We present a pose-driven framework for handedness prediction in ice hockey, designed to support reliable player identification in broadcast video. Building on extended 2-D human-stick pose representations, we introduce a lightweight multilayer perceptron (MLP) that classifies handedness from torso-normalized keypoint coordinates and confidence features. To improve stability over time, we further extend frame-level predictions to tracklet-level inference through a signed-confidence aggregation method that amplifies reliable predictions while suppressing ambiguous ones. Experiments on large-scale pose and handedness datasets demonstrate that our approach achieves accurate pose estimation and handedness classification. The resulting system provides appearance-agnostic discrimination of player handedness, which can be integrated to support identity preservation in hockey player tracking systems.
PDF