ConfSched: Confidence-Threshold Routing for Efficient Edge-Cloud Activity Recognition

Human activity recognition (HAR) on heterogeneous edge-cloud systems faces a fundamental tension: lightweight on-device models offer fast inference but limited accuracy, while powerful cloud-side models improve recognition quality at substantial resource cost. Static deployment strategies—always using the edge model or always using the cloud model—fail to exploit the complementary strengths of each tier. We present ConfSched, a confidence-threshold scheduling policy that routes each HAR inference sample to the cloud model only when the edge model’s softmax confidence falls below a threshold $\tau$ . We evaluate ConfSched against always-edge, always-cloud, and random-routing baselines on a 6-class, 10,299-sample HAR-structured synthetic dataset, using eight scheduling conditions across five random seeds. ConfSched at $\tau =0.80$ achieves 83.19% accuracy with only 12.3% cloud offload rate on this dataset, outperforming both the always-edge baseline (81.35%) and the always-cloud baseline (82.35%) while using approximately seven times less cloud compute than always-cloud. Paired t-tests confirm that these differences are unlikely to be due to seed-level randomness on this dataset ($p=0.002$ , Cohen’s $d=3.64$ ); we do not interpret them as evidence of dataset-level generalization. A threshold sensitivity analysis characterizes the accuracy-offload Pareto curve and is intended to inform—rather than prescribe—threshold selection in deployment. ConfSched is training-free and incurs negligible routing overhead (0.002 ms per sample). We position this work as a proof-of-concept: validation on real HAR benchmarks (UCI-HAR, PAMAP2) is required before recommending it as a deployment-ready policy.