Walking Speed and Uncertainty Estimation Using Mixture Density Networks for Dynamic Ambulatory Environments

Walking speed, often considered a representative indicator of activity levels, becomes notably reduced as muscle strength and cardiovascular function decline with aging. Wearable walking rehabilitation devices aim to alleviate the effort during walking or enhance the necessary muscles. Measuring the wearer's walking speed provides an objective assessment of rehabilitation progress. While various methods, such as GPS, model-based estimation, and deep neural network regression can estimate walking speed, they encounter challenges in diverse environments. This article introduces the CNN-based Mixture Density Network (CMDN) structure, which enhances accuracy and provides uncertainty information about estimated walking speed, indirectly reflecting the current walking environment. Validated with experiments involving 20 elderly individuals, CMDN demonstrated performance across flat and stair descent situations, showcasing its potential as a foundation for widespread use in diverse scenarios.