An AI driven approach to manage multiple raw input sources in the context of orchestrated services

Abstract

This Thesis proposes a scalable, AI-driven framework designed to optimise the orchestration of diverse input sources within Intelligent Transportation Systems (ITS), balancing performance, reliability, and sustainability objectives. Central to this study is the development of the Multi-Resolution Traffic Monitoring Dataset (MRTMD), a pioneering dataset that supports systematic multi-resolution analysis for effective traffic monitoring. The primary aims of this research are to optimise multi-resolution data handling, reduce the environmental impact of high-resolution data processing, and enhance system adaptability through a queryable interface. Experiments conducted on the MRTMD dataset reveal that lower resolutions, such as 720p, achieve high detection accuracy, particularly with the YOLOv7 model, while significantly reducing storage and bandwidth demands. Findings demonstrate that the variability in mAP and recall between high and low resolution was 3%, concluding that prioritising lower-resolution data for routine monitoring tasks can cut energy consumption and carbon emissions without compromising system reliability. Additionally, the framework’s adaptable design allows operators to flexibly allocate resources between edge and cloud processing, balancing real-time response with sustainability. A key contribution of this research is the development of a queryable traffic monitoring interface, allowing users to extract actionable insights on vehicle counts, flow, and congestion trends from processed data. This interface enhances the interpretability of traffic data, enabling operators to make informed decisions rapidly. The study establishes a foundation for future-ready, sustainable ITS deployments, demonstrating that an AI-driven, multi-resolution approach can yield significant environmental and operational benefits in urban mobility contexts.