1. Introduction

Location-based storytelling, a key strategy in smart tourism, helps visitors and residents deepen their understanding of local culture and history through urban exploration. By fostering a sense of attachment and pride, such experiences contribute to more comfortable and meaningful mobility within the city. A representative example is the proactive local guide, which delivers guides automatically when users arrive at key points of interest (POIs). These services rely on geofencing technology that triggers content delivery as users enter predefined spatial zones.

Manually designing effective and scalable geofences in urban settings is challenging. It requires consideration of varying GPS accuracy across locations and the complexity of pedestrian movement patterns, making manual implementation labor-intensive and often suboptimal (Garzon et al, 2017). Moreover, the appropriate timing and spatial conditions for triggering content depend on the nature of the storytelling itself (see Section 2). Thus, geofence design must address not only technical and behavioral constraints, but also the need to differentiate spatial triggers based on narrative intent. This study proposes a data-driven approach for identifying geofence zones by leveraging mobile sensor data collected from tourists.

2. Geofence Typology Based on Storytelling Needs

To align geofence design with different storytelling needs, we classify geofences into two types: enclosure-geofences and viewpoint-geofences (Figure 1).Enclosure-geofences associates situations where users physically enter museums. These geofences support immersive content tied to the space users are exploring. In contrast, viewpoint-geofences are placed at outdoor viewpoints from which users observe POIs, such as a building façade, from a distance. This distinction reflects different user behaviours—exploring from within versus observing from outside—and leads to two geofencing models: the ROI-based model for indoor cases and the viewpoint-based model for viewing cases. Clarifying this typology allows appropriate content to be delivered at the right place and moment, enhancing the relevance of location-based storytelling.

3. Classification Framework for Geofence Identification

This study investigates methods for extracting clusters corresponding to enclosure- and viewpoint-geofences from GPS trajectory data. Spatial clustering for detecting meaningful clusters from trajectory data have been proposed. We examined two widely adopted methods: HDBSCAN (Campello et al., 2015) and Stay Point Detection (Li et al., 2008). Figure 2 presents GPS trajectories recorded during approximately two hours of walking tours. HDBSCAN suffers from difficult parameter tuning (Yang et al., 2014), often resulting in overly large clusters including enclosure- and viewpoint-geofences, Stay Point Detection requires users to remain at fixed locations, making it unsuitable for detecting single clusters in large facilities.

This study proposes a classification framework utilizing mobile sensor data to identify geographic zones enabling to associate with each storytelling needs (Figure 3). First, GPS accuracy clustering extracts clusters from segments of the trajectory where the GPS horizontal accuracy falls below a certain threshold, indicating enclosure-geofence candidates. However, some of the extracted clusters did not accurately correspond to indoor stays. For example, low GPS accuracy may also occur outdoors, such as under tree cover or immediately after the GPS sensor is activated. In addition, areas like underground walkways, although technically indoor, often serve as mere transit routes rather than destinations of interest. To address these issues, we implemented a binary classification step for noise detection. This classifier aims to distinguish between meaningful indoor zones and spurious clusters caused by environmental noise or transient movement. Specifically, we examined combinations of features derived from smartphone sensor data, including the mean and variance of acceleration and horizontal accuracy values, as well as the number of GPS points contained within each cluster ().

4. Evaluation of Enclosure-Geofence Identification

To evaluate the proposed approach, we conducted an in-situ data collection study involving 12 participants who each carried an iPhone 11 while engaging in walking-based sightseeing activities for approximately 90 minutes on average. The resulting mobile sensor dataset was used to apply the GPS accuracy clusteringmethod. This process identified 35 clusters in total. Of these, 25 clusters were confirmed to correspond to actual user stays, while the remaining 10 were associated with either underground walkways or erroneous detections in outdoor environments. To effectively classify these clusters, we developed a decision tree classifier and examined various combinations of features through scatter plot analysis. The combination of mean acceleration and acceleration variance yielded the highest class separability, achieving an accuracy of 0.83 in leave-one-out cross-validation (LOOCV). These results suggest that users tend to slow down or pause intermittently around points of interest, leading to distinctive acceleration patterns that help distinguish indoor stays from transient movement.

5. Conclusions

This study introduces a systematic and computational framework for generating story-aware geofences from mobile sensor data. By classifying clusters based on narrative intent, and using data-driven techniques to detect and validate these regions, we provide a foundation for scalable, context-sensitive storytelling in urban environments. Future work includes the automatic delineation of geofence boundaries, improved alignment with POI datasets, and the development of a fully autonomous geofencing service deployable across diverse cities through crowdsourced mobile data collection.