CourtMotion: Learning Event-Driven Motion Representations from Skeletal Data for Basketball

TL;DR

The paper addresses the challenge of anticipating basketball events by coupling fine-grained skeletal motion with tactical intent. It introduces CourtMotion, a two-stage framework with a GNN-based skeleton encoder and a Transformer equipped with event projection heads, trained to predict trajectories and basketball events across past, present, and future windows. On NBA SportVU data with skeletal augmentation, CourtMotion achieves a 35% reduction in trajectory prediction error over position-only baselines and shows substantial improvements on downstream tasks such as shot taker, pick, assist, shot location, and shot type prediction, even seconds before events occur. The work demonstrates the value of skeletal motion information and explicit event-aware pretraining for multi-agent sports analytics, offering a solid pretrained foundation for broader analytics and tactical pattern analysis.

Abstract

This paper presents CourtMotion, a spatiotemporal modeling framework for analyzing and predicting game events and plays as they develop in professional basketball. Anticipating basketball events requires understanding both physical motion patterns and their semantic significance in the context of the game. Traditional approaches that use only player positions fail to capture crucial indicators such as body orientation, defensive stance, or shooting preparation motions. Our two-stage approach first processes skeletal tracking data through Graph Neural Networks to capture nuanced motion patterns, then employs a Transformer architecture with specialized attention mechanisms to model player interactions. We introduce event projection heads that explicitly connect player movements to basketball events like passes, shots, and steals, training the model to associate physical motion patterns with their tactical purposes. Experiments on NBA tracking data demonstrate significant improvements over position-only baselines: 35% reduction in trajectory prediction error compared to state-of-the-art position-based models and consistent performance gains across key basketball analytics tasks. The resulting pretrained model serves as a powerful foundation for multiple downstream tasks, with pick detection, shot taker identification, assist prediction, shot location classification, and shot type recognition demonstrating substantial improvements over existing methods.

Figures (10)

• Figure 1: CourtMotion architecture with three main components: (1) GNN Skeleton Encoder that processes 3D joint data into player motion embeddings, (2) Transformer-based trajectory predictor integrating motion with position data, and (3) Event Projection Heads that predict basketball events.
• Figure 2: GNN Skeleton Encoder processes 3D joint data through parallel vertex and edge processing streams with bi-temporal causal convolutions and strategic downsampling.
• Figure 3: Trajectory loss comparison across models at different timesteps, showing consistent performance advantages of our approach over previous methods throughout the sequence.
• Figure 4: Comparison for shot taker prediction across training set sizes.
• Figure 5: AP scores across time horizons for downstream prediction tasks, showing consistent performance advantages of our approach over baselines as predictions approach the target event.