Enhanced ETA Predictions with T-GCN on Optimized Road Segments

Accurate Estimated Time of Arrival (ETA) predictions play a critical role in improving the operational efficiency, traffic management, and reliability of services in transit systems. This work presents a comprehensive framework for predicting the ETA within an Intelligent Transit Management System (ITMS). We propose a robust framework capable of forecasting traffic delays considering both spatial and temporal dependencies of the traffic. We also propose an optimized road segmentation framework that partitions the entire road network into segments of non-uniform length ensuring every segment has an ideal length with sufficient recorded position data density. This segmentation framework yields reliable statistics for each segment on the road, especially in scenarios with sparse spatially distributed positional records. Using these segments, we use a traffic delay model based on the Temporal Graph Convolutional Network (TGCN), however, augment it with an adjacency matrix that captures correlations between unconnected segments. Finally, we use the segment-wise traffic delays in an end-to-end framework to estimate the arrival time of buses at a given point in the network.