The ultimate goal of any sports coach is to maximise an athlete`s performance at a specific time, usually for major competitions. Previously a coach`s ability to prescribe training has been based on their intuition and years of personal experience. More recently, efforts have shifted towards precise quantification of training loads in an attempt to model performance and therefore prescribe optimal training programs. However, the first step in optimizing training is to be able to easily and accurately quantify an individual athlete`s training. The development of heart rate monitors, personal GPS units and cycling power meters, have provided the necessary hardware to monitor athlete`s training loads in running and cycling based activities, but there is not currently any device which is able to accurately and easily monitor training load in swimmers. Methods A novel hardware device consisting of tri-axial micro-sensors (accelerometers, gyroscopes, and magnetometers) was developed in conjunction with Catapult Innovations. The micro-sensors are enclosed in a waterproof casing and worn by swimmers using a clip system which is attached to their swimming costumes as close as possible to the lumbar spine region. Data is transferred from the units and interpreted using a Swimming Algorithm developed by NICTA using machine learning techniques. The goal of the Swimming Algorithm is to identify and separate swimming and non-swimming related activities. Then within swimming related activities be able to detect each individual stroke of the four main swimming styles (freestyle, backstroke, butterfly, and breaststroke) as well as starts, tumble or lateral turns and finishes. From accurately identifying these events it is hoped the Swimming Algorithm will be able to calculate lap times and instantaneous stroke rates, in addition to stroke identification. The initial Swimming Algorithm was developed using only accelerometer data. Visual records and video footage were used as Gold Standard for stroke identification (550 laps) and lap times (496 laps). Lap times were all determined using video footage. Lap starts were identified as the last video frame a swimmers feet were in contact with the wall or blocks, and lap finishes, the first frame a swimmers hand contacted the wall. Results The Swimming Algorithm was able to correctly identify the swimming stroke used on 543 out of 550 laps. Of the 7 laps not correctly identified, the Swimming Algorithm did not detect a turn on six laps and combined the laps together. One lap of backstroke was not able to be identified. The number of laps of each swimming strokes used in this analysis are listed in Table 1. The Swimming Algorithm calculated lap times (sec) within 1 second of the Gold Standard time on 52% of the laps analysed and within 2 seconds of the Gold Standard time on 77% of laps. Figure 1 illustrates the distribution of differences between the Gold Standard times and those determined with the Swimming Algorithm. Discussion/Conclusion An algorithm developed to automatically quantify swimming training using accelerometer data has been successful in identifying swimming stroke type. Further work, including incorporating magnetometer data into the Swimming Algorithm, will continue in an attempt to improve the accuracy of lap times. Validation of instantaneous stroke rate data will also be performed in the near future. It is anticipated the use of the Swimming Algorithm will have a substantial impact in accurately quantifying swimming training load and in optimising training prescription and swimming performance
© Copyright 2009 National Elite Sports Council 2009 Athlete Services Forum - High Performance Programming for Success - 11-12th November - Satellite Program - Applied Physiology Conference 2009 - Australian Institute of Sport, Canberra - 10th, 11th and 13th November. All rights reserved.