AI-Driven Accident Detection and Emergency Response for Low-Resource Settings

Everlyne Nalule

Keywords: Deep Learning, Python, Machine Learning

Abstract:

AI-Driven Accident Detection and Emergency Response for Low-Resource Settings: A Case Study from Uganda
Abstract

Uganda witnessed 5,144 road traffic deaths in 2024—a 7% increase over 2023—with motorcycle taxis (boda bodas) accounting for nearly 47% of fatalities and in recent weeks peaking at 53% among crash victims, (Uganda Police Annual Crime Report, 2024). Over 420,000 traffic cases were reported in 2024 alone yet still the country’s emergency response remains ineffective with scarce ambulances that are often fee-based, police hotlines grossly underused, and many health centers lacking the capacity to handle trauma cases.
Existing accident detection and emergency alert systems in high-income settings assume strong infrastructure—reliable call centers, ambulance coverage, and trauma-equipped hospitals—which do not reflect realities in Uganda. There is an urgent need for AI-powered systems that both detect accidents and adapt to socio-technical constraints in low-resource contexts.

We address this gap by proposing an AI-driven accident detection and alert system designed for low-resource settings. Our novelty lies in coupling computer vision–based accident detection from roadside cameras with a hospital-capacity–aware routing and multi-channel alerting framework tailored to Uganda’s emergency ecosystem. The system employs convolutional neural networks with temporal anomaly detection to recognize motorcycle crashes, integrates OpenStreetMap and Ministry of Health data to identify the nearest trauma-ready hospital, and dispatches alerts through SMS/WhatsApp APIs.

The methodology begins with continuous video streams from roadside surveillance cameras, which are processed using a CNN backbone (a YOLOv5 variant) augmented with a temporal anomaly detection module based on GRUs. This architecture is designed to capture abrupt motion patterns such as falls, skids, or collisions, with a particular emphasis on motorcycle accidents that dominate Uganda’s traffic fatalities. To reduce false positives, we apply multi-frame confidence verification, ensuring that only consistent patterns across consecutive frames are classified as accidents.

Our contributions are threefold: (i) a spatiotemporal accident detection pipeline optimized for motorcycle crash scenarios; (ii) a geospatial hospital mapping algorithm prioritizing trauma capacity rather than geographic proximity; and (iii) coordinated alerting framework that directs notifications to the most relevant emergency provider, including government (KCCA, MoH), humanitarian and private responders (C-Care, City Medicals) ensuring a more efficient response.

Preliminary evaluations on synthetic and pilot Ugandan traffic datasets demonstrate over 83% accuracy in accident detection and alert dispatch times under one minute, significantly outperforming manual reporting. By enabling faster accident detection and hospital readiness, the proposed system has the potential to reduce fatalities in Uganda and offers a scalable framework for improving emergency response across similar settings in Sub-Saharan Africa and beyond.

References: Uganda Police Force (2024); Zheng et al. (2021); Nguyen et al. (2019); UC Berkeley EmergEye (2024).

Submission Number: 162