Application research of knowledge distillation and model compression technology in lightweight end-side running posture analysis

To address the trade-off between limited on-device computing resources and the high accuracy required for running posture analysis, this paper proposes an Aerodynamic Gait Analysis Network based on Knowledge Distillation. The model adopts a streamlined architecture based on adaptive bottleneck modules and contains 1.2M parameters, which is lower than MobileNetV3′s 4.2M. Through multidimensional collaborative knowledge distillation, AeroGait achieves a headline validation accuracy of 95.0% on the subject-wise validation split, while the teacher-student knee-angle prediction discrepancy is 0.2°. Network-only inference latency is 12.5 ms under the standard test setting and less than 5 ms on flagship mobile chips. At a parameter-compression ratio of 90%, the validation accuracy remains 88.6%, compared with 65% for the benchmark model, indicating an effective balance between computational efficiency and analysis accuracy.