Badminton Swing Analysis Using Lightweight OpenPose with an Inception-Time Module for Temporal Modeling

Abstract

In badminton training and technical and tactical analysis, accurately identifying and standardly evaluating racket swings are of great significance for improving athlete performance and preventing sports injuries. To analyze swing movements. This paper presents an integrated pipeline for badminton swing analysis, combining a lightweight OpenPose variant (with ShuffleNet-V2 backbone, adaptive resolution, and Kalman filtering) with an Inception-based architecture augmented by a temporal convolution branch (termed Inception-Time). The temporal branch specifically captures short-term phase transitions in swings, and a dual-branch output jointly optimizes action classification and normative scoring. The action recognition accuracy of LOP-NGN on the self-built BSD data set and the public SPB data set reached 96.5% and 97.3%, respectively, and the single frame response time was 28ms and 30ms, respectively. In the actual swing analysis scenario, the average absolute errors in elbow, shoulder, and wrist joint angle estimation were 1.2°, 1.4°, and 1.6° under low-speed swing conditions, and 1.5°, 1.8°, and 2.1° under high-speed swing conditions. The system power consumption was approximately 18.5 to 19.3W, the analysis frame rate reached 30~32 FPS, and the scoring error was 4.2~4.5 points. The results show that the method can efficiently and accurately identify and score badminton swings, significantly reducing computational overhead and joint estimation errors, and providing reliable technical support for intelligent badminton training.