Research on an Intelligent Evaluation System for Basketball-Specific Physical Conditioning Training Based on Deep Learning and Multimodal Sensor Data Fusion
Basketball-specific physical conditioning requires repeated high-power actions (jumps, sprints, and changes of direction) executed with stable technique under fatigue. While wearable sensors and video systems are widely available, practical training evaluation remains limited by three engineering bottlenecks: weak cross-device synchronization, brittle singlemodality models, and the lack of deployable algorithms that translate raw signals into coaching-grade quality scores in real time. This paper presents an intelligent evaluation system that fuses multimodal sensing (wearable inertial units, vision-based pose, and physiological signals) with deep learning. We propose (1) EdgeSync, a lightweight event-assisted alignment and qualityflagging module that compensates clock drift and packet loss ondevice; (2) UCT-FuseNet, an uncertainty-aware cross-modal transformer that performs multi-task learning for drill recognition and technique-score regression while remaining robust to missing modalities; and (3) an edge deployment pipeline based on teacher-student distillation and quantization-aware training that achieves low-latency inference on commodity mobile SoCs. Experiments on a basketball conditioning dataset ($\mathrm{N}=48$ athletes, 12 drills, 38.6 h total recordings) demonstrate that the proposed method improves macro-F1 by 3.2-6.1 points and reduces score MAE by $0.8-1.6$ compared with strong early/late fusion baselines, while meeting a 200 ms feedback budget. The system provides actionable feedback for coaches, including repetition segmentation, technique cues, and safety-risk alerts.
Paper
Full text
Research on an Intelligent Evaluation System for Basketball-Specific Physical Conditioning Training Based on Deep Learning and Multimodal Sensor Data Fusion
Semantic Scholar · 2026
Abstract
Basketball-specific physical conditioning requires repeated high-power actions (jumps, sprints, and changes of direction) executed with stable technique under fatigue. While wearable sensors and video systems are widely available, practical training evaluation remains limited by three engineering bottlenecks: weak cross-device synchronization, brittle singlemodality models, and the lack of deployable algorithms that translate raw signals into coaching-grade quality scores in real time. This paper presents an intelligent evaluation system that fuses multimodal sensing (wearable inertial units, vision-based pose, and physiological signals) with deep learning. We propose (1) EdgeSync, a lightweight event-assisted alignment and qualityflagging module that compensates clock drift and packet loss ondevice; (2) UCT-FuseNet, an uncertainty-aware cross-modal transformer that performs multi-task learning for drill recognition and technique-score regression while remaining robust to missing modalities; and (3) an edge deployment pipeline based on teacher-student distillation and quantization-aware training that achieves low-latency inference on commodity mobile SoCs. Experiments on a basketball conditioning dataset ($\mathrm{N}=48$ athletes, 12 drills, 38.6 h total recordings) demonstrate that the proposed method improves macro-F1 by 3.2-6.1 points and reduces score MAE by $0.8-1.6$ compared with strong early/late fusion baselines, while meeting a 200 ms feedback budget. The system provides actionable feedback for coaches, including repetition segmentation, technique cues, and safety-risk alerts.