Stumble Detection with LSTM Autoencoder for Walking Stability Evaluation in Pedestrian Simulator

Abstract

To improve the safety of vulnerable road users (VRUs) and enhance communication among road users, it is important to analyze complex and potentially hazardous traffic scenarios. Simulating these scenarios in a controlled and safe environment is crucial for testing. However, pedestrian simulators can introduce instability and unnatural walking patterns, making it necessary to assess whether participants have undergone sufficient training to walk stably within the simulation environment. This study presents an LSTM autoencoder model to detect stumble behavior while pedestrians interact with a pedestrian simulator. Walking data, including body position and rotation, are collected from participants during simulated walking tasks and used as input to the model. The proposed approach effectively learns from stable walking behavior, identifies stumble behavior, and thus provides quantitative feedback on gait stability. This feedback can be used to evaluate participant readiness and training adequacy for realistic walking within virtual simulation environments.