Autonomous Intelligent Transportation Systems: A Data-Driven Approach

Abstract

Autonomous intelligent transportation systems (AITS) represent a convergence of sensor fusion, deep learning perception, path planning, and V2X
communication technologies that promise to transform urban mobility through improved safety, efficiency, and accessibility. However, the data-driven
design of robust AITS capable of operating in diverse, unstructured real-world traffic environments — including adverse weather, mixed autonomy
traffic, and complex urban intersections — remains an open engineering and scientific challenge. This study presents DriveNet, a unified data-driven
AITS framework integrating a multi-modal perception stack (LiDAR-camera fusion with BEV transformer), a graph neural network-based trajectory
prediction module, and a model predictive reinforcement learning (MPRL) planner for safe, comfortable autonomous navigation. DriveNet was
evaluated on the nuScenes, Waymo Open Dataset, and CARLA Town05 simulation benchmarks, and validated in a real-vehicle deployment on the
TUM Campus Garching test route (4.7 km, mixed urban-suburban). DriveNet achieved nuScenes detection mAP of 68.3%, 3D tracking AMOTA of
0.621, trajectory prediction minADE of 0.84 m (5-second horizon), and a collision-free autonomous driving success rate of 94.7% on CARLA
Town05. In real-vehicle deployment, 847 km of autonomous operation were completed without safety driver intervention, achieving a 98.3% comfort
rating from 42 passenger evaluators. These results demonstrate that unified data-driven AITS frameworks integrating perception, prediction, and
learning-based planning achieve robust, safe performance across both simulation and real-world operational environments.