Terrain Classification From Human Gait Patterns Using Temporal Feature–Enhanced Multilayer Perceptron Model: A Cross-Participant Study

Abstract

Terrain classification constitutes a fundamental problem with wide-ranging applications in robotics, prosthetics, assistive technologies, and human gait analysis. The inherent variability of real-world environments, combined with the subtle biomechanical differences across terrain types, makes accurate classification particularly challenging. To address this problem, this study presents a terrain classification framework employing a Multilayer Perceptron (MLP) trained on an existing inertial measurement unit (IMU) data gathered from 30 participants across nine terrain conditions and one calibration condition, yielding 10 classification classes in total Two training strategies are investigated and compared: the first uses a fixed participant split (24 train / 6 test) without feature engineering, evaluated over 5 independent experiments; the second incorporates temporal statistical feature extraction from sliding signal windows prior to classification. The first method achieves a mean test accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$72.23 \% \pm 4.39 \%$</tex>, while the second reaches 87.31 % with a macroaveraged <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1}$</tex>-score of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 8 7}$</tex>, representing an improvement of +15 percentage points. These results demonstrate that handcrafted statistical and temporal descriptors provide a significantly more discriminative input representation for terrain classification than raw sequential data, and confirm the effectiveness of the proposed MLP architecture for cross-participant terrain recognition.