DBAS-SLSTM: A Multi-Sensor Fusion Approach for Real-Time Motion Attitude Error Detection and Correction

Abstract

Efficient and accurate posture assessment is essential in physical exercise training to enhance performance, prevent injuries, and support overall well-being. This research aims to develop a Motion Attitude Error Detection System capable of continuously monitoring and correcting human movement using multi-sensor data. The proposed system integrates inertial measurement units, accelerometers, gyroscopes, and vision-based key-point tracking to capture comprehensive motion information. Experiments were conducted using the Motion Attitude Error Detection Dataset, consisting of 4,800 labeled motion sequences representing correct and incorrect postures. Raw sensor signals were preprocessed using a Kalman filter for noise reduction and Z-score normalization for scale consistency. Feature extraction using Wavelet Transform (WT) was then performed to compute joint angles, limb orientations, stability indices, and angular displacement metrics from fixed-length temporal windows. The proposed framework combining rule-based constraints with a Dynamic Beetle Antennae Search–optimized Stacked Long Short-Term Memory (DBAS-SLSTM) model was employed to learn spatiotemporal motion patterns, detect deviations from reference postures, classify error severity, and predict corrective movements. The system also provides real-time visual and auditory feedback to guide users during training. Experimental validation in a MATLAB environment, using a 70%-30% training–testing split, achieved an accuracy of 98.12%, along with high precision, recall, and F1-score values, demonstrating statistically reliable performance. The results confirm that the proposed approach offers a scalable, accurate, and practical solution for posture monitoring in sports training, rehabilitation, and ergonomic applications.