Carla-Fabiana Chiasserini

Developing an edge-assisted UAV surveillance system that ensures low-latency, high-accuracy execution of computer vision tasks while optimizing energy consumption and network resources remains a complex challenge. In this paper, we address the limitations of existing research by leveraging image semantics, image ensemble processing, and mmWave UAV-edge channel statistics. We do so by focusing on joint optimization of UAV speed, camera images per second (IPS) rate, offloading policy, and transmission rates with the aim to minimize the UAV's energy consumption. Given the NP-hardness of the problem, we propose an algorithmic solution, named Intelligent UAV Network (IntUNe), which is based on an innovative constrained reinforcement learning strategy that dynamically and effectively adjusts to real-time conditions. Our results demonstrate that, in delay-constrained UAV surveillance networks, IntUNe closely matches the optimum in small-scale scenarios, and it increases inference accuracy by significantly reducing violations of the accuracy constraint by up to 96.19% compared to state-of-the-art alternatives. Also, it reduces UAV propulsion energy consumption by 34.67% and total UAV energy consumption by up to 37.3%.

We consider the collaborative inference acceleration task via cloud-edge-end collaboration, which involves a series of tightly coupled decision-making steps, including which DNN model to be selected, how much to compress model, how to partition model, and where to offload partitioned submodels. In practical deployments, these decisions jointly affect both fine-tuning and inference performance, and must jointly account for such aspects as the model being used, the computational resources and local datasets available at each device, as well as network latencies, which significantly increases the complexity of optimizing the problem. Yet, no existing studies focus on such joint optimization problem for these tightly coupled decisions. In this paper, we model this problem as a multi-dimensional optimization problem, jointly optimizing collaborative inference and fine-tuning by selecting the DNN model, compression level, partition strategy, and computational resource allocation, with the objective of minimizing the overall energy consumption of the learning-inference process, subject to accuracy and latency constraints. To this end, we propose an algorithmic framework called JQODI combining a time-energy tree diagram to represent the learning process, a dynamic programming solution strategy, and a data-driven theoretical approach to predict the expected total number of training epochs that meet the accuracy requirements. We prove that JQODI approximates the optimal solution with polynomial complexity. Numerical results demonstrate that JQODI surpasses state-of-the-art methods in both energy efficiency and latency.

The availability of powerful Artificial Intelligence (AI) algorithms boosts the development of advanced functionalities in the automotive domain and is essential to enable the deployment of autonomous and semi-autonomous decision-making vehicles. However, integrating such advanced and complex functionalities in automotive systems is challenging due to several factors, including: (i) the mandatory compliance with strict safety regulations, which require effective strategies to ensure timely development while allowing thorough dependability evaluations, and (ii) the short time-to-market imposing limited development, verification, and validation periods. In particular, the analysis of the effects of faults affecting the hardware executing an AI-based application is made challenging by the target system’s complexity (in terms of both hardware and software). Reliability analysis often resorts to Fault Injection techniques. However, Fault Injection experiments are often unacceptably time-consuming and limited to some components of the overall system, thus failing to consider the fault impact at the vehicle level. This work proposes a new method, named Two-steps IntegrAted Reliability Assessment ( TIARA ), for early estimation of the impact at the vehicle level of faults affecting the hardware running AI-based perception tasks in the automotive domain. TIARA allows for the early exploration and evaluation of algorithms, driving agents, and critical operational scenarios. TIARA can estimate the effects of faults affecting a subsystem up to the vehicle level, integrating a fault injection approach at the neural network level with a commercial automotive-grade virtual scenario generator. When compared to previous works, TIARA dramatically reduces the required computational effort by adopting a two-stage evaluation strategy. It first performs static analysis to determine fault vulnerabilities and identify the most vulnerable parts (code blocks) in a targeted application. Then, it focuses on the most susceptible parts of the neural network and estimates system-level effects on vehicle dynamics by combining the system’s perception, control, and driving features. We validated our methodology through the exhaustive evaluation of two applications: Lane Centering Assistance (LCA) and Emergency Lane Keeping Assistance (ELKA), using the YoloP model for perception. The experimental results show that TIARA allows for an efective early estimation of systems reliability through relevant driving dynamics and comfort metrics on the nine evaluated driving scenarios, as mandated by standards, while reducing computing complexity by up to 43.2X in comparison with a fully exhaustive evaluation approach. In addition, the validation of the TIARA methodology through a hardware-in-the-loop implementation shows that the results closely match the behavior of a real-world system, demonstrating the versatility of the TIARA strategy for the evaluation of automotive systems.

In vehicular communications, accurate modeling of real-world radio propagation channels is essential. To this end, we propose a novel stochastic model, named Vehicular-Tapped Delay-Line (V-TDL) that accurately captures the statistical behavior of multipath channels characterized by path-dependent gains, delays, and Doppler shifts. V-TDL supports diverse traffic conditions and road geometries by generating channel instances through well-established probability distributions. Also, it effectively models the parameters of the distributions through realistic geometry-based simulations, achieving the accuracy of a ray-tracing-based model while maintaining the low complexity of a purely stochastic approach. In contrast to existing models, V-TDL accounts for the correlation between propagation paths. Our findings show that this correlation is inherent in high-speed vehicular environments and neglecting it leads to a significant overestimation of channel diversity and system performance. We compare our model to existing alternatives to assess the performance of OTFS and OFDM modulations. The results demonstrate that, unlike traditional models such as the 3GPP EVA, V-TDL captures variations in channel diversity influenced by traffic intensity and road geometry, which impact the OTFS and OFDM performance. Although OTFS is penalized by path correlation, it consistently outperforms OFDM in all evaluated vehicular environments, confirming its suitability for high-speed vehicular communication scenarios.

Computer vision (CV) tasks are among the most pivotal, yet challenging, operations for uncrewed aerial vehicles (UAVs), especially in mission-critical applications. They require processing complex image data through deep neural networks (DNNs), which demand computational resources far beyond UAVs’ capacity. To address this limitation, Split DNNs offer a promising solution by partitioning the model into: 1) a lightweight Head, deployed on the UAV for rapid, albeit less precise, initial image representations, and 2) a more complex Tail, executed at the network edge for refined, higher accuracy results. However, this solution necessitates transmitting large tensor data from the UAV to the edge server, leading to significant bandwidth consumption. We tackle this challenge by introducing a goal-oriented framework named compressed tensor-based DNN split (CoTeD). Our framework integrates an application- and system-aware optimization model that orchestrates computing and transmission resources in real time. At the UAV, CoTeD dynamically selects relevant tensor information and optimally tradesoff between DNN detection quality and bandwidth consumption, guided by application requirements and system operational conditions. At the edge server, CoTeD reconstructs the tensor, enabling efficient inference by the Tail model. This approach effectively balances bandwidth usage with the quality of the CV task output. Experimental results, obtained through our hardware-software testbed and using datasets with different sizes and characteristics, show that CoTeD can reduce data transmission over the radio link by up to 90% without noticeable loss in object detection quality and inference latency by up to 70% compared to local DNN deployment onboard the UAV. Also, CoTeD yields an inference request success rate of at least 90%, with an increase of 20%–80% compared to direct DNN splitting, static Joint Photographic Experts Group (JPEG) compression, and DNN model quantization.