Evaluación de algoritmos basados en LSTM para la predicción de parámetros de marcha en personas con discapacidad visual
Loading...
Date
2026-04-03
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Tecnológica Centroamericana UNITEC
Abstract
La prevalencia creciente de personas con discapacidad visual en Honduras subraya la necesidad de tecnologías de asistencia para mejorar la movilidad de personas con discapacidad visual. Esta investigación se centra en el potencial de los algoritmos de aprendizaje automático, específicamente modelos híbridos LSTM+CNN y LSTM+FCN, para predecir parámetros de marcha esenciales, como la velocidad, la longitud de zancada, distancia y número de pasos. Usando una base de datos proporcionada por una investigación previa la cual contiene datos obtenidos por sensores inerciales que abarca a nueve participantes en diversas condiciones de caminata, estos modelos identifican patrones temporales y características espaciales. Se compararon las predicciones entre los modelos híbridos y el modelo biomecánico que utilizó configuraciones de TWS y THS. En relación con el tamaño de zancada, el modelo LSTM+FCN obtuvo un error de 0.015 ± 0.014 metros, superando al modelo biomecánico. En cuanto a la velocidad, los errores absolutos medios fueron de 0.018 ± 0.013 m/s y 0.044 ± 0.017 m/s para los dos modelos, respectivamente, siendo significativamente más bajos que los del modelo biomecánico. En la predicción del contador de pasos con el modelo Seq2One mostró una mayor variabilidad con un valor de 16.06 ± 32.11 en las métricas de los experimentos. Sin embargo, en comparación con el modelo biomecánico, presentó un margen de error más amplio, lo que destaca la necesidad de optimización futura. En cuanto al parámetro de distancia, el error absoluto medio global fue de 13.13 ± 7.93 m para el modelo LSTM+4FCN y de 21.37 ± 9.25 m para el modelo LSTM+CNN, El modelo Seq2One obtuvo un error absoluto medio global de 2.96 ± 3.3. En comparación con el TWS, este logra superarlo en términos de predicción, sin embargo, obtuvo mejores resultados en comparación con el modelo LSTM+FCN. Los algoritmos basados en LSTM pueden proporcionar predicciones más precisas de los parámetros de la marcha en relación con el método tradicional, por ende, se deberían de realizar pruebas adicionales para los parámetros de distancia y número de pasos.
The increasing prevalence of people with visual impairment in Honduras highlights the need for assistive technologies to enhance the mobility of individuals with visual disabilities. This research focuses on the potential of deep learning algorithms, specifically hybrid models LSTM+CNN and LSTM+FCN, to predict essential gait parameters such as velocity, step length, distance, and step count. Using a database provided by a previous study, which includes data obtained from inertial sensors and covers nine participants in various walking conditions, these models identify temporal patterns and spatial characteristics. Predictions were compared between the hybrid models and the biomechanical model that used TWS and THS configurations. Regarding step length, the LSTM+FCN model achieved an error of 0.015 ± 0.014 meters, surpassing the biomechanical model. As for speed, the mean absolute errors were 0.018 ± 0.013 m/s and 0.044 ± 0.017 m/s for the two models, respectively, which were significantly lower than those of the biomechanical model. In predicting step count with the Seq2One model, a higher variability was observed with a value of 16.06 ± 32.11 in the experiment metrics. However, compared to the biomechanical model, it exhibited a wider margin of error, emphasizing the need for future optimization. Concerning the distance parameter, the overall mean absolute error was 13.13 ± 7.93 m for the LSTM+4FCN model and 21.37 ± 9.25 m for the LSTM+CNN model, while the Seq2One model obtained an overall mean absolute error of 2.96 ± 3.3. Compared to TWS, it managed to outperform it in terms of prediction; although it achieved better results compared to the LSTM+FCN model. LSTM-based algorithms can provide more precise predictions of gait parameters compared to the traditional method; however, additional tests should be conducted for parameters like distance and step count.
The increasing prevalence of people with visual impairment in Honduras highlights the need for assistive technologies to enhance the mobility of individuals with visual disabilities. This research focuses on the potential of deep learning algorithms, specifically hybrid models LSTM+CNN and LSTM+FCN, to predict essential gait parameters such as velocity, step length, distance, and step count. Using a database provided by a previous study, which includes data obtained from inertial sensors and covers nine participants in various walking conditions, these models identify temporal patterns and spatial characteristics. Predictions were compared between the hybrid models and the biomechanical model that used TWS and THS configurations. Regarding step length, the LSTM+FCN model achieved an error of 0.015 ± 0.014 meters, surpassing the biomechanical model. As for speed, the mean absolute errors were 0.018 ± 0.013 m/s and 0.044 ± 0.017 m/s for the two models, respectively, which were significantly lower than those of the biomechanical model. In predicting step count with the Seq2One model, a higher variability was observed with a value of 16.06 ± 32.11 in the experiment metrics. However, compared to the biomechanical model, it exhibited a wider margin of error, emphasizing the need for future optimization. Concerning the distance parameter, the overall mean absolute error was 13.13 ± 7.93 m for the LSTM+4FCN model and 21.37 ± 9.25 m for the LSTM+CNN model, while the Seq2One model obtained an overall mean absolute error of 2.96 ± 3.3. Compared to TWS, it managed to outperform it in terms of prediction; although it achieved better results compared to the LSTM+FCN model. LSTM-based algorithms can provide more precise predictions of gait parameters compared to the traditional method; however, additional tests should be conducted for parameters like distance and step count.
Keywords
Análisis de la marcha, Discapacidad visual, O&M, Redes neuronales