.3±1.4 1.7±1.1

Right -0.1±0.8 -0.6±1.2 0.3±2.0 2.1±1.4

4 COMPARISON WITH

INERTIAL SENSORS

We have previously used foot mounted inertial sen-

sors to extract spatial (Bailey and Harle, 2014a; Bai-

ley and Harle, 2014b) and temporal (Bailey and Harle,

2015) measurements of running gait. This involves

integrating gyroscope and accelerometer data to form

a strapdown Inertial Navigation System (INS). These

systems require drift to be accounted and corrected

for, using de-drifting techniques (Mariani et al., 2010;

Bailey and Harle, 2014a), a Kalman Filter (Foxlin,

2005; Bailey and Harle, 2014a) or similar.

We ran a pilot study to compare our inertial results

to the stereo vision system. We collected 10 running

trials. During each trial the participant ran through

the camera ﬁeld of view allowing capture of a single

stride with the vision system. The videos were post-

processed to extract 3D foot trajectory by identifying

a coloured marker attached to the shoe and co-located

with the inertial sensor.

The inertial sensor was a modiﬁed Shim-

mer Shimmer3 Inertial Measurement Unit

(IMU,www.shimmersensing.com). We added a

200g accelerometer to the package following the dis-

A Portable, Inexpensive Point-Tracking System for Validation of Wearable Biomechanics Sensors

121

(a) Left Frame (b) Centre Frame (c) Right Frame

Figure 8: Field of view coverage of the three camera angles tested.

covery that the integrated 16g sensor sensor saturates

during a typical running stride (Bailey and Harle,

2014b). We generate individual step trajectories

from the inertial data using a de-drifted strapdown

algorithm (Mariani et al., 2010; Bailey and Harle,

2014a).

After processing we had 10 overground running

steps at a mean running speed of 5.4ms

−1

that were

captured by both the stereo vision and inertial subsys-

tems.

4.1 Results

The inertial system estimates both the foot position

and attitude of the foot during each stride, while the

stereo vision system is limited to foot position. Since

we have veriﬁed both the trajectory and attitude es-

timates in the laboratory, we focus on evaluating the

trajectory under the assumption that a good trajectory

estimate implies correct attitude.

We used spatial metrics to assess agreement be-

tween the two systems. Foot clearance, mean step

velocity and stride length were chosen as commonly

used statistics for gait assessment (Mariani et al.,

2010; Bailey and Harle, 2014a). The results in Ta-

ble 4 show strong correlation coefﬁcients from mea-

surements taken from each system implying that the

stereo vision system is working well and that the algo-

rithms used in treadmill running work as expected for

basic overground running on ﬂat ground. This clearly

does not constitute a full evaluation of the inertial de-

rived metrics, but serves to show the stereo system can

plausibly be used for such a purpose. The full evalua-

tion falls outside of the scope of this paper and is left

as future work.

Table 4: Mean and Standard Deviation of Error.

Measurement Error Correlation

Coefﬁcient

Mean Step Velocity −0.28 ± 0.07ms

−1

0.99

Stride Length −0.1 ± 5.2cm 0.96

Foot Clearance −4.2 ± 2.1cm 0.87

5 CONCLUSIONS AND FURTHER

WORK

We have shown that off-the-shelf commodity video

hardware can be used to track a point using a stereo

camera setup with centimetre level accuracy. This is

difﬁcult due to the issue of synchronisation between

cameras, but our novel approach to camera synchro-

nisation makes it possible to use such a set up for

athletic activities despite the faster motion inherent in

these scenarios. We have used the system to assess

foot kinematics in overground running and compared

them to a previously evaluated wearable system dur-

ing a small pilot study. Further work should use this

system to fully evaluate the ability of foot mounted

inertial sensors to assess spatial metrics in challeng-

ing environments including on slopes or undulating

ground.

ACKNOWLEDGEMENTS

This work was funded by Qualcomm Inc. In addition

to ﬁnancial support the authors would like to recog-

nise the technical contributions of the Advanced Lo-

cation Algorithms group at Qualcomm Cambridge.

REFERENCES

Bailey, G. P. and Harle, R. (2014a). Assessment of foot

kinematics during steady state running using a foot-

mounted IMU. In Procedia Engineering , volume 72,

pages 32–37.

Bailey, G. P. and Harle, R. K. (2014b). Investigation of Sen-

sor Parameters for Kinematic Assessment of Steady

State Running Using Foot Mounted IMUs. In Pro-

ceedings of the 2nd International Congress on Sports

Sciences Research and Technology Support, pages

154–161. SCITEPRESS - Science and and Technol-

ogy Publications.

Bailey, G. P. and Harle, R. K. (2015). Measuring Tempo-

ral Parameters of Gait with Foot Mounted IMUs in

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support

122

Steady State Running. In Proceedings of the 3rd In-

ternational Congress on Sport Sciences Research and

Technology Support, number icSPORTS, pages 24–

33. SCITEPRESS - Science and and Technology Pub-

lications.

Bichler, S., Ogris, G., Kremser, V., Schwab, F., Knott,

S., and Baca, A. (2012). Towards high-precision

IMU/GPS-based stride-parameter determination in an

outdoor runners’ scenario. In Procedia Engineering,

volume 34, pages 592–597.

Foxlin, E. (2005). Pedestrian tracking with shoe-mounted

inertial sensors. IEEE Computer Graphics and Appli-

cations, 25(6):38–46.

Hartley, R. I. and Zisserman, A. (2004). Multiple View Ge-

ometry in Computer Vision. Cambridge University

Press, ISBN: 0521540518, second edition.

Heikkil

a, J. and Silv

en, O. (1997). A Four-step Camera

Calibration Procedure with Implicit Image Correction.

Cvpr, pages 1106–1112.

Mariani, B., Hoskovec, C., Rochat, S., B

ula, C., Penders, J.,

and Aminian, K. (2010). 3D gait assessment in young

and elderly subjects using foot-worn inertial sensors.

Journal of Biomechanics, 43:2999–3006.

Sabatini, A. M., Martelloni, C., Scapellato, S., and Cav-

allo, F. (2005). Assessment of walking features from

foot inertial sensing. IEEE Transactions on Biomedi-

cal Engineering, 52:486–494.

Zhang, Z. (2002). A Flexible New Technique for Cam-

era Calibration (Technical Report). IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

22(11):1330–1334.

A Portable, Inexpensive Point-Tracking System for Validation of Wearable Biomechanics Sensors

123