串連電信與交通:推進圖神經網路於時空推理與多模態融合及預測

Abstract

交通預測對於現代城市規劃和智慧交通系統（ITS）至關重要，因為它在交通管理、資源分配和公共安全方面具有顯著的優勢。然而，傳統方法依賴部署實體的感測器和攝影機，成本高且覆蓋範圍不足，突顯了需要可擴展且高效的替代方案。為了解決這些挑戰，我們探索了電信數據作為交通預測的彈性且可擴展的替代方案潛力。 我們提出使用從各個地區收集的“基於電信的流量”來評估交通狀況並同時保護用戶隱私。然後，我們提出一個時空預測模型來基於電信流量來預測這些區域未來的變化，作為初步的交通指標。為了有效地提高預測準確性，我們提出了一個多模態數據融合框架，利用廣泛電信數據與稀疏基於影像數據的結合來改善預測效果。 進一步，為了解決電信流量與實際交通流量之間的差距，我們改進了融合框架並提出聯合損失函數，使預測結果對齊實際交通流量。此外，我們開發了新穎框架，透過跨模態方式將無方向性的電信數據轉為有方向性的移動流量，解決缺乏方向性的問題。這些方法確保了更準確的實際交通狀況、精確的交通動態，並為城市管理提供了寶貴的見解。 總體而言，本論文通過將電信數據與交通領域知識相結合，推進了交通預測的發展。我們的創新方法提供了多樣且可擴展的解決方案，超越了傳統基於傳感器的系統，並為城市規劃和智慧交通系統做出了重大貢獻。

Traffic prediction is crucial for modern urban planning and intelligent transportation systems (ITS) due to its significant benefits for traffic management, resource allocation, and public safety. Traditional methods, reliant on fixed-location sensors and cameras, are costly and provide insufficient coverage, highlighting the need for scalable and efficient alternatives. To address these challenges, we explore the potential of telecom data as a flexible and scalable source for traffic prediction. We propose the use of "telecom-based flow" collected from various areas to evaluate traffic conditions while protecting data privacy. We then present a spatio-temporal predictive model to predict future traffic conditions as an initial traffic indicator. To further improve prediction accuracy, we propose a multi-modal data fusion framework that combines extensive telecom data with sparse vision-based data. To address the gap between telecom-based and actual traffic flow, we develop and enrich the fusion framework with a joint loss function to more closely align predictions with actual traffic flow. Additionally, we develop a cross-modal framework to convert undirected telecom data into directed mobility flow to address the lack of directionality. These approaches ensure more accurate real-world traffic conditions, precise traffic dynamics, and provide valuable insights for urban management. Overall, this dissertation advances traffic prediction by integrating telecom data with traffic domain knowledge. Our innovative approaches offer diverse and scalable solutions, extending beyond traditional sensor-based systems, and provide significant contributions to urban planning and ITS.