Automatic Detection, Extraction, and Analysis of Landing During a Training Session, Using a Wearable Sensor System

Abstract

We demonstrate an end-to-end wearable inertial sensor-based system for use in a natural sports training environment, which can automatically extract different movement actions and their individual movement cycles to generate and statistically analyse representative joint angle and impact acceleration data. The Discrete Wavelet Transform in conjunction with a Random Forest classifier was able to successfully distinguish between the six training activities (98% accurate). Accurate sensor orientation in 3D space were estimated using a computationally efficient gradient decent algorithm and were utilized to calculate joint angles, which were temporally aligned using curve registration to facilitate inter-participant comparisons. An Analysis of Characterizing Phases was applied to the whole joint angle curve and showed a statistically significant difference in knee joint flexion-extension landing strategy between a participant with low back pain and nine uninjured participants. The presented end-to-end system has the potential to be used in the automatic extraction and analysis of movement technique and loading in various unconstrained environments for musculoskeletal injury management.