Anticipating Pedestrian Movements: Future Activity Prediction for Floor-Change Scenarios

Pavel Pascacio, Thomas Bonis, and Valérie Renaudin

Peer Reviewed

Abstract:	As pedestrian navigation systems evolve into dynamic, context-aware assistance systems, predicting upcoming user activities, such as approaching elevators or escalators in Global Navigation Satellite System (GNSS)-denied environments, has become essential. This capability enhances user experience, safety, and decision-making and can offer benefits for applications from mobility aids to real-time context awareness in crowded urban spaces. Despite advancements in Pedestrian Activity Recognition (PAR), predicting future activities remains challenging due to low-quality smartphone sensor data and the difficulty of collecting and labeling transition-rich datasets used for training. Reliable labeling of subtle and rapid transitions—such as switching between walking, stair climbing, and escalator use—is essential for training effective models. However, manual labeling is resource-intensive and error-prone, while real-time user labeling disrupts natural activity patterns, degrading data quality. To address this, we propose a novel two-stage methodology for activity prediction that leverages both labeled and unlabeled datasets to build reliable models. In the first stage, an XGBoost-based PAR model is trained to recognize five activities, including walking, standing still, and floor-changing activities (stair climbing, elevator and escalator riding). This model is then used in the second stage to automatically label datasets with complex activity transitions. By shifting sensor data and labels over time (aligning sensor data at t1 with labels at t2), the second stage enables the development of a predictive model to anticipate future activities. Experimental results demostrated that the XGBoost-based PAR model achieves up to 95.11% accuracy with fixed time window segmentation of 3-seconds and 1-second step size. Furthermore, the Pedestrian Activity Prediction (PAP) model successfully predicted floor-changing activities, such as escalator and elevator transitions, with up to 72.61% accuracy for 1- second ahead predictions and 54.76% accuracy for 2-second ahead predictions, showcasing the potential of self-labeling in advancing PAP systems. These findings highlight the feasibility of our methodology and its potential to improve real-time navigation systems based on predicted floor changes. Keywords—Pedestrian activity recognition, pedestrian activity prediction, XGBoost, machine learning, smartphone sensors, floor-change detection
Published in:	2025 IEEE/ION Position, Location and Navigation Symposium (PLANS) April 28 - 1, 2025 Salt Lake Marriott Downtown at City Creek Salt Lake City, UT
Pages:	35 - 43
Cite this article:	Pascacio, Pavel, Bonis, Thomas, Renaudin, Valérie, "Anticipating Pedestrian Movements: Future Activity Prediction for Floor-Change Scenarios," 2025 IEEE/ION Position, Location and Navigation Symposium (PLANS), Salt Lake City, UT, April 2025, pp. 35-43.
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In