Title: PathCRF: Inferring Possession Paths from Player Trajectories for Ball-Free Soccer Event Detection

Abstract

Even with significant progress in artificial intelligence, the collection of soccer event data continues to depend heavily on time-consuming manual labeling. While previous studies have investigated automatic event detection utilizing both ball and player movements, the scalability of ball tracking is hindered by substantial infrastructural and operational expenses. Consequently, comprehensive data acquisition in soccer is mostly restricted to elite-level competitions, which restricts the wider implementation of data-driven insights within the sport.

To overcome these limitations, this study introduces PathCRF, a novel framework designed to identify on-ball soccer events using solely player tracking information. The approach treats player trajectories as a fully connected dynamic graph, framing event detection as the task of identifying a single edge that represents the current possession state at every time step. To maintain logical coherence within the generated edge sequences, a Conditional Random Field (CRF) is utilized to prevent invalid transitions between successive edges. Both emission and transition scores are dynamically derived from edge embeddings generated by a socio-temporal backbone architecture.

During the inference phase, the most likely sequence of edges is determined through Viterbi decoding. Soccer events, such as passes or ball controls, are identified whenever the chosen edge shifts between consecutive time steps. Experimental results demonstrate that PathCRF generates accurate and logically sound possession paths, facilitating reliable downstream analyses and significantly diminishing the requirement for manual event annotation. The source code can be accessed at https://github.com/hyunsungkim-ds/pathcrf.git.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC