Trucks are essential to urban economies through facilitating the movement of goods and services, but their presence significantly impacts urban infrastructure, traffic congestion, air quality, and noise pollution. Effective urban freight monitoring and emission estimation require a detailed understanding of urban truck distribution by their classes at scale. However, obtaining such data remains a significant challenge, especially in complex urban settings. This study presents a novel drive-by sensing framework using Google Street View imagery (GSV) combined with advanced computer vision and deep learning methods to generate fine-grained truck classification data at scale in urban areas. The framework consists of four main components: first, a customized YOLO-based detector for truck detection and classification from GSV imagery; second, Depth Anything V2 integration for RGB-based depth estimation to filter out distant and less visible objects; third, EfficientNet for fine-grained truck classification; and fourth, a case study on census tract-level correlation analysis between detected truck distribution and environmental data to demonstrate the effectiveness of the drive-by sensing approach. On held-out, panorama-disjoint, geographically blocked test sets, the detector achieves 65.2% precision and 56.9% recall, and fine-grained classification reaches 80.0% overall accuracy (macro-F1 = 0.59, averaged over ten training seeds) across seven commodity-informed truck categories. The GSV-based approach provides extensive spatial coverage for large-scale urban freight monitoring. A case study in Manhattan demonstrates the framework’s capability to identify spatial patterns in truck distribution, with an exploratory census tract-level analysis (N = 90 tracts) revealing a positive association between near-field truck prevalence and PM2.5 concentrations that persists under spatial regression models accounting for spatial autocorrelation. This research demonstrates the potential of computer vision and drive-by sensing technologies for scalable urban freight monitoring and environmental impact assessment.
As rail networks face increasing pressure to improve punctuality and capacity under mixed-traffic conditions, digital support tools such as Connected Driver Advisory Systems (C-DAS) are increasingly deployed. Yet, empirical evidence of their operational effects in daily practice remains limited. This paper presents a data-driven two-level approach based on macroscopic and microscopic event graph representations, constructed from the same underlying corridor and observed train movements, to evaluate the performance of C-DAS. It fuses timetable event records, track section occupation and release data, and C-DAS usage logs to enable comparison between trains operated with and without C-DAS at timetable point and block section levels. Derived from the event graphs, differences in timetable deviation, running time deviation, and the occurrence of unplanned yellow signal aspects are assessed using non-parametric statistical tests. The approach is demonstrated through its application to RouteLint, a C-DAS deployed for freight operations in the Netherlands, using one year of data from a mixed-traffic mainline corridor comprising 23,573 freight train runs. The results show modest but consistent differences in operational performance between trains operated with and without RouteLint. Statistically significant differences in timetable deviation are identified at multiple key locations, with RouteLint-equipped trains showing approximately 30–70 s lower deviations at major stations and junctions. At the microscopic level, running time deviations exhibit substantial local variability, while lower deviations are observed on selected block sections for RouteLint-equipped trains. In addition, RouteLint-equipped trains show approximately 6% lower overall odds that an observed yellow signal aspect is unplanned, while the association varies across operating contexts.
Reliable fault diagnosis of the Multifunction Vehicle Bus (MVB) is essential for ensuring railway operational safety. Spectral analysis shows that MVB signal magnitude is unevenly distributed and mainly concentrated below 10 MHz, suggesting that conventional uniform spectral partitioning may not align well with this characteristic. To better exploit the non-uniform spectral characteristics of MVB signals, we propose a fault diagnosis framework centered on adaptive spectral partitioning and complemented by temporal dependency modeling. Specifically, a Cumulative Magnitude Partitioning Gabor (CMP-G) method is developed to adaptively partition the spectrum into multiple bands based on the cumulative spectral magnitude distribution, on the basis of which a Multi-band Attention Gabor-GRU (MBAGGRU) model is constructed to perform band weighting and capture temporal dependencies. On a test set from an experimentally acquired dataset containing nine simulated MVB physical-layer states, the proposed method achieves 99.80% classification accuracy. To further evaluate robustness, the method is tested on three independently acquired test sets from separate acquisition sessions and under five interference types, including random single-tone frequency-domain interference, bounded additive time-domain noise, additive white Gaussian noise (AWGN), burst-transient interference, and fractional (1/fβ) noise. Across these test scenarios, the proposed method achieves competitive diagnostic performance compared with the STFT-based and wavelet-based baselines. In addition, controlled comparisons with multiple spectral partitioning schemes and alternative backbone classifiers further validate the design. The competitive performance maintained across all tested interference conditions indicates that the proposed framework is effective for MVB physical-layer fault diagnosis and shows potential for Prognostics and Health Management (PHM) in railway systems.
In this paper, we introduce HARNES, a heuristic and Software-Defined Networking (SDN)-enabled routing framework for dynamic vehicular networks. The framework enhances classical search algorithms by developing cluster-based extensions of Theta∗ and D∗-Lite, which minimize computational overhead through localized decision-making and line-of-sight (LOS) optimization. By deploying SDN controllers at roadside units (RSUs), HARNES enables centralized coordination among clusters, thereby improving adaptability and scalability in large-scale networks. The proposed approach is evaluated against EG-Dijkstra (Evolving-Graph Dijkstra) and A∗ on grid-based vehicular topologies using six performance metrics: computation time, discovery messages, path length, expanded nodes, average degree of expanded nodes, and path optimality ratio. Simulation results demonstrate that the cluster-based Theta∗ achieves up to 42% reduction in computation time, 88% fewer discovery messages, and 47% fewer expanded nodes, while maintaining near-optimal path quality within 3.1% of optimality. Additional comparison with the original non-clustered Theta∗ and D∗-Lite further confirms the effectiveness of the proposed cluster-based design. These findings highlight the potential of HARNES as a low-latency, scalable, and SDN-compatible routing solution for real-time intelligent transportation systems.
Inadequate coverage and unreliable data backhaul plague traditional networks in remote mountainous regions, ecological reserves, and disaster-affected areas. To address this, this paper proposes a UAV-relayed Air-to-Ground (A2G) integrated monitoring and data transmission system. The architecture unifies low-power ground sensors, a multi-band airborne communication platform, and a base station. Custom hardware is developed for environmental sensing and edge-intelligent recognition, supported by a lightweight, extensible packet protocol. Extensive field experiments characterize A2G link performance under both Line-of-Sight (LoS) and Non-Line-of-Sight (NLoS) conditions. Quantitative analyses reveal the relationships among signal strength, signal-to-noise ratio (SNR), packet loss rate (PLR), and spatial parameters such as UAV altitude, horizontal distance, and obstruction. System validation in a mountainous scenario demonstrates stable acquisition and reliable relay transmission of environmental parameters and target recognition data. Experimental results confirm robust operation under communication-constrained and complex topographical conditions, offering a lightweight, high-reliability solution for distributed monitoring in infrastructure-limited transportation and emergency response scenarios.
Modern automotive systems leverage deep neural networks (DNNs) for semantic segmentation and operate in two key application areas: (1) In-car, where the DNN solely operates in the vehicle without strict constraints on the data rate. (2) Distributed, where one DNN part operates in the vehicle and the other part typically on a large-scale cloud platform with a particular constraint on transmission bitrate efficiency. Typically, both applications share an image and source encoder, while each uses distinct (joint) source and task decoders. Prior work utilized convolutional neural networks for joint source and task decoding but did not investigate transformer-based alternatives such as SegDeformer, which offer superior performance at the cost of higher computational complexity. In this work, we propose joint feature and task decoding for SegDeformer, thereby enabling lower computational complexity in both in-car and distributed applications, despite SegDeformer’s computational demands. This improves scalability in the cloud while reducing in-car computational complexity. For the in-car application, we increased the frames per second (fps) by up to a factor of 11.7 (1.4 fps to 16.5 fps) on Cityscapes and by up to a factor of 3.5 (43.3 fps to 154.3 fps) on ADE20K, while being on-par w.r.t. the mean intersection over union (mIoU) of the transformer-based baseline that doesn’t compress by a source codec. For the distributed application, we achieve state-of-the-art (SOTA) over a wide range of bitrates on the mIoU metric, while using only 0.14% (0.04%) of cloud DNN parameters used in previous SOTA, reported on ADE20K (Cityscapes).
Traffic scheduling in inland ports is particularly challenging due to constrained waterways, complex spatial layouts, and highly interactive vessel behaviors. Crossing maneuvers, especially those involving vessels departing anchorages and those navigating main channels, frequently trigger localized congestion and amplify conflict propagation in confluence areas and ferry-crossing zones. Yet, most existing inland waterway scheduling methods rely on simplified motion assumptions and overlook the coupled conflict dynamics of such mixed-traffic interactions. To address these limitations, this paper proposes a dynamic and conflict-aware vessel traffic scheduling method specifically designed for mixed crossing scenarios in inland port waters. The method consists of three major components. First, we introduce a dynamic-aware trajectory generation module that produces dynamically feasible and collision-free trajectories, enabling accurate prediction of crossing times and tighter scheduling windows. Second, building on these trajectories, we develop a multi-candidate global trajectory-topology generation mechanism that provides diverse feasible path alternatives and avoids local-optimal scheduling caused by independent path selection. Third, we design a centralized event-driven scheduling algorithm that jointly optimizes max-min fairness, waiting time, and collision risk to ensure efficient and safe traffic coordination. Extensive experiments using real AIS data and synthetic traffic scenarios on a customized simulation platform show that the proposed approach achieves substantial improvements in fairness, navigational efficiency, and safety compared with representative baseline scheduling policies.
Cross-domain generalization remains a critical challenge in LiDAR-based 3D object detection, primarily due to significant variations in point cloud distributions caused by different sensor configurations, installation angles, and environmental conditions. While prior research has focused extensively on network-level adaptations, the role of point cloud preprocessing has received limited attention. In this paper, we present a systematic study on how structural differences in raw point clouds affect the generalization performance of 3D detection models, and we propose a Point Cloud Structure Unification (PCSU) framework to mitigate these disparities. PCSU comprises three modular components: height normalization to align vertical distributions, reflectivity suppression to reduce overfitting to sensor-specific intensity patterns, and resolution standardization via voxel center alignment. Through extensive experiments across multiple datasets and detection models, we demonstrate that these structure-aware preprocessing strategies generally enhance cross-domain generalization without modifying model architectures. The improvements are most prominent for larger objects and near-to-moderate ranges. Our findings highlight the often-overlooked importance of data-level harmonization and provide actionable insights for designing more generalizable 3D detection pipelines.
Intelligent Transportation Systems (ITS) generate large volumes of real-time traffic data that require accurate forecasting, timely anomaly detection, and reliable deployment strategies to maintain operational efficiency. Existing ITS solutions often treat prediction, anomaly detection, and system deployment as separate processes, resulting in limited adaptability and increased deployment risk in dynamic traffic environments. To address this gap, this paper proposes a Deep Learning–Reinforced Self-Adaptive CI/CD Framework that integrates traffic forecasting, anomaly detection, and adaptive deployment strategies within a unified DevOps pipeline using GitLab and Ansible. The framework employs a sequence-to-sequence LSTM for short-term traffic forecasting and an LSTM autoencoder for anomaly detection, both embedded within a closed-loop CI/CD pipeline enhanced by reinforcement learning for deployment optimization. Experiments on the Metro Interstate, PeMSD7_228, PeMSD7_1026, and a combined heterogeneous dataset show that the proposed framework consistently outperforms statistical and deep learning baselines. It achieves up to 21.8% reduction in MAE and 19.0% reduction in RMSE compared with LSTM baselines, with a best RMSE of 20.5 on the Metro dataset and 22.2 on the combined dataset. The anomaly detection module achieves an AUROC of 0.94 and reduces detection delay to 6.4 seconds, outperforming GRU-AE and Isolation Forest baselines. The adaptive deployment mechanism reduces Change Failure Rate (CFR) to 4.3% and rollback rate to 3.1%, while improving system availability to 98.6% and reducing Mean Time to Recovery (MTTR) to 11.3 minutes. These results demonstrate that embedding AI-driven analytics into CI/CD workflows improves both predictive intelligence and deployment resilience, providing a scalable and self-adaptive solution for real-time ITS monitoring.
Intelligent transportation systems (ITS) generate vast heterogeneous data from roadside units (RSUs), traffic management centers (TMCs), and vehicular networks, posing challenges for privacy, scalability, and cross-domain learning. This paper proposes a Hierarchical Graph Federated Learning (HG-FL) framework that integrates multi-level aggregation, spatio-temporal graph modeling, and autoencoder-based feature compression to enable privacy-preserving, distributed ITS analytics. The framework mirrors real-world ITS hierarchy through three tiers: local RSU-level training, regional TMC-level aggregation, and global cross-domain coordination. Experiments conducted on the CIC-IoV 2024 Decimal Dataset and the NF-ToN-IoT-v2 dataset demonstrate that HG-FL achieves superior performance compared to centralized, flat federated, and graph-only baselines. Specifically, on the combined dataset, HG-FL attains an AUROC of 0.964, AUPRC of 0.947, and F1-score of 0.929, while reducing communication cost to 162 MB and convergence rounds to 15. For SLA violation risk prediction, it achieves an RMSE of 0.121 and a Brier score of 0.059, outperforming baseline methods. These results highlight the framework’s scalability, robustness, and effectiveness in achieving cross-domain generalization and service-level assurance within privacy-preserving ITS environments.
Planning public electric vehicle charging stations (EVCSs) in emerging urban environments requires balancing operator profitability and user accessibility while accounting for traffic congestion and uncertainty in future demand and economic conditions. However, most existing EVCS planning studies focus primarily on identifying nominal optimal solutions and provide limited insight into the reliability of deployment decisions under uncertain conditions. To address this gap, this study develops a congestion-aware bi-objective framework for EVCS siting and sizing that simultaneously maximizes annualized operator profit and minimizes aggregate additional user cost, including travel, queueing, and energy-consumption components. The study integrates metaheuristics with congestion due to peak period travel, queueing-based service assessment, Monte Carlo uncertainty analysis, reliability evaluation, and elasticity-based sensitivity assessment within a unified decision-support methodology. The approach is demonstrated for the city of Guwahati, India represented by a 53-node, 76-link road network. Results highlight that robust EVCS deployment solutions consistently converge to a narrow infrastructure envelope comprising approximately 16–18 charging stations, 366–390 chargers, and a peak demand of 22–23 MW despite uncertainty in key techno-economic parameters. Reliability analysis shows that selected compromise solutions maintain profitability and affordability targets with high probability under stochastic perturbations, while sensitivity analysis identifies charging demand, retail tariffs, and fleet battery composition as the principal drivers of outcome variability. The proposed framework provides a reproducible and reliable methodology for congestion-aware EVCS planning under uncertainty.
This work introduces “VIsual-COgnitive Distraction through Eye-tracking and Vehicular measurements” (VICODEV), an interpretable, statistically driven index combining five vehicular and behavioural measurements to estimate induced visual-cognitive distraction. The index was developed through an experimental campaign at the dynamic driving simulator of Politecnico di Milano using an Instrumented Steering Wheel, eye-tracking glasses, an electroencephalograph, and an electrocardiograph. Thirty-five novice drivers were recruited for the tests. Alongside the primary driving task, drivers performed five repetitions of a secondary “Clock Task”, inducing visual-cognitive distraction. During each secondary task, participants solved mathematical operations displayed on a tablet inside the cockpit. Twenty-one measurements were collected to describe induced distraction, including vehicular, behavioural, and physiological types. The Wilcoxon Signed Rank Test revealed twelve measurements with statistically significant trends between normal and induced distracted driving. Five measurements were then selected considering real-time in-vehicle applicability, namely Standard Deviation of Lateral Position of the vehicle’s center of gravity to the lane centerline, standard deviation of steering wheel grip force, Eyes Off-Road Time, rate of high-velocity eye movements, and mean duration of low-velocity eye movements. VICODEV combines these five measurements via canonical discriminant analysis, assessing induced visual-cognitive distraction every 40 ms. Leave-one-repetition-out cross-validation achieved an average recall of 99.47% and accuracy of 95.20%, confirming the index consistency across secondary task repetitions. Leave-one-subject-out cross-validation yielded a median Area Under the Receiver Operating Characteristic Curve of 82.9%, showing the index’s ability to identify shared patterns across the tested population. This study offers a standardised reference to support DMS design, evaluation, and comparison.
Accurate and real-time visibility estimation under blowing-snow conditions is crucial for ensuring roadway safety and maintaining operational continuity in cold-region transportation networks. This study proposes a lightweight visibility detection framework based on an improved YOLOv8 (You Only Look Once v8) architecture, termed YOLO-SDlite. Through module-level modifications and structural simplification, YOLO-SDlite achieves high classification accuracy while substantially reducing computational cost. In particular, we introduce a CSPDW (cross-stage partial (CSP) connections + depthwise separable convolution) module that uses depthwise convolution (DWConv) for local feature extraction and channel alignment, preserving CSP characteristics while cutting complexity. We also incorporate an improved omnidimensional dynamic convolution (ODConv Lite), bringing multidimensional dynamic attention into the feature extractor to adaptively recalibrate along the spatial domain, input channels, output channels, and convolution kernels. Experimental results demonstrate that the model effectively grades visibility levels in real time, supporting reliable performance on resource-constrained roadside devices. Specifically, it attains a 68.8% reduction in parameters and 66.7% reduction in model size compared with the baseline YOLOv8, while sustaining 93% classification accuracy. Building upon these perception capabilities, a coordinated variable message sign (VMS) control strategy is designed to translate model outputs into actionable traffic management measures, including dynamic speed limits, warnings, and closure operations. The integrated framework bridges environment perception and operational decision-making, providing a scalable solution for intelligent, weather-responsive traffic management.
The Internet of Vehicles (IoV) enhances transportation through connectivity and intelligence. However, this integration exposes IoV systems to various cyber threats, such as data breaches, hacking, vehicle control manipulation, and vehicle identity theft, which must be addressed to ensure system security. This study tackles security, scalability, and efficiency issues in the IoV framework by implementing a decentralized, blockchain-based authentication framework. It addresses cryptographic vulnerabilities, reduces computational and communication costs, and improves real-time performance. The framework also reduces latency and communication overhead, ensuring secure connectivity in high-traffic IoV scenarios. Experiments demonstrate the effectiveness of the techniques deployed, including secure mapping, distributed verification, verified connectivity, and decentralized validation. The results of performance evaluations demonstrate that the proposed framework has improved computing efficiency, high communication effectiveness, and reduced overall delay. The system is also shown to be adaptable to various node densities and communication conditions. The experimental results show that the proposed framework achieves a computation cost of 4.29 ms, representing a 48.19% improvement over the best-performing existing methods. Similarly, the total communication cost of the framework is 176 ms, showing reductions of 65.22%, 51.38%, 76.56%, and 74.04% compared to previous approaches. Lastly, a security analysis confirms that the framework establishes a solid security foundation, and the results of a comparative analysis indicate its potential to strengthen the security of IoV networks.
Shared steering control can reduce driver workload and improve vehicle stability, but its effectiveness depends on how control authority is allocated between the driver and automation. Existing allocation strategies often rely on geometric tracking errors, such as lateral offset and heading error, which may not reflect human sensitivity to lane deviations and typically do not adapt to individual driver capability. This paper proposes a driver-aware indirect shared steering framework using a Bayesian-Optimized Soft Actor-Critic (BO-SAC) authority allocator. The allocator incorporates a nonlinear driver-awareness error identified from manual-driving data, an activity-weighted driver skill factor updated only during active steering, and a situation severity index to personalize authority allocation according to both driver capability and driving demand. A Control Lyapunov Function (CLF)-guided reward-scheduling mechanism adjusts the tracking–comfort–stability trade-off according to the current stability condition. The automation channel is equipped with a time-varying gain fixed-time controller that provides predefined-time convergence under automation-channel operation and stabilizing assistance in the shared-control loop. Simulations and hardware-in-the-loop experiments show that the proposed framework reduces objective workload-, comfort-, and stability-related indices compared with fuzzy, RL-based, TD3-based, and sliding-mode-control baselines, while accepting a controlled tracking-accuracy trade-off. The results demonstrate the benefit of combining perception-aware error shaping, active driver skill estimation, and stability-aware reward scheduling for adaptive shared steering.
In this paper, we quantify the influence of varying road infrastructure and Human Machine Interface (HMI) types on cyclists’ overtaking behavior in a bicycle simulator. We specifically evaluate whether cyclists overtake an automated shuttle bus (ASB) or stop by varying bus and bicycle infrastructure, HMI type, and traffic conditions. We find that bicycle infrastructure and HMI type have a significant influence on cyclists’ overtaking behavior. The main contribution of this paper is quantifying how different types of road infrastructure and HMI designs affect cyclist-ASB interactions, providing evidence that road user interaction studies with automated vehicles must consider infrastructure variations for generalizable results.
Capturing and reconstructing dynamic urban scenes in 3D is central to real-world simulations and applications such as autonomous driving. Building on recent advances in 3D Gaussian Splatting (3DGS), which offers high-fidelity and efficient scene representations, we present EgoDepthGS, a camera-only approach for 3D reconstruction in dynamic driving environments. Unlike prior methods that require LiDAR supervision to initialize splat placement, EgoDepthGS achieves high-quality reconstruction from monocular image sequences alone, eliminating reliance on expensive depth sensors. We define a non-iterative depth estimation technique that derives per-pixel depth directly from optical flow, projecting it into world space using the pinhole camera model without iterative optimization. To ensure temporal consistency, we introduce a motion compensation module that corrects depth distortions induced by ego-motion. By disentangling the rotational component of the optical flow, our method remains robust during both straight driving and turns. EgoDepthGS matches existing approaches, including StreetGaussians, across standard metrics (e.g., LPIPS), demonstrating that integrating classical geometric vision with 3DGS provides a cost-effective, sensor-minimal, and scalable solution for modeling complex urban scenes. Our project page is at https://git4confer.github.io/ABC/.
The rapid growth of electric vehicles (EVs) has intensified demand for EV charging across time and space, overloading the current charging infrastructure and continuously increasing operational challenges for pricing and scheduling. Current studies on joint optimization of pricing and scheduling for EV charging often overlook critical real-world factors such as charging path selection, multiple aspects of charging costs, and the influence of future EV useers on the experience of current EV users, resulting in solutions misaligned with practical scenarios. To address these gaps, we consider an EV charging model that comprehensively captures scheduling decisions and charging costs. Then, we propose a novel reservation-based pricing and scheduling framework to mitigate service disruptions to current EV users caused by future EV arrivals. Next, we formulate EV charging pricing and scheduling as a complex Stackelberg game with a charging station (CS) company as the leader and an EV agent as the follower. Finally, we convert the game into a large-scale bi-level mixed-discrete-and-continuous optimization problem and tackle it using decomposition techniques and greedy methods. Numerical results on real-world datasets for Manhattan and Nanshan demonstrate that the proposed reservation-based framework outperforms the traditional reservation-free framework, and the proposed modeling, formulation, and algorithms together achieve a good tradeoff between performance and computation time compared to existing approaches, showcasing the practical value of the proposed solution framework.
Traffic flow prediction is a critical task in Intelligent Transportation Systems (ITS), essential for alleviating traffic congestion and optimizing resource allocation. While graph neural network-based methods have achieved significant success in modeling pairwise relationships between nodes, they often struggle to capture the dynamic, high-order spatio-temporal dependencies inherent in traffic networks. Furthermore, in distributed real-world scenarios, data heterogeneity among different clients often leads to negative transfer in traditional federated learning methods, thereby limiting model performance. To address these challenges, this paper proposes a novel Adaptive Parameter Coordination Dynamic Hypergraph Spatio-Temporal Prediction (ADH-STPC) framework for decentralized learning. This framework centers on the Client-Adaptive Parameter Sharing (CAPS) module, which coordinates decentralized collaboration by quantifying the similarity of global and local feature distributions, ensuring that parameter propagation occurs only between compatible clients, thereby suppressing negative propagation. Additionally, to accurately predict changes in traffic flow, we designed the Dynamic Hypergraph Convolution Architecture (DHCGA), which captures higher-order nonlinear dependencies by adaptively evolving hypergraph structures, thereby uncovering latent relationships among features. CAPS provides the core topology for adaptive collaboration, while DHCGA captures spatio-temporal traffic features, and together they drive the model to achieve optimal performance. Extensive experiments on two large-scale real-world traffic datasets validate the effectiveness of this framework. Compared to existing state-of-the-art baseline models, ADH-STPC demonstrates superior performance: achieving significant reductions across all key metrics on the BJSubway dataset, including a 15.1% decrease in MAE, a 15.2% reduction in RMSE, and a 7.1% improvement in MAPE.
Accurate city-wide traffic prediction is essential for intelligent mobility management in modern urban environments. This work introduces a novel approach that leverages aggregated cellular network activity as a large-scale, infrastructure-independent data source for predicting travel time and speed. We first develop the Network Insight Model, which demonstrates that cellular activity patterns alone can effectively capture city-wide traffic dynamics. We then propose the Integrated Mobility Model, a multimodal fusion architecture that combines cellular network activity with Bluetooth sensor data to further enhance prediction accuracy. Both models incorporate attention mechanisms, enabling interpretable insights into which intersections and road segments most strongly influence traffic conditions. Using real-world datasets from York Region, Canada, we evaluate our models against standard regression baselines and state-of-the-art deep learning approaches. Across multiple metrics, including MAE, RMSE, R2, MAPE, and the Travel Time Index (TTI), our models consistently achieve superior predictive performance while providing meaningful, attention-based explanations of traffic patterns.
