Abstract

This study aims to establish an automated analysis framework that quantitatively derives and visualizes latent movement path patterns within a public library, using real-time movement data collected through IoT-based camera sensors. To this end, continuous movement data captured by the sensors were structured using an N-gram approach and analyzed using Latent Dirichlet Allocation(LDA) topic modeling. Two term weighting methods— TF-IDF and Word2Vec—were each combined with bigram and trigram models, resulting in four analytical models. Topic distributions from each model were compared, and the structural characteristics of movement flows were visualized. To address the limitation of conventional LDA in capturing directionality and sequential information, the study also proposed an analysis method based on the Topical N-gram technique. The analysis results from each model were integrated using an ensemble approach based on cosine similarity and Jensen-Shannon Divergence (JSD). The experimental results revealed meaningful and repetitive movement patterns that are difficult to detect using simple statistical methods. In particular, key user routes centered around the ‘entrance’ and the ‘ information desk’—both serving as guidance and reference service hubs—were clearly identified. This study is significant in that it presents an integrated analysis framework capable of quantitatively interpreting user behavior based on real-time movement data, offering practical applications for the operation and planning of public services.