Conditional Generative Adversarial Network-Based Regression Approach for Walking Distance Estimation Using Waist-Mounted Inertial Sensors

Abstract

This article introduces a novel regression approach based on deep learning for estimating the walking distance using inertial sensors attached to the pedestrian’s waist. Walking step length can be estimated by using supervised learning. However, supervised learning commonly requires a large amount of labeled training data to achieve better performance. To tackle this issue, we propose the walking step length estimation method based on a conditional generative adversarial network (CGAN) used as a regression model. The CGAN-based regression model consists of a generator model for a step length regression task and a discriminator model for a classification task. Step segmentation is performed to extract acceleration amplitude data into step segments. These data are applied as additional input for both the generator and the discriminator. The generator model aims to generate walking step length as a label, while the discriminator model aims to classify an input label as either real or generated. Then, the step length prediction model using the CGAN-based regression approach is applied to calculate the walking distance. Two experiments are performed to evaluate the performance of the proposed method, where test walking paths include an 80-m straight corridor and a rectangular football field of about 1282 m. The results show that the proposed method using small labeled datasets, with 120–300 samples achieve estimation accuracy with an average error of 0.77% for straight paths and 0.88% for rectangular paths.