Yubo Jiao

High manual annotation costs and poor generalizability often limit the robustness of deep learning models for pavement crack segmentation. This study addresses these challenges by developing and validating a diffusion-based generative artificial intelligence framework to support data-efficient crack segmentation models for infrastructure inspection. The proposed framework formalizes the generation and utilization of synthetic, pixel-level annotated crack data and is empirically evaluated across multiple learning paradigms. To implement and validate the framework, an annotated synthetic dataset (AI500) was generated using Google’s Imagen 4 Ultra and designed to capture diverse crack morphologies, lighting conditions, and surface textures. The utility of the synthetic data was systematically assessed through three experimental strategies: synthetic-only training, hybrid dataset training, and sim-to-real transfer. Model performance was evaluated using three deep learning architectures and five real-world datasets. Results show that synthetic-only models provide a robust, transferable performance baseline, achieving cross-domain generalization comparable to the target-only baseline (TOB) for models trained on real-world datasets. Augmenting AI500 with only 10% real-world data yielded performance comparable to TOB. Furthermore, sim-to-real transfer using the same 10% fraction achieved near parity with TOB. Cost analysis revealed the potential for double-digit savings without compromising model robustness. Additionally, the development of G3P-500 using Gemini 3 Pro demonstrated superior photorealism and diversity, requiring minimal human verification, with only 3.2% of pixels added to generated annotations. These findings demonstrate that high-quality synthetic data within a structured generative framework can substantially reduce annotation requirements while enabling accurate, generalizable pavement crack segmentation models. However, further validation remains necessary in more challenging environments.

Transport emissions remain substantial, while eco-driving studies are often constrained by incomplete context information, inconsistent evaluation methods, and ineffective driver guidance. This study proposes an agentic framework integrating emission prediction, decision evaluation, and driver-facing intervention. The emission prediction component combines large language model (LLM)-enabled road-context and vehicle-proxy agents with a supervised machine learning model to estimate directed-link CO$_2$ emissions. The evaluation component couples reinforcement-learning policy and evaluation agents in a policy--evaluation feedback loop for route--style optimization through consistent CO$_2$ and travel-time calculations. The intervention component employs an LLM-enabled advice agent to translate selected decisions into natural-language guidance. A Montreal case study shows that enriched covariates improve CO$_2$ prediction, that emissions and travel time should be evaluated jointly, and that optimized decisions can be communicated as practical guidance. This framework delivers a transparent, deployable workflow for eco-driving strategies, bridging the gap before electric and autonomous vehicles dominate the market.

Urban origin–destination (OD) data often come from two distinct sources: traditional surveys and global positioning system (GPS)-based records. The former offer rich behavioral detail but are infrequent and costly, while the latter provide continuous coverage but lack semantic context. They often lead to different findings and are rarely examined together, making it difficult to build a consistent understanding of urban travel behavior. To address this gap, we propose a dynamic joint latent factor model that decomposes multi-source OD matrices into shared spatial structures and source-specific temporal dynamics. The model identifies latent movement patterns by jointly factorizing both datasets while allowing each source to retain its unique temporal characteristics. Applied to GPS and survey OD data from Ottawa, Canada, the model achieves strong reconstruction accuracy ( R 2 ≈ 0 . 8 overall). The results uncover stable spatial patterns alongside clear differences in the traveler groups driving each latent mode. GPS factors emphasize younger and higher-income travelers, especially in the more flexible afternoon period. However, survey factors reflect routine commuting by middle-aged and middle-income households. These patterns show that the joint model isolates behavioral differences from sampling bias and provides a coherent representation of OD demand across sources. Together, these results show that the joint factorization quantifies behavioral differences from sampling bias and provides a coherent representation of OD demand across sources. This framework helps compare and interpret multi-source OD data in settings where traditional surveys are limited.

Drivers’ mental and physiological states critically influence traffic safety, making reliable estimation essential for timely warnings and interventions. However, these states evolve over time and are only indirectly observable through noisy, modality-dependent responses, which can induce biased labels and fragile detectors. To improve the accuracy and reliability of state estimation, we propose a unified nonlinear state-space framework that jointly models the dynamics of latent driver states, the influence of driving context, and their effects on multi-sensory responses. Within this framework, an offline multi-view clustering method fuses complementary indicators to establish robust benchmark states, while an online ResNet-Transformer maps multi-sensory data to states in real time. We evaluated the proposed framework in a driver-fatigue case study, where fatigue labels were inferred offline from self-reports, reaction-time variability, and eye-aspect-ratio variability; and were subsequently detected online using blood volume pulse, electrodermal activity, and facial images. Our results show that the offline-derived levels exhibit statistically significant separation across multiple fatigue indicators, validating the benchmark labels, while the online model achieved 82.5% mean accuracy with 10-second windows and up to 89.5% with 30-second windows, outperforming representative baselines without evidence of class-specific bias. These findings demonstrate that the proposed state-space representation effectively captures latent driver-state dynamics and that the online-offline estimation approach offers reliable and effective means of inferring these states from diverse observational sources, supporting proactive warnings and safer human–vehicle interaction.

Identification of causal factors in traffic crashes has always been a significant challenge in road safety studies. Traditional crash prediction models are limited in elucidating the underlying causal mechanisms in road crashes. This research explores the application of three graphic models, namely, the Gaussian graphical model (GGM), causal Bayesian network (CBN) and graphic extreme gradient boosting (XGBoost), through a case study using highway–railroad-grade crossing (HRGC) inventory and collision data from Canada. The three modelling approaches have generally yielded consistent findings on various risk factors such as crossing control type, track angle, and exposure, showing their potential for identifying causal relationships through the interpretation of causal graphs. With the ability to make better causal inferences from crash data, the effectiveness of safety countermeasures could be more accurately and reliably estimated.

Understanding the driving state (e.g., fatigue state) is critical in improving the safety of urban rail transit operations. This study reports a study that used simulation experiments to evaluate the effects of fatigue on professional drivers’ psychological, biological, and physical responses and thus develop a fatigue detection model, enabling early warning and ensuring operational safety. Specifically, we applied statistical analytics on self-reported fatigue feelings, the Mackworth Clock Task, the Stroop Test, and the Iowa Gambling Task to evaluate the impact of fatigue on driver performance. The results show that fatigue reduced participants’ reaction time and cognitive ability and led professional drivers to make more conservative decisions. Moreover, we utilized heart rate signals, electrodermal activity, and eye movements from wearable devices and cameras to build a driver fatigue detection model with machine-learning methods. We analyzed the impact of different time window lengths on fatigue detection. The results indicate that as the length of the time window increases, the data captures more comprehensive information, leading to improved fatigue detection accuracy. Furthermore, multi-feature fusion significantly enhanced model performance. With feature fusion, even short-term time windows can encapsulate sufficient information for effective fatigue detection. These findings not only contribute to a deeper understanding of the risks posed by fatigue in professional driving, but also pave the way to developing a real-time, non-invasive fatigue detection system, enhancing traffic safety.

This paper investigated the feasibility and reliability of employing various physiological measures -for determining drivers’ fatigue levels, which may ultimately lead to a solution for real-time detection of driver fatigue state for improving driving and traffic safety. An experimental study was conducted to collect the data, including fatigue levels assessed via the Karolinska sleepiness scale and heart rate variability (HRV) and electrodermal activity (EDA) features. Based on an extensive statistical analysis of the collected data, significant differences in numerous HRV and EDA features were found across varying fatigue levels. Employing several machine learning techniques for classification purposes, the most favorable binary classification performance was achieved using the Light Gradient Boosting Machine classifier, with an accuracy rate of 88.7% when HRV and EDA features were utilized as inputs. Meanwhile, for three-class classification, the accuracy decreased slightly to 85.6% when employing the Random Forest classifier. These outcomes underscore the potential of HRV and EDA feature fusion in capturing diverse physiological responses to fatigue, thereby bolstering fatigue detection performance. Besides, subject-independent classification yielded an accuracy of 52.0% and 53.3%, reflecting the potential bias introduced by unobserved heterogeneity in classification models. Moreover, feature selection should be prioritized over dimensionality reduction in feature fusion endeavors to diminish feature redundancy and prevent information loss. The findings of this study could contribute to the development of reliable driver fatigue detection methodologies utilizing readily available measures of physiological response measures, such as HRV and EDA features.

Pilots’ excessive mental workload could reduce their ability to perform concurrent tasks during emergency flights, which is one of the most critical aviation safety concerns. Several past efforts have attempted to investigate the underlying issues, but all had limited success owing to the challenge of collecting representative data under realistic operating conditions. This study aimed to address this challenge by conducting a flight simulator study involving a comparatively large number of participants, who were pilot cadets with flight experience, and using noninvasive functional near-infrared spectroscopy (fNIRS) to collect the pilots’ brain activity data. Pilots’ subjective ratings and brain activity records were collected over a total of 75 simulated flights under three subtask scenarios comprising different equipment failures. A statistical analysis was carried out on the subjective ratings and on the changes observed in the saturation of the oxyhemoglobin (ΔOxyHb) of individual fNIRS channels. The mental workload of the pilots was classified using a support vector machine hierarchical combination classifier, focusing on the question of whether it is feasible to classify pilots’ mental workload using brain activity signals (i.e., ΔOxyHb). The results suggested that the pilots’ mental workload levels were highly associated with the ΔOxyHb measures as well as with the activities of different brain regions, including the prefrontal-, motor-, and occipital cortex. The findings from this study could provide a reference for optimizing pilot training systems and improving pilot performance during emergency flight operations.