EdgeFusionNet: real-time multimodal feedback for table tennis training via lightweight cross-modal attention fusion...

Real-time biomechanical feedback during table tennis training demands both low latency and high recognition accuracy, yet existing systems sacrifice one for the other due to cloud-transmission delays and the computational constraints of edge devices. This paper presents EdgeFusionNet, an integrated edge-cloud collaborative architecture that delivers actionable stroke-level feedback within 32 ms under realistic network conditions. At its core, a lightweight cross-modal attention fusion network (LCA-FNet) fuses temporally aligned features from high-speed vision, inertial measurement, and surface electromyography streams through shared-projection cross-modal attention and adaptive channel gating, achieving 93.6 ± 0.4% recognition accuracy (Macro-F1 = 0.927 ± 0.005, mean ± SD over five seeds) across seven canonical stroke types with only 1.48 million parameters, and retaining 90.4 ± 2.3% accuracy under a strict leave-one-subject-out evaluation. A hardware-triggered synchronization mechanism maintains sub-millisecond cross-modal alignment, while a two-level knowledge distillation strategy recovers 97.8% of the cloud-resident teacher model's accuracy after aggressive structural compression. An adaptive computation offloading agent, trained via Q-learning with explicitly defined state, action and reward spaces, dynamically partitions inference between the edge node and cloud server based on prediction entropy and network quality, sustaining sub-32 ms P95 end-to-end latency even at 5 Mbps uplink bandwidth. Field deployment over thirty training sessions with twelve athletes per study arm confirmed ecological validity, yielding 91.8 ± 0.7% accuracy under uncontrolled gymnasium conditions, a Cohen's kappa of 0.874 against the consensus of two expert coaches (whose own inter-coach kappa was 0.892), and a 10.3-percentage-point gain in standardized multi-ball hit rate over a matched control group after four weeks (p < 0.001). These results demonstrate that principled co-design of multimodal fusion, model compression, and adaptive offloading can bridge the gap between laboratory-grade recognition performance and the stringent latency requirements of live athletic training.