Urban Exploration of Içərişəhər: Data Integration and Pathfinding System

Abstract

Navigating through historical urban districts as İçərişəhər or the Old City of Baku ushers in new challenges never met in the typical navigation systems. The usual tools are inclined upon path analysis on the basis that the path be least in distance from one spot to another, leaving aside the myriad amounts of cultural references, scenic beacons, and the specific needs of exploring heritage places on behalf of visitors. This thesis went on to fill that gap by proposing data backend support, data processing, and deployment of multi-criteria urban exploration navigation system for İçərişəhər. The aim was to build a system able to generate distinct paths such as the standard shortest path by distance, "cultural" path for optimizing historical and cultural context, and "geocaching" path for including points of interest from this geocaching game. A comprehensive approach was used to source data from OpenStreetMap (for the road network) and Google Places API for Points of Interest-POIs with respect to a clearly defined area of study. A robust data processing pipeline was designed with Python's geospatial libraries and was used to clean network geometry, calculate segment lengths accurately, find the intersections in question efficiently through R-tree spatial indexing, and generate network graphs accurately. Eventually, the gemini-2.5-pro Large Language Model was used for uniquely filtering over 11,000 POIs to 5,689 really relevant POIs according to context and exploration objectives. Now the "interestingness" of different road segments was reckoned with a new "relevance" metric that came into play after refining of filters and network-based smoothening. A mix of PostgreSQL/PostGIS and Neo4j hybrid database models was setup with the former used for the exhaustive storage of spatial objects and attributes and the latter for network topology modeling and prompt pathfinding using its Graph Data Science (GDS) library. Proceeding from GDS, the A* search algorithm was employed to do the shortest path search and find a valid geographical path using segment length and relevance score, respectively. The geocaching path is constructed by combining two shortest paths directed through the closest to original shortest path geocache. A FastAPI backend controls all database layer interactions and expose the routing logic, while Docker along with Docker Compose supported containerization of the system, hence achieving reproducibility and ease of continuous deployment. Experimental outcomes show the generation of various and more than feasible paths for each of the models, which is a clear indication of the success of the relevance factor and hybrid architecture. Testing for observation proved it had reasonably fast response times with an average of 200ms for the API which demonstrates adequate performance. This work has contributed to a carefully crafted data processing workflow using an unexpected LLM POI filter application, a custom metric for relevance, and a functional plug-and-play containerized backend system that certainly paves the way for enhanced exploration and navigation tools for cultural heritage in mixed and rich historical urban environments.