Engineering Features from Raw Sensor Data to Analyse Player Movements during Competition

Abstract

Research in field sports often involves analysis of running performance profiles of players during competitive games with individual, per-position, and time-related descriptive statistics. Data are acquired through wearable technologies, which generally capture simple data points, which in the case of many team-based sports are times, latitudes, and longitudes. While the data capture is simple and in relatively high volumes, the raw data are unsuited to any form of analysis or machine learning functions. The main goal of this research is to develop a multistep feature engineering framework that delivers the transformation of sequential data into feature sets more suited to machine learning applications.

Keywords: feature engineering; machine learning; wearable devices.

Figure 1

Transformation methodology.

Figure 2

Event identification from the raw sensor data.

Figure 3

Left: Frequency of events per duration. Right: Frequency of events per duration ≥ 30 s.

Figure 4

Left: Frequency of events per distance. Right: Frequency of events per distance ≥ 130 m.

Figure 5

Left: Frequency of events per high-speed distance. Right: Frequency of events per high-speed distance ≥ 35 m.

Figure 6

The time series of high-speed distance for the remaining games are merged to form a unique time series. The first half of the tested game is placed in the last position. These data constitute the training set for the predictive experiments.