Urban Currents

Crossing the Road in Front of Autonomous Vehicles: An Investigation of Pedestrian Behavior Using Survival Analysis

This study analyzed pedestrian waiting times before crossing in front of autonomous vehicles (AVs) and human-driven vehicles (HDVs) using a random intercepts accelerated failure time survival model. The analysis utilized a large-scale, real-world AV dataset spanning Singapore and three US cities (Boston, Las Vegas, and Pittsburgh) to evaluate how vehicle type, traffic volumes, and regional differences affect crossing decisions.

Why it matters — The findings demonstrate that pedestrians wait longer to cross in front of human-driven vehicles than AVs, indicating a higher level of comfort with autonomous technology. It also establishes that city-level cultural and regulatory contexts significantly alter pedestrian caution, with Singaporean pedestrians waiting longer than their US counterparts.

Caveat: The study relies on observational interactions from a specific AV fleet's dataset, which may not represent pedestrian behavior around all types of autonomous driving systems or deployment stages.

Journal of Transportation Engineering Part A Systems · doi

ST-KFNet-Based Framework for Online Metro Passenger OD Demand Forecasting under Uncertainty

The researchers developed ST-KFNet, a forecasting framework that combines ARIMA, a Kalman filter, and a CNN-based variational autoencoder to predict online metro passenger origin-destination demand. Tested on a metro dataset from Hangzhou, China, the model integrates statistical and deep learning methods to capture spatial and temporal dynamics while handling travel uncertainties.

Why it matters — The framework enables real-time, online metro demand forecasting that remains robust during unexpected disruptions, such as unplanned station closures, outperforming traditional GCN, ARIMA, and standard deep learning benchmarks.

Caveat: The model's performance and generalizability are demonstrated using a dataset from only a single metro network in Hangzhou.

Journal of Transportation Engineering Part A Systems · doi

Linear ruptures: two faces of the High Line (New York) and the urban void of the Line (Saudi Arabia)

This paper presents a comparative theoretical analysis of two prominent 21st-century linear urban developments: the High Line in New York City and the Line in Saudi Arabia. It examines how the High Line repurposes existing infrastructure to create ruptures in the urban fabric while striving for social connectedness, whereas the Line ruptures the desert landscape through a massive, exclusionary infrastructural investment. The analysis applies theories of urban disorder, global urbanism, and the 'just city' to evaluate both projects.

Why it matters — It establishes a conceptual framework for understanding how linear urban forms can simultaneously act as forces of spatial rupture and as potential sites for social mending and public resistance.

Caveat: The study is entirely theoretical and conceptual, relying on qualitative literature and urban theory rather than empirical spatial data or field measurements.

Journal of Urbanism International Research on Placemaking and Urban Sustainability · doi

The vehicle routing problem with cross-docking for perishable products under freshness-life and travel time uncertainties

This study introduces a robust optimization model and an adaptive large neighborhood search (ALNS) algorithm to solve the vehicle routing problem with cross-docking for perishable products under travel time and freshness-life uncertainties. Tested on benchmark instances with up to 200 requests, the ALNS algorithm maintains stable performance across 10 runs, while sensitivity analysis shows that accounting for freshness constraints reduces freshness loss by 47.12% at an average total cost increase of 9.61% for robustness.

Why it matters — It provides logistics planners with a computationally viable method to balance the trade-offs between delivery costs, product quality, and schedule robustness when distributing highly perishable goods through cross-docking networks.

Caveat: The performance and trade-offs of the model are demonstrated using simulated benchmark instances rather than real-world distribution data.

Transportation Letters · doi

A glass-box approach to driving risk: Explainable Boosting Machines for interpreting transient behaviors in trajectory data

The study trained an Explainable Boosting Machine (EBM) on 16 million GPS records from 10,350 taxis in Beijing, segmenting the trajectories into 5.87 million 300-second windows. Using DBSCAN and K-Means clustering, 14.85% of these windows were classified as risky based on speed volatility, sudden stops, and abrupt heading changes. The EBM achieved 99.65% accuracy, matching the performance of black-box models like XGBoost while directly quantifying how individual driving behaviors contribute to risk scores.

Why it matters — It demonstrates that glass-box machine learning models can match the predictive accuracy of complex black-box models in traffic safety, providing a transparent, auditable framework suitable for real-time driver warning systems and usage-based insurance.

Caveat: The classification of 'risky' versus 'safe' driving windows relies entirely on unsupervised clustering of GPS-derived movement metrics rather than ground-truth accident or collision data.

Transportation Letters · doi

Automatic Kiwifruit Segmentation From Terrestrial LiDAR Point Clouds for Precision Agriculture

This study segmented kiwifruit from non-fruit elements in terrestrial LiDAR point clouds of a kiwi tree using Support Vector Machine (SVM) and Random Forest (RF) classifiers. The models utilized RGB color values and 12 geometric features, such as sphericity, local density, and normal vector components, optimized via LinearSVC-based Recursive Feature Elimination. Evaluation on a labeled test dataset showed that the RF model outperformed SVM across all metrics, notably achieving a 10% higher Precision score, while SVM was better at preserving the physical integrity of individual fruits.

Why it matters — It demonstrates that combining terrestrial LiDAR geometric features with machine learning can successfully isolate individual fruits in complex agricultural canopies, providing a foundational step for automated yield estimation and robotic harvesting.

Caveat: The models were trained and tested on a dataset derived from only a single scanned kiwi tree, which may limit immediate generalization to orchards with different training systems or foliage densities.

Transactions in GIS · doi

Modeling Expressway Crash Severity with Heterogeneity: Behavior-Specific Analysis of Speeding and Drowsy/Inattentive Driving with Text-Derived Features

The study analyzed 17,876 crash records from South Korean expressways between 2019 and 2024 using a Bayesian random threshold random parameter ordered probit model. It integrated structured crash data with unstructured police report narratives processed via TF-IDF and Latent Dirichlet Allocation to model severity across speeding, drowsy, and inattentive driving. The analysis revealed that drowsy driving on mainline segments and involving trucks yielded the most severe outcomes, while speeding crashes involving heavy vehicles also increased severity, though rainy conditions unexpectedly correlated with lower severity.

Why it matters — It demonstrates that integrating natural language processing of police narratives with heterogeneity modeling reveals distinct, behavior-specific risk factors that are otherwise masked in aggregate crash analyses.

Caveat: The findings are derived specifically from South Korean expressway data and may not directly generalize to urban arterials or different national road networks.

Transportation Research Record Journal of the Transportation Research Board · doi

Video Geospatial Mapping of Large-Scale Tower-Based Cameras Based on 3D GIS and Gradient Descent

The study introduces a geospatial mapping method that georeferences tower-based camera video by aligning it with a 3D GIS scene built from public DEM and high-resolution remote sensing data. It uses a constrained L-BFGS-B optimization framework to jointly optimize camera intrinsic and extrinsic parameters, followed by an outlier-removal and re-optimization step. Compared to standard methods like PnP, RANSAC-PnP, SQPnP, and DLT, this approach reduces the root mean square error of 3D spatial errors by 84.6%, 86.2%, 83.1%, and 69.4% respectively.

Why it matters — It enables accurate 3D video georeferencing and spatial analysis for large-scale tower cameras without requiring manual, on-site field calibration.

ISPRS International Journal of Geo-Information · doi

Relationship Between Courtyard Enclosure Forms and Autumn Dynamics of Façade Greening: A Field Study Based on Time-Series Monitoring of Microclimatic Fluctuations

This exploratory study monitored Boston ivy phenology and microclimates at six sample points across two-sided, three-sided, and four-sided courtyard enclosures with south-, west-, and north-facing façades. Using 5-minute interval temperature and relative humidity logs alongside fixed-ROI time-series imagery, the research tracked leaf color changes and defoliation, finding that the 50% defoliation index (DI50) occurred between October 20 and November 1. Spearman analysis of 60 matched daily observations revealed that median relative humidity correlated positively with leaf color retention (correlation coefficient of 0.70) and negatively with the defoliation index (-0.50).

Why it matters — It demonstrates that near-wall microclimatic fluctuations vary more significantly across different courtyard enclosure geometries than mean thermal-humidity conditions do, directly influencing the autumn phenology of vertical greening.

Caveat: Because courtyard enclosure forms and façade orientations were partly coupled in the experimental setup, their individual effects on the microclimate and plant phenology could not be isolated.

Urban Science · pdf · doi

Embark Now: User Demand Oriented Framework for Multi-day Urban Travel Itinerary Planning

A new multi-day travel itinerary planning framework integrates Large Language Models to capture user preferences with an enhanced Greedy Randomized Adaptive Search Procedure (GRASP) algorithm. Tested on 5,040 simulated user cases across real-world urban datasets from Beijing and Tianjin, the framework improved average total itinerary scores by at least 4.52% and 11.09% respectively compared to state-of-the-art methods, while reducing computation times.

Why it matters — It demonstrates that combining semantic LLM inputs with heuristic search algorithms can generate highly customized, multi-day urban travel plans faster and with higher satisfaction scores than existing iterative planning methods.

Caveat: The performance improvements are evaluated using simulated user cases on historical datasets rather than real-world deployment with active travelers.

arXiv · pdf

Unsupervised Detection of Entry and Exit Regions from Vehicle Trajectories for Camera-Agnostic Turning Movement Counts

An unsupervised pipeline was developed to identify intersection entry and exit regions directly from raw vehicle trajectories without requiring manual camera calibration or geometric annotations. The system clusters the start and end points of trajectories to generate persistent spatial polygons, which were evaluated across 17,100 pipeline runs using 9 surveillance cameras in Bengaluru, India, and 10 sequences from the UA-DETRAC dataset. Tested on 16 held-out cameras in Bengaluru, the pipeline achieved a median classification error of 3.4% and a median turning-movement GEH statistic of 2.43.

Why it matters — It automates the generation of turning movement counts from traffic cameras without the labor-intensive manual region-of-interest annotation typically required for each new camera deployment. By producing static spatial polygons rather than clustering individual trajectories, the system processes future vehicle paths at a linear computational cost.

Caveat: While more computationally efficient and stable across different camera views than trajectory clustering baselines, this pipeline yields a higher median classification error.

arXiv · pdf