Document Type

Conference Paper

Publication Date

9-2-2026

Abstract

This project showcases the development and implementation of an artificial intelligence framework trained from a large dataset of street-level images to categorize new images with similar attributes.  The framework is implemented into an application to be used as a mobile navigation software.  The application is constructed using three main languages: Dart, Python, and Java, and uses several SDKs including Google Maps, Firebase, geolocators, shared preferences, and other Flutter resources.  The framework is constructed using a main artificial intelligence model, TensorFlow, trained from hundreds of Google Maps street view images and personally collected images.  The TensorFlow model is provided these images, pre-categorized into an “attractiveness” rating (1-5), along with a prompt telling the model what features to look for when evaluating the images.  The trained model categorizes street-view images along a route and calculates the path’s average scenery attractiveness rating.  That score determines whether the user will be prompted to change from their current route to an alternate, more scenic route.  The trained model is then downloaded onto the user’s device within the application, and TensorFlow Lite allows the model to be altered based on preferences for what a specific user considers “scenic.”  The user preference modeling is accomplished firstly through an initial questionnaire displaying a series of street-level images and secondly through users rating completed routes.  Those images are integrated back into the training algorithm, making the model an LLM Twin specified per-user.  This sort of trained model could be implemented into applications to categorize images based on specific attributes according to a user’s personalized preferences.

Keywords

ai, artificial intelligence, navigation, llm, llm twin, framework, app, mobile app

Disciplines

Artificial Intelligence and Robotics | Software Engineering

Download

Share

COinS