Classification of team sport activities using a single wearable tracking device

Abstract

Wearable tracking devices incorporating accelerometers and gyroscopes are increasingly being used for activity analysis in sports. However, minimal research exists relating to their ability to classify common activities. The purpose of this study was to determine whether data obtained from a single wearable tracking device can be used to classify team sport-related activities. Seventy-six non-elite sporting participants were tested during a simulated team sport circuit (involving stationary, walking, jogging, running, changing direction, counter-movement jumping, jumping for distance and tackling activities) in a laboratory setting. A MinimaxX S4 wearable tracking device was worn below the neck, in-line and dorsal to the first to fifth thoracic vertebrae of the spine, with tri-axial accelerometer and gyroscope data collected at 100 Hz. Multiple time domain, frequency domain and custom features were extracted from each sensor using 0.5, 1.0, and 1.5 s movement capture durations. Features were further screened using a combination of ANOVA and Lasso methods. Relevant features were used to classify the eight activities performed using the Random Forest (RF), Support Vector Machine (SVM) and Logistic Model Tree (LMT) algorithms. The LMT (79–92% classification accuracy) outperformed RF (32–43%) and SVM algorithms (27–40%), obtaining strongest performance using the full model (accelerometer and gyroscope inputs). Processing time can be reduced through feature selection methods (range 1.5–30.2%), however a trade-off exists between classification accuracy and processing time. Movement capture duration also had little impact on classification accuracy or processing time. In sporting scenarios where wearable tracking devices are employed, it is both possible and feasible to accurately classify team sport-related activities.

Introduction

Objective measurement of sports activities is essential for understanding the physical and technical demands related to sports performance (Aughey and Falloon, 2010). It is also important in evaluating the effectiveness of training programs designed to increase performance as well as those targeting both the prevention and rehabilitation of injury (Neville et al., 2010). Fundamental to furthering these understandings is the need to accurately collect specific information relating to the type, intensity and frequency of activities performed (Carling et al., 2009). Consequently, attempts to improve the techniques related to activity analysis in sports have been made in recent years.

At least partially responsible for these improvements are the considerable developments that have occurred in commercially available wearable tracking device technology. Wearable tracking devices typically integrate multiple sensors (e.g., global positioning system [GPS], accelerometer and gyroscope) into a single, versatile unit often worn on the upper back in a sports vest (Kelly et al., 2012). To date, the majority of research has focused on the GPS sensor contained within these devices to obtain basic descriptors of sports activities, such as speed, distance travelled, and the number of high-intensity efforts performed (Cummins et al., 2013). However, evidence suggests that more detailed analysis can be obtained using the accelerometer sensor (Ermes et al., 2008). Specifically, different activity types can be classified based on the features of the accelerometer signal.

McNamara et al. (2015) developed a bowling detection algorithm for cricket. The researchers found that the algorithm was able to classify cricket bowling more effectively in training than game-play, with a maximum accuracy of 98.1% (training). Kelly et al. (2012) applied support vector machine (SVM) and hidden conditional random field algorithms to automatically detect tackling in rugby. The algorithm was able to consistently classify collisions, with a maximum accuracy of 95%. Similarly, Gastin et al. (2013) assessed the concurrent validity of a manufacturer-developed tackle detection algorithm (Catapult Sports), which was compared against video-replay and coded into three intensity categories. The researchers found a maximum classification accuracy of 78%, with tackled players more accurately detected than the players initiating the tackle. However, during game-play the algorithm was only able to correctly detect tackles 18% of the time. Although these findings are promising, more sophisticated and generalisable sport and activity specific algorithms are required (Gastin et al., 2014).

Mitchell et al. (2013) recently proposed a method using a single accelerometer contained within a smartphone worn on the upper-back, with the aim of identifying seven different sporting activities (stationary, walking, jogging, sprinting, hitting a ball, standing tackle, dribbling a ball). An overall activity classification success rate of 75% was achieved using classification approaches that included SVM, Logistic Model Tree (LMT), and a range of Neural Network/Optimisation type classifiers. With the aim of achieving higher classification accuracy, multiple sensors (i.e., both accelerometer and gyroscope) have also been considered in the literature, rather than a single accelerometer sensor alone (e.g. (Leutheuser et al., 2013; Najafi et al., 2003)). Gyroscopes are insensitive to linear accelerations and gravity, and provide essential information pertaining to the rotational motions of the body during human activity (Kunze et al., 2010). As a gyroscope sensor is typically contained within most wearable tracking devices, this would appear to be a feasible approach to aid in the ability to classify of sporting activities.

Another important methodological consideration in the classification literature relates to the duration over which the activity is measured (movement capture duration) for a given classification algorithm (Trost et al., 2012). The optimal duration will ideally be long enough to capture the entire activity as it occurs, while also being short enough to not include any additional activities (Mitchell et al., 2013). Previous work classifying activity type has extracted features in accelerometer data from movement capture durations as short as 0.1 s (Bulling et al., 2014) or as long as 60 s (Trost et al., 2012). In team sports, however, most activities (sprinting, jumping, tackling etc.) can be performed over much shorter durations. For example, the lowest intensity movement (walking) occurs approximately 1.4–2.2 times per second (e.g. (Peacock et al., 2014)). Therefore, much shorter movement capture durations (e.g. 1.5 s or less) may be required to capture activities in team sports. Further, this may improve classification accuracy of these activities, as more periods are available for training (Mannini et al., 2013).

The aims of this study were threefold. First, to determine whether data obtained from wearable tracking device inputs (specifically, gyroscope and accelerometer sensors) alone or in combination can be used to classify team sport-related activities. Second, to determine the ability of three classification algorithms (LMT, RF, SVM) and movement capture durations (0.5, 1.0 and 1.5 s) for feature extraction to classify activities in team sports. Third, to consider the processing time and data collection burdens associated with these methods and identify the best option for practitioners.