Smart Tourist Safety System: Integrating Machine Learning, Spatial Data, and Human-Centred Emergency Workflows

Abstract

Tourists visiting unfamiliar locations may face difficulties identifying nearby emergency services, communicating their location, and obtaining timely information about the status of an emergency request. The Smart Tourist Safety System (STSS) is designed to address these challenges through a software platform that combines location-aware services, emergency-service directories, machine-learning models, and responder communication workflows. The system includes district-level crime-risk estimation, audio-based distress classification, and natural-language intent recognition. It also uses geospatial calculations to support distance estimation and location-based processing. According to the available project audit, the calibrated crime-risk model achieved 82.54% accuracy, while the TF-IDF and LinearSVC intent classifier achieved 84.92% test accuracy with a macro F1-score of 0.8437. The audio distress classifier reported 37.72% precision and 54.43% recall at a distress threshold of 0.80. At this threshold, the reported number of false positives decreased from 791 to 71. These results illustrate both the potential and the limitations of applying machine learning to tourist safety. In particular, prediction errors, incomplete geographic data, and differences between offline model artifacts and active application components must be considered. STSS is therefore presented as a developing decision-support platform rather than a proven replacement for emergency services. Further reproducible evaluation and real-world testing are required before its effectiveness can be established.

Introduction

The Smart Tourist Safety System (STSS) is a web-based application designed to improve tourist safety by integrating emergency-service information, location-based services, natural-language processing, and machine-learning techniques into a single platform. Tourists travelling in unfamiliar locations may struggle to find nearby police stations or hospitals, communicate emergencies, and obtain reliable assistance because of language barriers, limited connectivity, and fragmented digital services.

The proposed system provides a centralized platform connecting tourists with police and hospital directories, incident-reporting functions, and responder-oriented interfaces. It uses a React and TypeScript frontend, a Node.js and Express backend, MongoDB for data storage, and Socket.IO for real-time communication where configured. External services, such as OSRM routing, support location-related functionality.

The system incorporates three main machine-learning components:

1. District-level crime-risk estimation: A Random Forest model with 150 trees and a maximum depth of eight estimates crime risk using district-level crime data. Platt scaling is applied to calibrate the predicted probabilities.

2. Audio-based distress classification: A two-dimensional Convolutional Neural Network (CNN) analyses 64-band log-Mel spectrograms to classify audio into three categories: distress, normal, and background.

3. Tourist-query intent recognition: A TF-IDF and LinearSVC classifier identifies the intent behind tourist queries. A pretrained MiniLM sentence-embedding model is also used for semantic matching.

The reported evaluation of the tourist-query classifier achieved 84.92% accuracy and a macro F1-score of 0.8437 on a dataset containing 839 annotated queries across 24 intent categories. The other machine-learning components require appropriate evaluation evidence before their real-world effectiveness can be established.

The system also employs deterministic geospatial algorithms, including the Haversine formula for calculating geographic distances and point-in-polygon checks for determining whether locations fall within specified boundaries. These functions support the identification of nearby emergency facilities and location-aware services.

A major challenge identified in the project is data quality. Many police stations, hospitals, and tourist attractions lack verified facility-level coordinates. Approximate administrative centroids may produce misleading distance calculations, so geographic records need verification, clear source information, and confidence indicators.

Conclusion

This paper presented the Smart Tourist Safety System as an integrated platform for tourist safety information, location-aware processing, machine-learning-assisted classification, and responder communication. The system combines trained models with deterministic geospatial calculations and application-level workflows. The available project audit reports promising results for the intent classifier and calibrated crime-risk model, while the audio classifier demonstrates the difficulty of balancing false alerts against missed distress events. The audit also identifies limitations in coordinate coverage and differences between certain offline model artifacts and the active runtime. These findings indicate that STSS is a developing research platform with potential for further investigation, rather than a system whose real-world safety effectiveness has already been established. Reliable deployment will require reproducible experiments, verified directory data, transparent communication of uncertainty, careful privacy controls, human confirmation for consequential alerts, and testing under realistic conditions.

References

[1] National Crime Records Bureau, Crime in India. Government of India. Verify the exact editions, publication years, tables, and URLs used. [2] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” in Proceedings of NAACL-HLT, 2019, pp. 4171–4186. [3] A. Vaswani et al., “Attention Is All You Need,” in Advances in Neural Information Processing Systems, vol. 30, 2017. [4] F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation Forest,” in Proceedings of the IEEE International Conference on Data Mining, 2008, pp. 413–422. [5] J. Platt, “Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods,” in Advances in Large Margin Classifiers, 1999, pp. 61–74. [6] Project OpenStreetMap, OSRM: Open Source Routing Machine. Verify the documentation URL and access date. [7] Sentence Transformers, all-MiniLM-L6-v2 Model Card. Verify the exact model repository and version used by the project. [8] Add the primary references and licensing information for the exact ASVP_UPDATE and NIGENS dataset versions used in the audio experiments.

Copyright

Copyright © 2026 S. Jadesha Gowda, T. Dinesh Reddy, Sharath Bhoopal B J, Chanabasava G, Shashikanth Reddy. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.