A COMPONENT BASED INTEGRATED SYSTEM FOR SIGNAL

PROCESSING OF SWIMMING PERFORMANCE

Tanya Le Sage

Sports Technology Institute, Loughborough University, LE11 3TU, Loughborough, U.K.

Paul Conway, Laura Justham, Siân Slawson, Axel Bindel, Andrew West

Wolfson School of Mechanical and Manufacturing Engineering, Loughborough University

LE11 3TU Loughborough, U.K.

Keywords: Swimming, Components, Integrated system, Signal processing.

Abstract: Research presented in this paper details the development of an integrated system, which allowed

presentation of meaningful data to coaches and their swimmers in a training environment. The integrated

system comprised of a wireless sensor node, vision components, a wireless audio communication module

and force measurement technologies. A trigger function was implemented onto the sensor node which

synchronized all of the components and that allowed relative processing of the data. Filtering approaches

and signal processing algorithms were used to allow real-time data analysis on the sensor node.

1 INTRODUCTION

The majority of methods used to analyse swimming

technique are vision-based systems. Quintic is an

example of vision-based software where the analyst

uses a pre-recorded video file and then manually

digitises key occurrences within the recording

(Quintic). The disadvantage of this and other vision

systems are the parallax errors introduced by the use

of video cameras, inaccurate measurements due to

light reflections on the water surface and the large

amount of time it takes to process the data. Manual

digitisation is a time consuming process and does

not allow real-time feedback to the coaches or

swimmers. The process provides limited quantitative

data and requires operator expertise. There is

inherent variability within the results due to the

reliance of human judgement.

Force measurement platforms are an additional

technology used for measuring swimmer

performance. Force data can be integrated with

video data during the block phase (time from the

start trigger to leaving the block) of the dive to

enable more complete analysis.

Accelerometer sensor devices have also been

developed for use in a swimming environment. An

example of this was presented by Davey (2005),

where a system was developed using a tri-axis

accelerometer to monitor stroke technique. Ohgi

used a similar system to measure wrist acceleration

of swimmers (Ohgi, 2002). Both systems used a data

logging accelerometer system to capture the data,

which meant that the data could not be viewed in

real time. These existing systems focus on post

processing that again increases the analysis time

significantly and subsequently coaches are unable to

offer immediate feedback to the swimmers based on

these data. Neither case used a wireless sensor

network (WSN) to allow data to be captured from

multiple swimmers, nor an integrated system to

allow full analysis of the stroke technique.

Research presented within this paper, carried out

at Loughborough University, UK, was concerned

with the development of a component based

integrated system for monitoring elite athletes in the

water. The main results of the initial feasibility study

are presented in this paper. This study considered a

variety of different sensing and measurement

devices and an integrated system was constructed to

capture the data. The integrated system comprised of

a WSN, real time audio communications to the

swimmer, a vision analysis system using real-time

image processing, an underwater camera and a force

measurement platform. The WSN was chosen due to

Le Sage T., Conway P., Justham L., Slawson S., Bindel A. and West A. (2010).

A COMPONENT BASED INTEGRATED SYSTEM FOR SIGNAL PROCESSING OF SWIMMING PERFORMANCE.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 73-79

 SciTePress

Figure 1: Integrated system.

its ability to transmit and feedback data in real-time.

It also allowed multiple swimmers performances to

be analysed simultaneously. Furthermore it was

possible to synchronize the network with other data

capture methods used within the integrated system.

The high-speed and underwater cameras were used

because of their ability to provide the coach with

visual information with regards to the athletes’

performance. This made analysing the data from the

accelerometer and the force platform much simpler.

The audio communications were chosen because

they allowed the coach to feedback information to

the swimmer in real-time, based on the data

gathered.

The WSN was designed with a star topology, a

number of nodes communicated with a poolside

personal computer (PC) via an “Access Point” that

collated the wireless data transmissions from the

wireless nodes and had a hardwired connection to

the PC. A trigger function was implemented onto the

sensor node to allow synchronised processing of

data obtained from all components. A Butterworth

filter and signal processing algorithms to extract the

relevant swimming features were embedded onto the

node which allowed the coach to extract useful data

with regards to each individual swimmer’s

performance in real time.

2 METHODOLOGY

The force measurement system was used to augment

the information available from a high speed camera

and the WSN during the start process. The force

measurement system was comprised of a start

platform instrumented with four Kistler force

transducers (9317B) sampling at 100Hz. The force

measurement platform was used to ascertain the

following parameters:

 Horizontal force

 Vertical force

 Time: to first movement, to back foot leaving,

to front foot leaving, overall block time

 Centre of pressure

A high-speed camera and a WSN were

synchronised with the force measurement platform

using a TTL trigger function. The function was

implemented in the embedded programming of the

node which sent an interrupt to the access point (AP)

when the trigger was enabled. Sending a TTL signal

to a port on the AP triggered the system. The

embedded code initialised the trigger, starting the

trigger on the rising edge of the signal. The

integrated system can be seen in Figure 1 and has

been used to determine the characteristics of an

accelerometer trace based on the data gathered from

the high speed video camera and force plateThe

camera used was a Photron SA1 colour camera with

a 1024x1024 resolution, sampling at 50 frames per

second (fps). Automated vision processing was used

to track wearable LED markers placed on body key

body landmarks, for example, the hip. This was done

via spatial thresholding algorithms, developed in

Matlab. An underwater bullet CCTV camera

sampling at 25 fps was also used. Vision data was

used to supplement the data gathered from the WSN,

SIGMAP 2010 - International Conference on Signal Processing and Multimedia Applications

allowing stroke recognition and real-time analysis of

the accelerometer signal.

For many low-g (<2g) inertial sensing

applications the signal-to-noise ratio is low and thus

any un-modelled error in the physical parameters

undermine the effectiveness of the intended

application over time (Ang, 2004). A common

method to minimize the errors associated with the

accelerometer signal is the use of filtering (see for

example Koukoulas 2005, Jo 2004, Hernandez

2000). For the current system a low-pass finite

impulse response (FIR) filter was implemented to

filter out frequencies greater than a pre-defined

threshold while retaining the low frequency

components (Ketharnavaz, 2005). Filtering also

reduces the errors associated with integration of a

signal, in this case integration of the accelerometer

data in order to obtain velocity and double

integration to obtain position. Edwards (2005)

demonstrated that seemingly small aliased content

could cause appreciable errors in the integrated

waveforms.

The raw accelerometer values were fed into a

real-time Butterworth filter and signal processing

equations, which were embedded onto the node.

This enabled analysis to take place robustly, in real-

time, so that the results could be sent directly from

the node rather than sending raw data. This was

preferable because the raw data file was large and

therefore filled the available bandwidth. A low pass

Butterworth filter was chosen to smooth the data

collected and to minimize the noise components of

the signal. It was chosen over a Chebyshev filter due

to its ability to be implemented in real time and

embedded on the sensor node. Lap count

identification was automatically determined by

setting a low filter frequency on the Butterworth

filter and using a ‘zero crossing’ algorithm. Signal

processing algorithms were developed to analyse

filtered data, including a ‘zero crossing’ algorithm to

determine the stroke durations and stroke rates,

which were identified to be the variables of most

interest to the end users. Pulse analysis of the

filtered data was also calculated and used to

determine the rise and fall times of each stroke.

Circular buffers were used to allow real-time

implementation of the filter and signal processing

algorithms.

A wireless audio communication module was

attached to the swimmer and a UART interface to

the host device was used to configure the module

operation and then transfer data between the host

and the communication end-point via the wireless

interface. Once the devices were connected the

coach used the microphone input to provide

feedback to the swimmer (who wore earphones

attached to the wireless module) on their

performance throughout their training. The module

transmitted wirelessly up to a depth of 10cm over a

distance of more than 50m underwater.

3 RESULTS

Initially the results are used to highlight the

implementation of the synchronised system and data

capture. The filtering technique used on the

synchronised data is then considered. Finally

determination and analysis of the stroke

characteristics from the filtered data are reviewed.

The TTL trigger function embedded on the

sensor node was used to capture data simultaneously

from the force measurement platform, the high

speed video camera and the sensor node. These data

can be seen in Figure 2. The high speed video was

used to supplement the data gathered from the force

platform and allowed determination of the key

points that occurred during the dive, for example,

time of back foot leaving the force platform. The

times on the video were correlated with those of the

accelerometer data, identifying time to entry, the

point where the stroke was initiated and the time at

15m (where the start officially ends). In addition, the

WSN was used to consider elements such as lap

count, stroke rate, stroke duration in free swimming

and to distinguish the different phases of the turn.

Initially a Butterworth filter was used to

ascertain the lap count of the swimmer. Setting a

low filter frequency achieved this. A comparison of

the raw unfiltered data and the real-time embedded

filtered data on 4 lengths of front crawl stroke can be

seen in Figure 3. The largest peak in the data was

identified as the swimmer’s turn at the wall at the

halfway point. By setting a threshold the filter and

signal processing algorithms were used to pick out

the lap count. For these data the four laps were

identified.

Different filter frequencies were required for the

different swimming strokes. In order to retain the

peaks in the breaststroke and butterfly data a higher

cut off frequency was used. The signal processing

technique used for one length of front crawl can be

seen in Figure 4. It was found that pre processed

data could be analysed to establish timing

information, stroke count, stroke durations, rise

times and fall times. This analysis may then be

collated to give an indication of the swimmers

performance. Four 100m trials have been analysed

A COMPONENT BASED INTEGRATED SYSTEM FOR SIGNAL PROCESSING OF SWIMMING PERFORMANCE

Figure 2: Integrated system for starts.

Figure 3: Butterworth filter on 4 lengths of front crawl data.

SIGMAP 2010 - International Conference on Signal Processing and Multimedia Applications

Figure 4: Analysis of the front crawl stroke using video and accelerometer data.

to derive all of the discussed parameters, Figure 5.

Automated timing was found to be within 1 second

of hand timing on average. Hand timing is

undesirable since it is subject to human judgement

and variability and cannot be readily scaled to

support the monitoring of multiple swimmers in

training sessions. Average stroke durations gave an

idea of a swimmer’s typical stroke and provided a

measure to determine if they had changed their

technique.

The underwater video camera was used to

chracterise phases of the turn with the accelerometer

data. The x axes represented the forward motion, the

y axis the roll of the swimmer laterally and the z axis

the vertical movement of the swimmer. On the

swimmer’s approach to the wall the acceleration in

the z axis remained fairly constant. When the

swimmer initiated the turn the z axis rotated through

90 degrees, which meant that the x axis experienced

the major gravity component and the z axis tended

towards zero. When the swimmer turned onto their

back the z component experienced a negative

contribution from gravity. As the swimmer turned

back onto their front the z acceleration returned to

fluctuating about 1g. This process can be seen from

the video and accelerometer data in Figure 6.

A comparison of manual and automatic tracking

was carried out to determine the efficiency of the

automated code. The time it took to analyse one

100m IM manually, i.e. to determine lap count,

stroke rate, stroke duration and rise and fall times,

was approximately 45 minutes. An elite swimmer

swims around 4-6km in a two hour session. If each

length was analysed for the parameters discussed it

would take up to 45hours to analyse one swimmer’s

two hour training session! The embedded coding

enabled these same values to be obtained in real-

time throughout the swimmer’s training session.

A COMPONENT BASED INTEGRATED SYSTEM FOR SIGNAL PROCESSING OF SWIMMING PERFORMANCE

4 CONCLUSIONS

A multimedia system and signal processing

techniques for monitoring swimmer performance has

been presented in this paper. It provides a significant

advantage over current methods used because it

allows results from multiple components to be

integrated and analysed simultaneously in real-time.

The signal processing techniques used on the

accelerometer offer feedback to swimmers in real-

time and parameters are derived automatically on the

sensor node.

5 FUTURE WORK

An inertial navigation system (INS) will be used in

which measurements from embedded accelerometers

and gyroscopes will be used to track the position and

orientation of a swimmer relative to a known

starting point, orientation and velocity. An INS

comprising of a tri-axis accelerometer and a tri-axis

gyroscope, measuring angular velocity and linear

acceleration respectively, will be attached as a

strapdown system to a swimmer. By processing

signals from these devices it is possible to track the

position and orientation of a device (Woodman,

2007). The output of the gyroscope provides the

attitude of the swimmer. Strapdown navigation

equations will be used to combine the accelerometer

and gyroscope data, compensating for the effect of

gravity on the system. The output will then be

integrated twice (once in order to obtain velocity,

and again in order to obtain position).

The results from the IMU will then be fed into an

extended Kalman filter. The Kalman filter combines

noisy sensor outputs to estimate the state of a system

with uncertain dynamics (Grewal, 2007). The noisy

sensors in this research will be INS accelerometers

and gyroscopes. The system state includes position,

velocity and attitude rate of the swimmer. It also

includes the accelerometer and gyroscope biases and

scale factors. The uncertain dynamics includes

unpredictable disturbances of the swimmer, for

example, waves in the water. A GPS receiver may

be used to calibrate the system initially (before the

swimmer enters the building), increasing the

accuracy of the initial error predictions.

The integrated system will be presented in a

graphical user interface (GUI) thus allowing the

coaches and swimmers to visualise the results with

ease, allowing unique insight into the skill and

performance capabilities of elite swimmers.

REFERENCES

Ang, W., 2004. Physical model of a MEMS accelerometer

for low-g motion tracking applications. Proceedings of

the 2004 IEEE International Conference on Robotics

and Automation. New Orleans, LA.

Arellano, R., et al., 2005. A comparison of CMJ,

simulated and swimming grab-start force recordings

and their relationships with the start performance.

Proceedings of XXI International Society of

Biomechanics in Sport. China

Brown, R., Hwang, P., 1992. Introduction to random

signals and applied Kalman filtering. John Wiley &

Sons, Inc.

Davey, N., 2005. An accelerometer-based system for elite

swimming performance analysis. Proceedings of SPIE

the International Society for Optical Engineering., vol.

5649, no. 1.

Edwards, T., 2005. Effects of aliasing on numerical

integration. Mechanical systems and signal

processing, vol. 21. Elsevier.

Galbraith, H., et al., 2008. Biomechanical comparison of

the track start and the modified one-handed track start

in competitive swimming: an intervention study.

Journal of Applied Biomechanics, vol. 24.

Grewal, M., 2007. Global positioning systems, inertial

navigation, and integration. Hoboken, New Jersey.

John Wiley & Sons, Inc.

Hernandez, W., 2000. Improving the response of an

accelerometer by using optimal filtering. Sensors and

actuators A, vol. 88. Elsevier.

Jo, G., 2004. Underwater navigation system with velocity

measurement by a receding horizon Kalman filter.

Seiken symposium conference 38, vol. 3.

Ketharnavas, N., 2005. Real-time digital signal

processing. Bulington, MA, USA. Elsevier.

Koukoulas, T., 2005. Binary low, high and band pass

amplitude filters will full and quantized phase in the

presence of disjoint noise. Lasers in Engineering. Old

City Publishing Inc, vol. 15.

Maglischo, E., 1993. Swimming even faster. Mountain

View, CA. Mayfield Publishing Company.

Mason, B., et al., 2007. A kinematic analysis and

recommendations for elite swimmers performing the

sprint start. Proceedings of XXV International Society

of Biomechanics in Sport. Brazil.

Ohgi, Y., 2002. Microcomputer-based data logging device

for accelerometry in swimming. Engineering of Sport.,

vol.4.

Quintic Consultancy Ltd. Putting sports science into

practice. Available: http://www.quintic.com

[Accessed: 10

October 2009]

Woodman, O., 2007. An introduction to inertial

navigation. Technical Report, no. 696. University of

Cambridge.

A COMPONENT BASED INTEGRATED SYSTEM FOR SIGNAL PROCESSING OF SWIMMING PERFORMANCE