Edge computing architectures have greatly benefited from high-performance end systems and new-generation cellular networks. Smart cities now have the potential to synthesize all the aspects of a heterogeneous communication and computing infrastructure, leveraging high-definition sensors to propose multiple applications to a variety of road users. However, with the advent of connected vehicles and due to the short lifetime of OBU-to-RSU links, the issue of mobility threatens the reliability of task and data offloading. We thus propose a multi-RAT data offloading scheme exploiting device-to-device communications. In particular, considering the large amounts of data generated by an on-board lidar, we establish a bilateral offloading scheme where the RSU hosts a management algorithm. Based on this algorithm, different packet sizes may be requested based on the networking performance of each vehicle. Through extensive field trials and simulations, performance evaluation shows clear performance discrepancies between C-V2X and DSRC, with C-V2X being more appropriate for dynamically-managed offloading. 4192-byte packets sent at a 10 ms interval achieve the highest throughput with 96 % delivery rates. The proposed algorithm then achieves 30-to-45% increases in throughput when compared to an OBUcentered scheme due to improved resource allocation.

Vehicular connectivity is becoming an integral part of the automotive industry, pushed forth by edge computing and smart cities. However, in urban environments, wireless links are prone to physical disturbances. To maintain their service with these conditions, vehicular networking architectures need the ability to anticipate phenomena and adapt proactively. The multitude of variables involved in this problem calls for deep learning approaches. As such, predictive Quality-of-Service has shown promising contributions to the performance of low-latency infrastructures, but a field study of appropriate recurrent neural networks has yet to be conducted for vehicular networks. Moreover, the inherent heterogeneity of urban networks can be leveraged for service continuity purposes, by developing a radio access technology selection algorithm based on the predictions made for each technology. In that regard, while standalone 5G has been publicized as a key technology for very demanding applications like tele-operated driving, previous works have not studied its contribution in a real-life environment. This work thereby proposes a predictive Quality-of-Service infrastructure with proactive radio access technology selection for urban vehicular networks, evaluated by field trial and involving 5G SA, C-V2X and ITS-G5. Experimental observations show the enhanced contribution of 5G SA, with a minimal latency of 16 milliseconds and higher overall reliability than device-to-device communications. These measurements pave the way to cellular vehicular communications in applications with high service level requirements. Moreover, our system is able to ensure 99.9% success rates from 72% of the vehicle travel time to more than 90% of the time. Average reliability increases from 95.7% to 99.4%.

The future of smart cities depends on the ability of devices to share contextual data in a quick and reliable manner. Deploying access points in close proximity to the side of the road greatly reduces the propagation distances of messages sent over the sidelink, achieving extremely low latencies. However, in challenging situations encountered in urban contexts, multiple obstacles hinder the availability of direct links between a device and its destination. While the usual fallback solution resides in cellular connectivity, the gap in performance is immense. We propose relaying packets sent using Cellular-V2X (C-V2X) sidelink, where the initial message is received by an intermediary with better connectivity to the targeted access point, then retransmitted from this new position. To explore the performance and reliability of such communications in urban V2X architectures, we present experimental campaigns exploiting C-V2X PC5, 5G, autonomous vehicles and state-of-the-art hardware. We study multiple scenarios, extending or shortening line-of-sight situations to analyze coverage and performance. Results analysis highlights high reliability of relayed C-V2X packets at up to 200-meter ranges from one peer to another, with observed latencies being, at worst, 42% more favorable than the cellular fallback.

Internet of Things (IoT) infrastructures can undoubtedly derive significant benefits from content-centric networking (CCN). Thus, a flurry of research focuses on their integration for better content retrieval. Despite the plethora of caching strategies available today, none stands out by efficiently taking into account the limitations of sensors and IoT systems such as low data-rate and energy, as well as limited memory and processing capability. In this work, we focus our attention on the effectiveness of caching on IoT edge gateways. To this end, we propose a novel Deep Reinforcement Learning (DRL) strategy that establishes an intelligent placement of contents and promotes the consideration of content popularity as well as the diversity of data in the network. Through extensive experiments and a real-world testbed on our campus, we perform an in-depth comparative analysis. We demonstrate that our approach decreases the data retrieval distance, assuredly leading to energy efficiency and reduced latency. Moreover, our DRL-based solution outperforms other strategies in terms of cache hit, content diversity, and content popularity support.

To enhance road safety and traffic fluidity, vehicles need to communicate amongst themselves and transfer large amounts of data to the infrastructure with extremely low latencies, through access points deployed closely to the road. However, in obstacle-heavy scenarios, latency increases when devices shift coverage areas. Though networks are heterogeneous by nature in the smart city era, very few mechanisms are provided in connected vehicles for multiple Radio Access Technology (multiRAT) selection. We thus present a networking approach to improve link stability, exploiting mobile and V2X technologies: LTE, 5G, Cellular-V2X, DSRC. Unlike opportunistic handover schemes taking advantage of Device-to-Device (D2D) communications when coverage is detected, we develop a proactive scheme leveraging the performance of received packets, to avoid link failure against pathloss and shadowing. An autonomous shuttle roams the playground, filling grids associating positions with QoS indicators. The grids then provide knowledge on RAT performance to incoming vehicles, facilitating selection according to position. Experimental studies show high performance of DSRC in short range, line-of-sight situations, while C-V2X offers up to 60% longer D2D ranges and more resilient coverage. Field trial analysis of our solution highlights a 27% reduction of high-latency packets, by extending V2I communications and limiting reliance on cellular connectivity.

The next step towards vehicular networks in smart cities would be the deployment of autonomous shuttles with multiple on-board applications. Their need to offload task towards Road Side Units (RSUs) is inevitable, especially with a certain proportion of urgent data needing to be processed in the shortest possible delay. Therefore, QoS-provisioning appears as imperative, along with the optimization of resource requesting at RSUs. To this end, we propose CAVTOMEC, a multi-RAT location-aware, context-aware task offloading solution with QoS provisioning for MEC vehicular networks. Our solution consists of three intertwined mechanisms: traffic classification, location-awareness exploiting the contents of CAM beacons and V2N-enhanced resource polling. Traffic classification identifies high, low, and indifferent task priorities, while location and resource awareness help to select the most appropriate RSU to offload tasks to depending on these priorities. Performance evaluations show that our proposal offers better load balancing at the RSUs than traditional offloading schemes, thus satisfying high priority task offloading at better rates and freeing up more resources in case an unexpected event occurs.