BPTD-Net: A Unified Framework for Player Detection and Ball Trajectory Prediction in Football Matches

Abstract

Introduction: Accurate perception of player locations and ball trajectories is fundamental for tactical analysis and intelligent decision‑making in football matches. Existing studies typically focus on either player detection or event‑level understanding, lacking continuous modeling of ball trajectories, and their robustness degrades under small‑scale ball appearance, dense occlusions, and frequent camera view changes. These issues are particularly severe during fast movements and heavy occlusion, where conventional detection‑and‑prediction pipelines fail to maintain spatio‑temporal consistency.

Objectives: To address these challenges, we introduce BPTD Net (Ball Player jointDetection andTrajectoryPredictionNetwork), which integrates a Multi Scale Contextual Enhancement (MSCE) module and a Motion Consistent Trajectory Predictor (MCTP).

Methods: MSCE leverages cascaded dilated convolutions and spatio temporal attention to enrich features of small or occluded objects, markedly improving the detection accuracy of both players and the ball. MCTP combines state filtering with gated recurrent units to jointly capture short term motion cues and long-term dependencies, refining per frame detections and extrapolating future positions to ensure trajectory coherence and physical plausibility.

Results: Experiments on the SoccerNet Tracking and SoccerTrack Challenge datasets show that BPTD Net improves player mAP by 2.8%, 2.4% and ball mAP by 3.8%, 3.3%, while reducing the Average Displacement Error of ball trajectories by 12.2%, 19.4%, thereby demonstrating strong robustness and practical value across diverse settings.

Conclusions: This study presents BPTD-Net, a unified framework for joint player detection and ball trajectory prediction in football video analysis. By incorporating the Multi-Scale Contextual Enhancement (MSCE) module and Motion-Consistent Trajectory Predictor (MCTP), BPTD-Net effectively addresses challenges such as small-object appearance, occlusions, and dynamic camera shifts. The model achieves notable improvements in detection accuracy and trajectory prediction quality on benchmark datasets, demonstrating its robustness and applicability to real-time football analytics. These findings highlight the potential of BPTD-Net as a practical tool for enhancing tactical understanding and intelligent decision-making in sports scenarios.