This paper proposes a stereo-based dynamic visual simultaneous localization and mapping (dynamic-vSLAM) approach tailored for autonomous driving. Conventional vSLAM systems are fundamentally constrained by the static-world assumption, resulting in performance degradation under dynamic conditions, such as environments with moving vehicles, pedestrians, or other objects. To address this limitation, we develop a front-end motion estimation scheme that explicitly models multiple independent motions arising from these moving objects in image sequences. ORB (Oriented FAST and Rotated BRIEF) key points are employed as observations, where their spatial locations are treated as time-varying variables. We propose a motion-clustering strategy, followed by probabilistic refinement to ensure motion consistency. In the back end, a unified bundle adjustment (BA) jointly optimizes camera poses, local map structure, and dynamic key point trajectories, thereby preserving the geometric relationships between dynamic observations and the camera motion. The proposed approach is validated using both publicly available benchmarks and custom datasets representative of autonomous driving scenarios. Quantitative evaluations demonstrate significant improvements in localization accuracy and robustness compared to conventional vSLAM systems in dynamic environments.
Generative artificial intelligence (GenAI) is rapidly advancing from text and image synthesis into domains with direct societal impact, including transportation. To examine its implications, the AI in transportation committee of ASCE Transportation & Development Institute convened the workshop Pioneering Transportation Innovation with Generative AI at the University of Maryland in March 2025, bringing together more than 170 participants from academia, industry, and government. Discussions spanned traffic operations, safety, institutional governance, and systemic adoption, highlighting both the promise of new capabilities and the fragility of current validation and policy frameworks. This forum paper presents a structured synthesis of the expert discussions, recurring themes, and cross-cutting insights that emerged during the workshop, highlighting research and policy priorities needed to guide the responsible advancement of generative AI in transportation practice.
Scaled robotic experiments offer a cost-effective and observable means to study collective traffic behavior, yet most existing testbeds lack dimensional consistency with full-scale systems. This paper establishes a dimensionally consistent framework for reproducing intersection-level operations in scaled experiments. Starting from the Lighthill-Whitham-Richards (LWR) conservation model, a complete set of dimensionless Π-groups is derived to represent macroscopic flow, wave propagation, and signal-timing relationships. These macroscopic quantities are then linked to microscopic control parameters through normalized car-following dynamics, enabling physical realization with small robotic vehicles. The framework is implemented in the Miniature Connected, Cooperative and Automated Mobility (MiniCCAM) laboratory, an indoor intersection testbed equipped with a fleet of scaled robots. Experiments demonstrate that the proposed approach preserves the relative wave propagation, inflow scheduling, and discharge behavior predicted by the Π-group analysis, while quantifying practical deviations caused by sensing limits and actuation latency. The results establish both the theoretical foundation and the experimental viability of dimensionally consistent intersection studies using scaled robotic platforms.
Evaluating the safety of autonomous vehicles (AVs) requires testing across diverse conditions to characterize their Operational Design Domain (ODD) and its boundaries. Scenario-based testing has emerged as a key approach, yet interpreting its results remains challenging due to the complexity and scale of simulation data. We present a visual analytics framework for behavior-centric analysis of scenario-based AV testing, linking AV behaviors with scenario parameters (e.g., oncoming vehicle speed and start delay) to support interpretable safety assessment. By integrating dimensionality reduction, clustering, and coordinated visualizations, the framework reveals diverse interaction patterns and identifies safety-critical regions, such as collision boundaries and regions of high risk variation. We demonstrate the framework through three case studies—intersection navigation, overtaking, and multi-lane cut-ins—which illustrate its ability to uncover nuanced behavioral differences, analyze boundary conditions, and compare AV planning systems. Expert feedback indicates that the framework provides interpretable and actionable insights to support safety-oriented behavioral analysis in simulation-based AV testing workflows.
Topographic flow estimates ego-motion from temporally adjacent top-down range-intensity observations and can provide odometry for unmanned surface vessels operating in limited or denied global positioning environments. This is done by estimating the planar-motion of a continuous topographic data stream, such as that provided by a maritime radar. A radio detection and ranging data stream was captured whilst traversing through the coastal waters of Sweden. The stream was converted into cylindrical coordinate images, and an optic flow estimator was used to track pixel intensity across temporally adjacent images. Analysis of this estimator was used to find angular velocity, which was removed, and the subsequent residual flow vectors were determined to be lateral and longitudinal velocities. Validation was performed by testing the method against inertial, velocity, and position reference sensors. Compared to the reference sensors, our results from topographic flow showed a standard deviation of less than 0.008 rad s−1 for angular velocity. The system also showed a standard deviation of less than 0.25ms−1 when solving for vessel velocity in longitudinal and lateral vessel movement. When used as an open-loop dead-reckoning component for a navigation system, the method was able to maintain the dominant yaw motion of the vessel, although fully open-loop topographic flow dead-reckoning remains sensitive to accumulated translational and heading errors. The results indicate that topographic flow is a viable onboard odometry component for GNSS-free navigation in confined waters with sufficient static structures.
