Corbett, David (2021) Temporal analysis of physical and skilled performance in professional Australian Rules football. PhD thesis, Victoria University.

Abstract

Australian Football requires physical and skilled output from its participants for more than ninety minutes of play. In both research and practice, physical output is typically described using aggregate parameters extracted from wearable technologies. Parameters include volume measures (eg., total distance), work rates (volume expressed relative to time, eg. metreage per minute) and output bands, which bin either accelerations or velocity into a smaller number of thresholds. Similarly, skilled output may be described using coaches’ ratings, player rankings and counts of skilled actions, termed involvements. Involvements refer to skilled actions when players are both in possession and not in possession of the ball. These parameters are typically aggregated across pre-set windows, including stints, quarters, and training drills. However, there are periods of altered physical and skilled output within training drills and stints, which are not captured by aggregate parameters. It is also difficult to determine when output meaningfully changes within sessions using these aggregate parameters. Consequently, it is difficult to use aggregate parameters to inform time-based decisions, including substitutions and stint-to-rest, and training drill length prescription. The aim of this thesis therefore was to develop an alternative method to aggregate parameter profiling, which can identify changes— either increases or decreases-- in physical and skilled output within training drills and matches. Study One quantified the relationship between physical output, skilled output and stint duration in elite Australian football matches. Physical output was quantified using aggregate parameters, extracted from Global Navigation Satellite Local Positioning System devices. Skilled output was quantified using individual player involvements. Random effect models showed negative relationships between duration, high intensity running, and involvements per minute. Metreage per minute had a positive relationship with involvements per minute for most players. Three conditional inference trees were computed. These models described the impact of factors, including round (ie., game number within a season) and rotation number, and how individuals react to outputs, along with a general set of thresholds for the data. All models demonstrated a weak relationship between physical, skilled output and time. This suggests that wearable technology data and notational analysis feeds could be analysed differently to improve their use in team sports. Study Two proposed a combined time-series/frequency domain approach to profiling physical and skilled output in team-sport. A binary segmentation change point algorithm was applied to the velocity time-series, collected via wearable technologies of Australian football players during matches. This method overcame the need for pre-set aggregation windows by identifying different segments of physical output through the mean and variability of velocity. Spectral and involvement features were extracted for each segment to describe physical and skilled output respectively. Spectral features were able to describe aspects of output that are not captured using aggregate parameters. For example, spectral kurtosis may describe whether physical output is continuous or intermittent. Between five and seven change points were able to give more insight into physical and skilled output than aggregate parameters, whilst identifying sufficiently different segments of play. Study Three applied the time-frequency approach of Study Two to match profiling in team- sport. This study demonstrated how a time-frequency approach may identify differences in physical output between matches, that are not apparent from aggregate parameters. Additionally, the time-frequency approach was able to identify changes in physical and skilled output within matches. Alongside the change-point algorithm, k-means clustering allowed for segments of movement to be classified through both their time elapsed within a match, and their physical and skilled output. These methods could therefore be used, to increase the specificity of load monitoring and physical activity prescription in team-sports. Study Four illustrated how a time-series/frequency-domain can be applied to physical output to assess the sequence, specificity and difficulty of team-sport training drills. By condensing velocity data from training drills into a similarity metric relative to match segments, a drill sequence resembling physical output at differing points of a match was generated. This study identified challenge points for each drill, where the mean and variance of velocity within training drills changes. The location and features of challenge points varied substantially by drill. Aggregate work rate parameters may therefore misrepresent the influence of training drill length on physical output. Movement paths were further analysed to explore how players accrue total volume measures such as total distance. These movement paths may reveal differences in physical output between training drills to match outputs, despite similar aggregate parameters. This thesis demonstrated how a time-frequency analysis of physical and skilled output may increase the sophistication of match and training drill profiling in team-sport. The methods presented in this thesis can identify periods of high physical output late in a match and the movement paths completed by athletes, with differences in physical output between matches. This information may assist practitioners to identify difficult matches (ie., matches with high physical outputs), without relying on typical aggregate parameters. These methods may also increase the specificity of training drill prescription to match outputs. The methods presented may also inform training considerations that are not addressed with aggregate parameters, including training drill sequence and duration.

Search Google Scholar

Repository staff login