Myeonggi Hong, Eunbi Cho, Journal of Internet Computing and Services, Vol. 27, No. 1, pp. 157-165, Feb. 2026
10.7472/jksii.2026.27.1.157, Full Text:  HTML
Keywords: Environmental criminology, CPTED, Sementic Segementation, clustering, Street Environment

Abstract

This study examines street-level sexual offense locations from an environmental criminology perspective by combining deep-learning–based semantic segmentation with K-means clustering to quantify and typologize micro-environmental features. Using 122 real incidents, we collected forward- and backward-facing Kakao Street View images for each site and applied PSPNet to extract pixel proportions for 19 classes such as road, sidewalk, building, vegetation, and vehicles. To secure spatial discriminability, candidate cluster solutions were evaluated with the Elbow criterion and the Silhouette score before fitting K-means. The analysis identified three spatial types of offense locations: open spaces characterized by higher shares of vegetation and sky and lower building proportions; market/commercial areas with higher proportions of people, buses, and trucks and a mid-level building share; and dense residential areas with higher building and wall proportions indicating stronger enclosure. Across clusters, incidents were concentrated during late-night hours, 00:00–05:00. The study complements city-scale, macro-indicator approaches by deriving a typology of crime-prone environments from image-based quantification of micro-level streetscape features.

Statistics
Cite this article