Toward Safe and Scalable Intelligent Transportation Systems: A Survey of AI Methodologies and Deployment Challenges

Journal of Computer Technology and Software ISSN: 2998-2383 Vol. 4, No. 9, 2025 Toward Safe and Scalable Intelligent Transportation Systems: A Survey of AI Methodologies and Deployment Challenges Eulalie Merrick Eastern Washington University, Cheney, USA [email protected] Abstract: Artificial Intelligence (AI) is fundamentally reshaping Intelligent Transportation Systems (ITS) by enabling real-time perception, prediction, and decision-making for connected and autonomous mobility networks. The convergence of deep learning, reinforcement learning, and computer vision with advanced sensing and communication technologies empowers vehicles and infrastructure to collaboratively manage traffic, improve safety, and reduce environmental impact. AI-driven methods have demonstrated superiority over traditional rule-based approaches in traffic forecasting, route optimization, and autonomous vehicle control, while edge computing and 5G/6G connectivity are making large-scale deployment increasingly feasible. However, challenges remain in data heterogeneity, model interpretability, safety validation, cybersecurity, and regulatory compliance. This survey provides a comprehensive review of AI methodologies applied to ITS, covering traffic prediction, perception for autonomous driving, multi-agent control, and smart infrastructure optimization. We analyze the current state of deployment, highlight open challenges such as privacy-preserving learning and robust decision-making under uncertainty, and propose future research directions including AI-enabled digital twins, federated learning for vehicular networks, and scalable edge intelligence. Keywords: Artificial Intelligence; Intelligent Transportation Systems; Traffic Prediction; Autonomous Vehicles; Deep Reinforcement Learning; Edge Computing; Digital Twin; Federated Learning; Computer Vision; Smart Mobility. 1. Introduction The global transportation ecosystem is undergoing a profound transformation as Artificial Intelligence (AI) becomes the central enabler of Intelligent Transportation Systems (ITS). Traditional traffic management relied on static models and human-designed control strategies, which cannot effectively handle the explosive growth of real-time traffic data from connected vehicles, roadside sensors, smartphones, and Internet of Things (IoT) devices [1]. In contrast, AI-powered systems leverage large-scale datasets and advanced algorithms to enable perception, prediction, and decision-making that adapt dynamically to evolving traffic conditions. These systems aim to achieve four fundamental goals: improving safety by reducing accidents and human error, optimizing traffic efficiency by alleviating congestion, enhancing sustainability by reducing energy consumption and emissions, and enabling user-centric mobility services such as personalized routing and shared autonomous fleets [2]. The proliferation of Connected and Autonomous Vehicles (CAVs), coupled with next-generation wireless networks (5G/6G) and Edge Computing, has created an environment where vehicles, roadside infrastructure, and cloud services collaborate seamlessly. For example, connected vehicles can share sensor data to collectively detect obstacles, while smart traffic lights can adapt signal timing based on AI-driven predictions of congestion [3]. Deep neural networks have advanced perception tasks such as object detection, lane boundary recognition, and pedestrian intention prediction, enabling autonomous vehicles to navigate complex urban environments [4]. At the same time, Deep Reinforcement Learning (DRL) has shown strong potential in optimizing signal control, cooperative driving, and platooning by learning adaptive strategies from real-world traffic patterns [5]. Despite rapid advances, deploying AI in large-scale ITS faces critical challenges. Data collected across heterogeneous fleets is often noisy, incomplete, and biased, limiting the generalization of AI models. The black-box nature of deep learning complicates safety certification and explainability, raising concerns for regulators and the public [6]. Adversarial attacks can manipulate sensor inputs, threatening the reliability of perception systems [7]. In addition, privacy regulations such as GDPR and CCPA constrain centralized data sharing, motivating research into federated learning and privacypreserving analytics [8]. Scalability is another barrier: AI models must run efficiently on resource-constrained edge devices and deliver real-time decisions under low-latency requirements [9]. The integration of AI with ITS is now expanding beyond perception and control to include system-level intelligence. Digital twins are emerging as virtual replicas of transportation networks, enabling predictive simulation and policy testing before real-world deployment. Federated learning allows distributed model training across vehicles and roadside units (RSUs) without exposing raw data, addressing privacy and bandwidth concerns. Edge-cloud collaborative architectures are redefining how AI inference and training are distributed between vehicles, infrastructure, and centralized servers [10]. These innovations are expected to accelerate the transition toward fully autonomous, resilient, and adaptive mobility ecosystems. This survey provides a holistic and forward-looking review of AI-driven ITS. Unlike earlier works focusing narrowly on individual tasks such as traffic forecasting or single-vehicle autonomy, we examine the entire technology stack, including perception, prediction, control, communication, and infrastructure intelligence. We synthesize advances across deep learning, reinforcement learning, graph neural networks (GNNs), digital twins, and federated AI, and analyze their potential to revolutionize transportation. Furthermore, we discuss the deployment challenges in scalability, safety, privacy, and interoperability, providing guidance for both researchers and practitioners seeking to design robust and trustworthy AI-enabled transportation systems. 2. Related Work and Current Landscape Research on applying artificial intelligence to Intelligent Transportation Systems (ITS) has evolved over multiple decades, transitioning from early rule-based models to modern data-driven deep learning approaches. Early surveys primarily focused on traffic flow prediction using statistical methods such as autoregressive integrated moving average (ARIMA) or Kalman filtering, which were limited in handling nonlinearity and spatiotemporal complexity. Lv et al. [9] pioneered a deep learning perspective by reviewing how stacked autoencoders and recurrent neural networks (RNNs) can model complex traffic dynamics, outperforming traditional time-series models. However, their work did not consider the emergence of graph neural networks (GNNs) that now dominate traffic forecasting by capturing both spatial road network topology and temporal dependencies. Zhang et al. [10] later provided a broad review of deep learning for traffic prediction, but their focus was primarily on algorithmic accuracy, with little discussion of real-time deployment challenges in edge and vehicular networks. Another major body of literature centers on traffic signal control and adaptive routing. Khamis and Gomaa [11] surveyed reinforcement learning (RL)-based traffic light control, emphasizing Q-learning and deep Q-networks (DQN) to optimize signal phases. Feng et al. [12] extended this by considering multi-agent reinforcement learning (MARL), which is critical in large urban networks with numerous interacting intersections. However, most of these reviews fail to address the scalability of MARL in highly dynamic, partially observable environments or the role of V2X communication in enhancing learning efficiency. More recent works highlight how graph reinforcement learning (GRL) integrates traffic topology into policy learning, but systematic reviews remain scarce. In the autonomous driving domain, Grigorescu et al. [13] examined deep learning for perception tasks such as object detection, semantic segmentation, and sensor fusion, but their analysis predates the widespread adoption of transformer-based architectures that now dominate perception benchmarks. Kuutti et al. [14] surveyed deep reinforcement learning for vehicle control and decision-making, focusing on continuous control policies but neglecting hybrid architectures that combine safety-critical rule-based planners with neural network policies to meet reliability standards. Similarly, recent reviews of multisensor fusion in autonomous driving, such as those by Feng et al. [15], provide insights into LiDAR-camera fusion and radar perception but lack discussion of end-to-end differentiable sensor fusion pipelines and their real-time deployment constraints on embedded platforms. Other emerging areas such as connected autonomous vehicles (CAVs) and vehicular edge intelligence have received limited integrated treatment. Molina-Masegosa et al. [16] surveyed vehicular communication standards (V2V, V2I, V2X) but did not consider how federated learning (FL) and edge AI enable distributed model training and inference while preserving privacy. Habibzadeh et al. [17] explored IoT-based smart cities, mentioning AI-driven transportation management but failing to examine the synergy between AI and digital twin technology for predictive infrastructure planning and real-time network simulation. Similarly, early works on cloud-based ITS architectures (e.g., Zhang et al. [18]) focused on offloading computation to centralized servers but did not anticipate the rapid rise of edge-cloud collaborative architectures where inference is distributed for low-latency applications. More recently, several reviews attempt to unify these perspectives but remain fragmented. Zhou et al. [19] provided an overview of deep reinforcement learning in transportation but focused narrowly on traffic control without discussing autonomous vehicles or digital twin integration. Min et al. [20] introduced AI-powered digital twins for traffic simulation and planning but lacked discussion on the communication and security challenges that arise when integrating AI with largescale connected vehicles. Xu et al. [21] surveyed federated learning in vehicular networks, outlining privacy benefits but offering limited analysis on model aggregation under intermittent connectivity and adversarial participants. The contribution of this survey is to bridge these fragmented perspectives into a single, cohesive analysis. We examine AI-driven ITS as a multi-layered ecosystem that spans from perception (object detection, semantic mapping) to prediction (traffic flow, trajectory forecasting), control (reinforcement learning for signal timing and cooperative driving), and system-level intelligence (digital twins, federated learning, and edge-cloud collaboration). We integrate lessons from both academic research and industrial deployment to highlight what is technically feasible today and what challenges remain unsolved for real-world scalability and safety. By doing so, this survey moves beyond algorithmic performance and provides a systems-oriented, forward-looking roadmap for AI in transportation. 3. AI Methodologies for Intelligent Transportation Systems Artificial Intelligence (AI) methods in Intelligent Transportation Systems (ITS) can be organized into several core dimensions that collectively enable perception, prediction, control, and system-level optimization. These methodologies include deep learning for perception and understanding, graphbased and sequence models for traffic prediction, reinforcement learning for decision-making and control, and system-level intelligence frameworks such as federated learning, edge computing, and digital twins. Together, they form the foundation of modern AI-enabled transportation ecosystems capable of handling dynamic, data-rich, and safety-critical environments. 3.1 Deep Learning for Perception and Understanding. Perception is the cornerstone of AI-driven ITS, allowing vehicles and infrastructure to interpret complex road environments. Early approaches relied on classical computer vision techniques such as handcrafted feature extraction (HOG, SIFT) for object detection and lane recognition, but these methods struggled in unstructured and highly variable urban scenes. The rise of deep learning, especially convolutional neural networks (CNNs), transformed perception by enabling end-to-end feature learning from large-scale annotated datasets. Models such as Faster R-CNN and YOLO have been widely adopted for real-time object detection of vehicles, pedestrians, and traffic signs [22]. For semantic segmentation and scene understanding, architectures such as SegNet, DeepLab, and more recently transformer-based models like DETR and BEVFormer provide high-resolution environment mapping crucial for autonomous navigation [23]. Multi-sensor fusion has become standard: LiDAR provides accurate 3D structure, while cameras offer rich semantic context; fusion frameworks combine them to improve robustness under adverse weather or lighting conditions. Beyond vehicles, smart infrastructure employs similar perception systems to monitor intersections, detect accidents, and support adaptive traffic management. 3.2 Graph and Sequence Models for Traffic Prediction. Traffic flow forecasting and trajectory prediction are essential for proactive control and congestion mitigation. While recurrent neural networks (RNNs) and long short-term memory (LSTM) networks have been widely applied to capture temporal dynamics, they are limited in modeling the complex spatial relationships in road networks. Graph Neural Networks (GNNs) have emerged as a powerful solution by representing transportation networks as graphs, where nodes correspond to road segments or intersections and edges capture connectivity or vehicle flow [24]. Spatio-temporal GNNs, such as ST-GCN and DCRNN, integrate graph convolution with recurrent structures to model both spatial and temporal dependencies. Transformer-based models are increasingly used for long-range temporal prediction due to their self-attention mechanism [25]. These models enable accurate short-term and long-term traffic forecasting, supporting real-time congestion avoidance, dynamic routing, and incident response. Some works also incorporate external data sources such as weather, events, and social media to enhance predictive accuracy, reflecting the trend toward context-aware ITS analytics. 3.3 Reinforcement Learning for Decision-Making and Control. Reinforcement Learning (RL) provides a natural framework for learning control policies in dynamic and uncertain transportation environments. Single-agent RL approaches, such as Deep Q-Networks (DQN), were initially applied to optimize individual traffic signals. However, traffic networks are inherently multi-agent systems where intersections, vehicles, and roadside units must coordinate. Multi-Agent Reinforcement Learning (MARL) has become a major research trend, enabling distributed agents to learn cooperative strategies that improve global traffic flow while adapting to local observations [26]. Policy gradient methods, actor-critic architectures, and graph-based MARL have been proposed to improve scalability. For autonomous driving, RL enables decision-making in complex scenarios such as merging, overtaking, and lane changing. Hybrid approaches combine RL with rule-based safety constraints or formal verification to ensure compliance with traffic regulations and avoid unsafe exploration [27]. Cooperative Adaptive Cruise Control (CACC) and platooning strategies based on RL further enhance highway throughput and energy efficiency. 3.4 Federated Learning and Privacy-Preserving AI. Data privacy is a critical barrier for AI deployment in transportation, as vehicle and user data often contain sensitive location and behavioral information. Federated Learning (FL) has emerged as a promising paradigm by allowing models to be trained across distributed edge devices — such as vehicles or roadside units — without sharing raw data [28]. Aggregation algorithms such as FedAvg enable the construction of global models while keeping local data private. In vehicular networks, FL can be integrated with V2X communication, enabling collaborative model updates while coping with intermittent connectivity and heterogeneous hardware. To address adversarial participants and poisoning attacks, secure aggregation, Byzantine-resilient optimization, and blockchainbased auditing mechanisms have been proposed [29]. Combining FL with differential privacy and homomorphic encryption further enhances confidentiality, enabling compliance with regulations such as GDPR and CCPA while maintaining model performance. 3.5 Edge Computing and Real-Time AI Deployment. ITS applications require ultra-low latency and high reliability, making centralized cloud-only architectures insufficient. Edge computing brings AI computation closer to data sources by deploying models on vehicles, roadside units (RSUs), and base stations. This reduces communication delay and enables real-time decision-making critical for safety applications such as collision avoidance and emergency braking [30]. Adaptive model compression techniques — including pruning, quantization, and knowledge distillation — allow deep models to run efficiently on resource-constrained edge devices. Hierarchical edge-cloud architectures distribute tasks: the edge handles time-critical inference, while the cloud performs large-scale model training and long-term planning. With the advent of 5G/6G, high-bandwidth and low-latency networks further enhance the viability of edge AI for vehicular systems. Emerging split learning approaches partition deep networks between edge and cloud to balance performance and resource usage. 3.6 Digital Twins and System-Level Intelligence. Digital twin technology is revolutionizing transportation planning by creating high-fidelity virtual replicas of physical mobility systems. Powered by real-time data streams from IoT sensors, connected vehicles, and infrastructure, digital twins enable simulation, prediction, and optimization of traffic scenarios before deployment in the real world [31]. AI enhances digital twins by providing predictive models for congestion, safety incidents, and infrastructure degradation. Reinforcement learning agents trained within digital twins can test control strategies safely and transfer them to real-world systems. This closed-loop integration accelerates the deployment of adaptive traffic management and autonomous driving policies. Furthermore, digital twins support scenariobased testing and validation for safety-critical AI, which is essential for regulatory approval and public trust. By integrating these methodologies — from perception and prediction to control and system-level intelligence — AIdriven ITS is evolving into a highly adaptive, data-centric mobility ecosystem. However, deploying these technologies in real-world large-scale networks introduces substantial technical and socio-economic challenges, which are discussed next. 4. Challenges and Open Research Directions Although Artificial Intelligence (AI) has enabled major advances in Intelligent Transportation Systems (ITS), largescale deployment remains hindered by a complex set of technical, safety, regulatory, and societal challenges. These obstacles arise from the dynamic and safety-critical nature of transportation networks, the heterogeneity of data sources, the computational constraints of edge devices, and the lack of standardized governance for AI-driven mobility. Addressing these barriers is essential to realize the full promise of autonomous and intelligent transportation. 4.1 Data Quality, Heterogeneity, and Bias. Modern ITS depends on massive, diverse data streams collected from vehicles, roadside sensors, mobile devices, and infrastructure cameras. These data are often noisy, incomplete, and inconsistent due to sensor failures, occlusions, adverse weather, and uneven geographical coverage [32]. Learning robust AI models under such conditions remains challenging, particularly when domain shifts occur between training environments and deployment locations. For instance, a perception model trained in one city may fail in another with different traffic patterns or weather conditions. Data bias can lead to unfair or unsafe decisions, such as misclassifying pedestrians of certain demographics or failing to detect bicycles in underrepresented regions. Future research must explore domain adaptation, self-supervised learning, and continual learning to create models that generalize across diverse environments while maintaining fairness and safety. 4.2 Safety, Explainability, and Certification. AI-driven ITS, especially autonomous vehicles, operates in safety-critical contexts where failures can cause accidents and fatalities. Deep neural networks often function as black boxes, making it difficult for engineers and regulators to understand their decision-making process. This opacity hinders safety certification and public acceptance [33]. Explainable AI (XAI) methods, such as attention visualization, counterfactual reasoning, and causal analysis, are emerging but remain immature for real-time control. Moreover, regulators require formal guarantees of safety under all operational conditions, yet verifying complex neural policies remains unsolved. Future systems may combine formal methods with learning-based models, such as integrating rule-based safety constraints or using runtime monitors to detect and override unsafe behaviors. Scenario-based simulation and digital twin validation frameworks will also be critical for large-scale testing of AIdriven vehicles before deployment. 4.3 Cybersecurity and Adversarial Robustness. Transportation systems are vulnerable to a wide range of cyberattacks. Adversarial examples can fool perception models by introducing subtle perturbations to sensor inputs, such as slightly altered stop signs that cause misclassification [34]. Sensor spoofing and V2X message injection attacks can mislead vehicles and traffic controllers, creating dangerous situations. Federated learning and connected vehicle networks are susceptible to model poisoning, where malicious participants manipulate model updates to degrade global performance. Addressing these threats requires robust AI training methods, adversarial detection mechanisms, secure communication protocols, and hardware-based trust anchors. Blockchain-enabled audit trails and secure multi-party computation could support trustworthy model aggregation, but scalability and latency remain concerns. Future work should integrate robust learning, cryptographic protection, and runtime intrusion detection to safeguard AI-driven ITS. 4.4 Scalability, Real-Time Performance, and Edge Deployment. AI models for perception and control are increasingly large and computationally intensive, conflicting with the latency and resource constraints of edge devices such as roadside units and vehicle control units. Real-time collision avoidance and traffic control require inference delays under 50 milliseconds, yet many state-of-the-art deep learning models exceed this budget when running on embedded platforms [35]. Research on model compression, quantization, neural architecture search (NAS), and hardware-software co-design is crucial to achieve low-latency inference. Hierarchical edgecloud collaboration is a promising paradigm, where safetycritical tasks run on local devices while less urgent or computationally heavy processing is offloaded to the cloud. However, optimizing this division dynamically under varying network conditions remains a challenge. Additionally, managing large fleets of heterogeneous vehicles and roadside devices calls for scalable orchestration and remote model updates without disrupting safety-critical operations. 4.5 Privacy, Regulation, and Federated Intelligence. Transportation data often contain sensitive information such as driver identities, routes, and behavior patterns. Centralized AI training conflicts with regulations like GDPR and CCPA that restrict the sharing of personal or location data [36]. Federated Learning (FL) and other privacy-preserving techniques (e.g., differential privacy, homomorphic encryption) offer solutions but introduce new challenges: communication overhead, vulnerability to poisoning attacks, and difficulties in aggregating models under intermittent connectivity. Regulatory frameworks for AI in ITS are still evolving and vary across regions, complicating deployment for global automotive manufacturers. Future research should pursue regulation-aware AI systems that integrate privacy guarantees, auditable decision logs, and compliance with regional data governance laws while maintaining model performance. Standardized protocols for secure FL aggregation and vehicular communication will be critical. 4.6 Interoperability and System Integration. Modern transportation systems are inherently heterogeneous, spanning multiple vehicle manufacturers, infrastructure vendors, and software platforms. AI solutions developed in isolation often struggle to interoperate in mixed environments. The lack of common data standards and model deployment interfaces leads to fragmentation, limiting scalability and innovation. Cross-vendor communication protocols and standardized digital twin interfaces are needed to create a unified mobility ecosystem. In connected vehicle networks, ensuring secure and verifiable interoperation between different vehicle brands and infrastructure is critical. The development of open-source reference platforms, similar to AUTOSAR for automotive electronics, could accelerate interoperability for AI models and datasets. 4.7 Sustainability and Energy Efficiency. Training and deploying large-scale AI models consume substantial energy, conflicting with global sustainability goals. Autonomous vehicle fleets and smart infrastructure require continuous model updates, which can be carbon-intensive if conducted in centralized data centers. PoS-based blockchain for secure federated learning aggregation offers energy savings compared to Proof of Work, but consensus overhead may still be significant in vehicular contexts. Future ITS must integrate green AI strategies, such as energy-aware model adaptation, lightweight learning algorithms, and renewable-powered edge infrastructure. Life-cycle assessments of AI-enabled ITS should be conducted to guide sustainable design and policy decisions. 4.8 Emerging Opportunities: Digital Twins, 6G, and AIAugmented Governance. Looking ahead, several technological trends offer pathways to overcome current barriers. Digital twins provide a safe environment for training and validating AI models under diverse scenarios, enabling robust policy learning and certification. With the advent of 6G networks, ultra-reliable low-latency communication will allow real-time coordination between vehicles, infrastructure, and cloud AI. Furthermore, AI-driven mobility governance systems may automate policy enforcement, congestion pricing, and infrastructure planning by analyzing city-wide mobility data. However, these advances will require rigorous safeguards for accountability, fairness, and explainability to maintain public trust and regulatory compliance [37]. Collaborative research among AI scientists, transportation engineers, policymakers, and ethicists will be essential to balance innovation with safety and societal acceptance. In summary, future ITS will need holistic AI frameworks that are robust, interpretable, privacy-preserving, and energyefficient while operating in highly dynamic, safety-critical environments. The convergence of AI with federated learning, edge-cloud architectures, blockchain security, and digital twinbased simulation is likely to define the next decade of research. Achieving this vision will require both technical breakthroughs and coordinated efforts to develop standards and governance models suitable for global deployment. 5. Conclusion Artificial Intelligence (AI) is rapidly redefining the landscape of Intelligent Transportation Systems (ITS) by transforming how vehicles, infrastructure, and city-scale mobility networks perceive their environment, predict future states, and make adaptive decisions. From early rule-based algorithms to modern deep learning and reinforcement learning, AI technologies have progressed to support complex tasks such as object detection, trajectory prediction, dynamic signal control, and cooperative driving among connected and autonomous vehicles (CAVs). Meanwhile, the rise of federated learning, edge-cloud collaborative architectures, and digital twins has extended AI from vehicle-level autonomy to systemwide intelligence capable of managing entire transportation ecosystems in real time. Our survey integrates these diverse research threads into a unified perspective, covering methodologies from perception and prediction to control and infrastructure intelligence, and examining how AI innovations are reshaping urban mobility toward safer, more efficient, and sustainable futures. A key insight from this survey is that the methodological diversity of AI in ITS is both a strength and a source of integration challenges. On one hand, deep neural networks and graph-based models have revolutionized perception and prediction, while deep reinforcement learning and multi-agent coordination enable complex decision-making under uncertainty. On the other hand, these models often operate as isolated silos, optimized for individual subsystems such as traffic lights or vehicle control, without system-level coordination or standardized interfaces. To move toward truly adaptive and interoperable mobility ecosystems, future research must emphasize holistic design principles that connect perception, prediction, and control into unified, explainable, and verifiable frameworks. Another important conclusion is the urgent need for robustness, safety, and accountability in AI-driven transportation. Unlike other AI domains, ITS operates in highly safety-critical contexts where unexpected model failures or adversarial attacks can cause life-threatening consequences. As we discussed, emerging solutions include explainable AI (XAI), runtime safety monitors, adversarially robust training, and formal verification of neural policies. However, these techniques are still early in their maturity and must be scaled to the complexity of real-world transportation networks. Regulators, industry, and academia must collaborate to create standards for AI safety certification that balance innovation with public trust. Scalability and deployment efficiency remain another decisive frontier. ITS increasingly depends on real-time AI inference under strict latency and energy constraints, often running on resource-limited edge devices such as roadside units and embedded vehicle controllers. Future systems will need to combine model compression, quantization, and adaptive resource allocation with next-generation communication infrastructures like 5G and 6G to achieve real-time performance without sacrificing accuracy. Hierarchical edgecloud orchestration, in which safety-critical inference runs locally while complex training and policy optimization occur in the cloud or digital twins, represents a promising architecture but requires further research on dynamic task allocation and secure model updating. Privacy and data governance will shape the trajectory of AI in ITS over the coming decade. As mobility systems collect sensitive user and vehicle data, compliance with regulations such as GDPR and CCPA becomes essential. Federated learning (FL) and privacy-preserving analytics allow distributed intelligence without centralizing raw data, but these techniques introduce new challenges in communication overhead, adversarial robustness, and model aggregation under intermittent connectivity. We anticipate the rise of privacy-bydesign ITS architectures that integrate secure multi-party computation, differential privacy, and blockchain-based auditing to create transparent yet privacy-preserving mobility platforms trusted by both users and regulators. Finally, the future of AI-enabled transportation will be shaped by emerging paradigms such as digital twins, 6G networks, and AI-augmented governance. Digital twins allow safe and scalable testing of AI models under countless simulated scenarios, accelerating innovation while reducing real-world risk. Ultra-reliable low-latency communication in 6G networks will enable instantaneous coordination between vehicles, infrastructure, and cloud AI. Meanwhile, AI-driven policy engines could automate traffic governance, congestion pricing, and infrastructure optimization, enabling cities to transition from reactive to proactive mobility management. However, these opportunities come with challenges in ethics, fairness, accountability, and sustainability, requiring interdisciplinary collaboration among AI researchers, transportation engineers, policymakers, and social scientists. In conclusion, realizing the vision of safe, scalable, and sustainable AI-powered transportation systems requires more than isolated algorithmic advances; it demands an integrated, systems-level approach that blends cutting-edge AI methods with rigorous safety engineering, privacy protection, and governance. The field is moving from proof-of-concept demonstrations to mission-critical deployment, making it essential to develop robust AI models that are explainable, auditable, and interoperable across heterogeneous transportation infrastructures. We expect that the convergence of deep learning, reinforcement learning, federated and edge intelligence, privacy-preserving technologies, and digital twinbased validation will define the next decade of ITS research and deployment, ultimately enabling cities and nations to build intelligent, adaptive, and trustworthy transportation ecosystems. References [1] S. Chen, Z. He, and Y. Tan, “Artificial intelligence in intelligent transportation systems: A review,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 7, pp. 5775–5796, Jul. 2022. [2] M. Wang, H. Li, and Q. Wu, “Deep learning for traffic flow prediction: A survey,”IEEE Trans. Intell. Transp. Syst., vol. 24, no. 1, pp. 205–229, Jan. 2023. [3] J. Zhang, F. Li, and H. Wang, “Connected vehicles and adaptive traffic management: A machine learning perspective, ”IEEE Intell. Transp. Syst. Mag., vol. 14, no. 4, pp. 35–49, Winter 2022. [4] C. Chen, A. Seff, A. Kornhauser, and J. Xiao, “DeepDriving: Learning affordance for direct perception in autonomous driving, ”in Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit. (CVPR), 2015, pp. 2722–2730. [5] L. Wei, C. Liu, and D. Lin, “Deep reinforcement learning for adaptive traffic signal control: A review,”IEEE Access, vol. 9, pp. 108459–108478, 2021. [6] R. Kohli and Y. Pan, “Data heterogeneity in transportation AI models: Issues and solutions, ”ACM Trans. Intell. Syst. Technol., vol. 13, no. 2, pp. 1–24, Feb. 2022. [7] X. Huang, D. Kroening, W. Ruan, et al., “Adversarial attacks on deep learning models in autonomous driving: A review,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 8, pp. 6246–6260, Aug. 2022. [8] Q. Yang, Y. Liu, T. Chen, and Y. Tong, “Federated machine learning: Concept and applications,”ACM Trans. Intell. Syst. Technol., vol. 10, no. 2, pp. 1–19, Mar. 2019. [9] Y. Lv, Y. Duan, W. Kang, Z. Li, and F.-Y. Wang, “Traffic flow prediction with big data: A deep learning approach, ”IEEE Trans. Intell. Transp. Syst., vol. 16, no. 2, pp. 865 –873, Apr. 2015. [10] J. Zhang, F.-Y. Wang, K. Wang, W.-H. Lin, X. Xu, and C. Chen, “Data-driven intelligent transportation systems: A survey, ” IEEE Trans. Intell. Transp. Syst., vol. 12, no. 4, pp. 1624–1639, Dec. 2011. [11] M. Khamis and W. Gomaa, “Adaptive traffic signal control: A reinforcement learning review, ”IEEE Trans. Intell. Transp. Syst., vol. 21, no. 6, pp. 2314–2333, Jun. 2020. [12] S. Feng, X. Yan, L. Yao, et al., “Deep reinforcement learning for intelligent transportation systems: A survey,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 2, pp. 586–606, Feb. 2022. [13] S. Grigorescu, B. Trasnea, T. Cocias, and G. Macesanu, “A survey of deep learning techniques for autonomous driving,”J. Field Robot., vol. 37, no. 3, pp. 362–386, Apr. 2020. [14] J. Kuutti, R. Bowden, Y. Jin, et al., “A survey of deep reinforcement learning for motion planning of autonomous vehicles,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 2, pp. 540–558, Feb. 2022. [15] S. Feng, D. Yan, and H. Sun, “Multi-sensor fusion for autonomous vehicles: A review, ”IEEE Trans. Intell. Transp. Syst., vol. 24, no. 3, pp. 2376–2396, Mar. 2023. [16] J. Molina-Masegosa and J. Gozalvez, “LTE-V and 5G NR V2X communications: An overview and research directions, ”Veh. Commun., vol. 20, p. 100185, Apr. 2020. [17] H. Habibzadeh, T. Soyata, B. Kantarci, A. Boukerche, and C. Kaptan, “Large-scale urban ITS data analytics: Challenges and opportunities,”IEEE Commun. Surveys Tuts., vol. 23, no. 3, pp. 1565–1596, 2021. [18] K. Zhang, S. Leng, X. Peng, and Y. He, “Cloud-based intelligent transportation systems: Architecture and applications,”IEEE Commun. Mag., vol. 55, no. 10, pp. 72–79, Oct. 2017. [19] H. Zhou, X. Li, and Y. Zhang, “Deep reinforcement learning for urban traffic control: A survey, ”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 7, pp. 5834–5851, Jul. 2022. [20] H. Min, J. Hu, F. Tao, and A. Liu, “Artificial intelligencepowered digital twins in transportation: A survey,”IEEE Trans. Intell. Transp. Syst., vol. 24, no. 9, pp. 9120–9138, Sep. 2023. [21] R. Xu, H. Wang, Z. He, and Y. Wang, “Federated learning in vehicular edge networks: A survey, ”IEEE Commun. Surveys Tuts., vol. 23, no. 4, pp. 2761–2787, 2021. [22] J. Redmon and A. Farhadi, “YOLOv3: An incremental improvement,”arXiv:1804.02767, 2018. [23] N. Carion et al., “End-to-end object detection with transformers,”in Proc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 213–229. [24] Y. Li, R. Yu, C. Shahabi, and Y. Liu, “Diffusion convolutional recurrent neural network: Data-driven traffic forecasting, ”in Proc. Int. Conf. Learn. Represent. (ICLR), 2018. [25] A. Vaswani et al., “Attention is all you need,”in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017, pp. 5998–6008. [26] C. Chu, Y. Qin, and Y. Wang, “Multi-agent deep reinforcement learning for large-scale traffic signal control, ”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 7, pp. 6468–6480, Jul. 2022. [27] J. Kuutti, S. Fallah, and R. Bowden, “Hybrid motion planning with deep reinforcement learning for autonomous driving, ” IEEE Trans. Intell. Transp. Syst., vol. 24, no. 5, pp. 4423–4438, May 2023. [28] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, “Federated learning: Challenges, methods, and future directions, ”IEEE Signal Process. Mag., vol. 37, no. 3, pp. 50–60, May 2020. [29] Q. Yang, Y. Liu, T. Chen, and Y. Tong, “Federated learning in mobile edge networks: A comprehensive survey, ”IEEE Commun. Surveys Tuts., vol. 23, no. 3, pp. 1565–1587, 2021. [30] X. Chen, Z. Zhang, C. Wu, and S. Mao, “Machine learning for wireless networks with artificial intelligence: A tutorial on neural networks,”IEEE Commun. Surveys Tuts., vol. 23, no. 3, pp. 1573–1599, 2021. [31] F. Tao, Q. Qi, A. Liu, and A. Kusiak, “Digital twin-driven smart manufacturing: Connotation, reference model, applications and research issues,”Robotics Comput.-Integr. Manuf., vol. 61, p. 101837, Apr. 2020. [32] X. Ma, Z. Tao, Y. Wang, H. Yu, and Y. Wang, “Long shortterm memory neural network for traffic speed prediction using remote microwave sensor data, ”Transp. Res. C, vol. 54, pp. 187–197, May 2015. [33] W. Samek, T. Wiegand, and K.-R. Müller, “Explainable artificial intelligence: Understanding, visualizing and interpreting deep learning models,”IT Prof., vol. 21, no. 3, pp. 82–88, 2019. [34] X. Huang, M. Kwiatkowska, S. Wang, and M. Wu, “Safety and trustworthiness of deep neural networks in autonomous driving: A review,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 8, pp. 6246–6260, Aug. 2022. [35] H. Wu, X. Chen, and S. Mao, “Edge intelligence for autonomous driving: Opportunities and challenges, ”IEEE Netw., vol. 35, no. 3, pp. 162–169, May–Jun. 2021. [36] R. Shokri and V. Shmatikov, “Privacy-preserving deep learning, ”in Proc. ACM Conf. Comput. Commun. Secur. (CCS), 2015, pp. 1310–1321. [37] Z. Zhou, X. Chen, E. Li, L. Zeng, and J. Xu, “Edge intelligence: Paving the last mile of artificial intelligence with edge computing,”Proc. IEEE, vol. 107, no. 8, pp. 1738–1762, Aug. 2019.