Open Collections

Abstract

Occupational slips, trips, and falls on the same level remain a major source of workplace injury. Wearable inertial measurement unit (IMU) sensors have been proposed for near-fall and fall detection, but many systems are sensitive to sensor placement and may not generalize well beyond controlled laboratory settings. This thesis evaluates a placement-adaptable near-fall and fall detection model in real-world occupational environments and extends it using decision-level sensor fusion. A total of 510 hours of data were collected from 30 adults (20 female, 10 male; age 26.8 ± 3.9 years) wearing the sensors on their lower back and both thighs. A total of 4 near-falls and 1 fall were reported. Due to extreme class imbalance, evaluation focused on activities of daily living (ADLs) and false positive fall and near-fall detections, where ADLs were incorrectly classified as events. Individual sensors achieved 97.21% ADL specificity but produced 36.07 near-fall and 0.98 fall false alarms per hour, with thigh sensors generating the highest false-positive rates. Decision-level fusion reduced false positives across sensors. Using three sensors, unanimous voting achieved 99.90% ADL specificity with 0.50 near-fall and 0 fall false alarms per hour, but required full agreement across sensors, reducing decision coverage. Majority voting achieved 98.47% ADL specificity with 7.60 near-fall and 0.05 fall false alarms per hour, but required all three sensors to be available to determine a majority, reducing robustness to sensor failure and user preferences. In contrast, Bayesian fusion maintained full decision coverage and reduced uncertainty from individual sensor predictions by 85.6%, enabling a measurable reduction in classification ambiguity while achieving 98.89% ADL specificity with 5.55 near-fall and 0.05 fall false alarms per hour. These findings demonstrate that decision-level Bayesian fusion improves reliability by reducing uncertainty and false positives in real-world occupational data without retraining the base models. While individual-level alerting remains limited by false positives, the framework enables population-level monitoring and establishes a scalable foundation for future occupational near-fall and fall risk assessment with larger datasets.