ME-BATNet: A Lightweight Multi-Task Edge Intelligence Network for Real-Time Badminton Stroke Recognition and Shuttlecock Trajectory Tracking on Mobile Devices

Mobile badminton video analysis faces three coupled challenges: fine-grained strokes differ mainly in brief local body motions, the high-speed shuttlecock is easily degraded by motion blur and occlusion, and cascaded task-specific models impose excessive latency and energy overhead on mobile devices. To address these issues, this paper proposes a lightweight multi-task edge intelligence network, ME-BATNet. The network uses a shared visual encoder to extract multi-scale features and constructs a badminton-specific skeleton topology and lightweight spatiotemporal graph convolution to jointly model natural skeletal relations, racket-side coordination, and dynamic human-shuttlecock connections. The tracking branch integrates high-resolution heatmaps, sparse template interaction, and motion priors, while stroke events are identified by jointly considering action probabilities, wrist velocity, and shuttlecock trajectory changes. Experiments show that the model achieves a pose AP of 73.8%, an action Macro-F1 of 93.0%, a tracking F1 of 96.3%, and an event F1 of 93.4%, with a center localization error of 2.8 pixels; the INT8 model on Snapdragon 8 Gen 3 reaches 53.5 FPS. These results demonstrate a favorable balance among recognition accuracy, real-time performance, and resource consumption. ME-BATNet is therefore suitable for real-time on-device feedback in mobile coaching, self-training, and lightweight match analysis without relying on continuous cloud inference.