Measuring Cricket Fast Bowling Workload Using Inertial Measurement Units
Fast bowlers have the highest incidence of injury compared to any other position in cricket. A bowling volume that is too high or too low – measured by the number of deliveries bowled in a session – can increase the chance of injury. Despite the importance and the simplicity of measuring bowling volume, it is rarely done at an amateur level due to the monotonous task of manually counting bowls and the effort required to analyse the data. Bowling volume by itself is also not a true measure of bowling workload, as it does not consider the intensity of each delivery. Bowling workload is now considered a combination of bowling volume and bowling intensity. Due to tactical and motivational reasons, fast bowlers will often bowl at different intensities during a match and in training. There are also variations in bowling technique, run-up speed, and anthropometrical characteristics of players. Therefore, the forces exerted on the body are not constant across players or deliveries.
This thesis explores whether an inertial measurement unit (IMU) can predict bowling volume and different intensity metrics – ball release speed, perceived intensity, and ground reaction forces (GRF). These metrics allow a more comprehensive picture of intensity, as each captures slightly different constructs related to bowling workload. This can provide researchers with a mechanism to determine possible links between workload and injury, leading to a more personalised approach to injury management, and give coaches and players a tool to monitor fatigue and performance.
In Chapter 2, a systematic literature review was conducted to examine methods for activity classification in court and field-based sports using IMUs. A key finding was that machine learning techniques had shown promising results across a range of sports. However, only user-defined algorithms had been used in cricket, meaning the application of machine learning had yet to be tested.
Chapter 3 was the first of five studies to develop a system that could estimate bowling workload. A standard IMU was positioned on the upper back in a training setting, and five different machine learning models were used to estimate bowling volume. When tested against outfield throws, several models achieved an F-score of 1.0, meaning perfect differentiation of bowling versus throwing. The analysis was repeated with several down-sampled datasets (i.e., 250 Hz to 25 Hz) to simulate a low-cost IMU that samples data less frequently. A minimal drop in accuracy was observed (F-score = 0.97).
In chapter 4, bowling intensity was quantified by predicting two metrics, (1) ball release speed, and (2) perceived intensity, using the same IMU as Chapter 3 located on the upper back. The gradient boosting algorithm (XGB) was the most consistent machine learning model for measuring ball release speed (mean absolute error (MAE) = 3.61 km/h at 25 Hz) and the perceived intensity zone (F-score = 0.88 at 25 Hz). The results were again consistent across different sampling frequencies, meaning a range of different IMUs might be able to quantify these bowling intensity parameters, including consumer-grade wearables.
The aim of Chapter 5 was to examine whether an IMU placed on the dominant (bowling arm) and non-dominant wrist (instead of the upper back) could improve the previously observed results in Chapters 3 and 4. For practical application, a research-grade IMU (capable of measuring 100 g) was compared against a consumer-grade Apple Watch (32 g). XGB models had the best results across all bowling volume and bowling intensity measures. A slight improvement was observed compared to the previous study (bowling volume: F-score = 1.0; ball release speed: MAE = 2.76 km/h; perceived intensity: F-score = 0.92). There was no significant difference between the research-grade IMU and Apple Watch; however, IMUs on the dominant wrist classified perceived intensity significantly better than on the non-dominant wrist.
In Chapter 6, another component of bowling intensity was introduced – the GRF experienced during the front foot contact of the delivery. Peak force and loading rate, measured by a force plate, were significantly different across three perceived intensity zones in the horizontal and vertical axes (Cohen’s d range = 0.14–0.45, p < 0.01). When ball release speed increased, peak force and loading rate also increased in the horizontal and vertical axes (ηp2 = 0.04–0.18, p < 0.01). Lastly, moving from high to medium intensity, or medium to low intensity, was associated with a larger relative decrease in GRF compared to the relative decrease in ball release speed. For example, reducing bowling effort from high to medium intensity resulted in a 7–17% decrease in the horizontal GRF compared to only a 5% decrease in ball release speed. This finding could influence bowlers’ strategies during an unlimited overs match, as they could conserve energy and reduce workload with only a small reduction in bowling speed.
Similar machine learning techniques as the previous chapters were used in Chapter 7 to estimate GRF. As earlier research had only used accelerometer data to estimate GRF in other sports, this study also assessed whether the addition of gyroscope data could improve accuracy. Research-grade IMUs were attached to the upper back and bowling wrist. A mean absolute percentage error (MAPE) of 22.1% for vertical and horizontal peak force, 24.1% for vertical impulse, and 32.6% and 33.6% for vertical and horizontal loading rates were observed, respectively. The linear support vector machine model had the most consistent overall results. In general, there were no significant differences between using data only from the accelerometer compared to data from the accelerometer and gyroscope. Although the results were similar to previous studies that estimated GRF, the magnitude of error would likely prevent its use in individual monitoring. However, due to the large differences in raw GRFs between participants, researchers may be able to help identify links among GRFs, injury, and performance by categorising values into levels (i.e., low and high).
It is hoped that the methods explored in this thesis can be used as a foundation for future applications that automatically estimate bowling workload across weeks or seasons. As the access to smart devices is increasing in developing nations, such a system has the potential to reach most of the cricketing population.