Most existing studies using electrostatic sensors for gait analysis have been conducted in tightly controlled laboratory settings, which poorly reflect everyday living environments. In contrast, this work deliberately exploits a dataset acquired under realistic conditions, where low-cost electrostatic sensors are installed along a corridor in a home-like environment and record people walking naturally as part of their daily activities. The article focuses on estimating footstep impact instants. Two complementary strategies are investigated: an analytical, physics-based approach that detects local minima in electrostatic signals, and a statistical, machine-learning approach based on neural networks. An automated acquisition and labeling system, combining electrostatic sensors with a depth camera and skeleton tracking, provides large-scale reference data without explicit instructions to participants. This paper proposes to use possibility theory to handle the imprecise knowledge about the presence or absence of footsteps, particularly regarding the impact instants during the learning phase. Although its use is uncommon, this theory offers a natural and consistent framework for the adopted approach. The use of unsupervised learning via a statistical method allows us to achieve results similar to those obtained in other studies that used supervised learning via an analytical method. • Automatic dataset labeling using a depth camera and skeleton tracking. • Uncontrolled labeling with people walking naturally. • Two approaches for footstep impact instants: analytical (local minimum) vs. statistical (neural-network-based). • Possibility theory to handle uncertain footstep labels during learning. • Best statistic model comparable to state of the art analytic methods in controlled environment.
Lapique et al. (2026) studied this question.