Hazard detection is an important problem for Intelligent Transportation Systems (ITS), although, developing and deployment of such systems is a complicated task, given human presence in the loop and a large account of non-controlled variables (e.g. traffic levels, weather conditions, etc). In addition, a distributed hazard detection system will require a communication network for a large number of sensors, making this network a crucial element of the system and a potential bottleneck for its performance. Therefore, a suitable platform is needed for the development and validation of vehicular applications that handle both the vehicular and network aspects of the system to seamlessly migrate an application to an urban environment. This paper addresses these gaps by designing a hardware-in-the-loop architecture for a realistic ITS hazard detection system. The platform relies on CARLA simulator to provide a safe, controlled, and repeatable environment to test the application, and a software-defined Open RAN network that allows studying the effect of communications nuisances on the service and the possibility to implement network aware services to enhance performance. In addition, a hazard detection service is devised using the described architecture. Our approach uses a combination of machine learning detections with Kalman (or particle) filters to aggregate data from multiple cameras. We assume detections are asynchronous and data transmission should be minimized to ensure scalability. A data association criterion based on Mahalonabis distance is proposed to automatically associate filters with trajectories. Our experiments provide insights into the throughput and latency of the network and their impact on the ITS service. The results show the service is robust to end-to-end latency, making comparisons among KFs, PFs and unscented KFs.

The EU-funded Horizon 2020 5GMED project aims to promote Cooperative Intelligent Transportation Systems (CITS) in Europe by deploying Cooperative, Connected, and Automated Mobility (CCAM) services across international borders. In this context, this paper presents a 5G-based road traffic safety & management service based on the development of multiple actors and components capable of assessing the traffic status and providing the proper information to Connected Vehicles (CVs) through 5G and Vehicle-to-Everything (V2X) communication technologies. Performance evaluation has been carried out through real-world experiments and trials conducted along the 5GMED cross-border corridor (CBC) between Spain and France, leading to interesting findings and lessons learned presented in the conclusion of this paper.

As autonomous driving (AD) systems grow more complex, their rising computational demands pose significant energy and sustainability challenges. This paper investigates spiking neural networks (SNNs) as low‐power alternatives to convolutional neural networks (CNNs) for regression tasks in AD. We introduce a membrane‐potential ( V mem ) decoding framework that converts binary spike trains into continuous outputs and propose the energy‐to‐error ratio (EER), a unified metric combining prediction error with energy consumption. Three CNN architectures (PilotNet, LaksNet, and MiniNet) and their corresponding SNN variants are trained and evaluated using delta, latency, and rate encoding across varied parameter settings, with energy use and emissions logged. Delta‐encoded SNNs achieve the highest EER, substantial energy savings with minimal performance loss, whereas CNNs, despite slightly better MSE, incur 10–20 × higher energy costs. Rate encoding underperforms, and latency encoding, though improving relative error, demands excessive energy. Parameter tuning (threshold θ , temporal dynamics ( S ), membrane time constant ( τ ), and gain G ) directly influences eco‐efficiency. All experiments run on standard GPUs, showing SNNs can surpass CNNs in eco‐efficiency without specialized hardware. Paired statistical tests confirm that only delta‐encoded SNNs achieve significant EER improvements. This work presents a practical, energy‐aware evaluation framework for neural architectures, establishing EER as a critical metric for sustainable machine learning in intelligent transport and beyond.

The rapid progress of autonomous driving (AD) has triggered a surge in demand for safer and more efficient autonomous vehicles, owing to the intricacy of modern urban environments. Traditional approaches to autonomous driving have heavily relied on conventional machine learning methodologies, particularly convolutional neural networks (CNNs) and recurrent neural networks (RNNs), for tasks such as perception, decision-making, and control. Presently, major companies such as Tesla, Waymo, Uber, and Volkswagen Group (VW) leverage neural networks for advanced perception and autonomous decision-making. However, concerns have been raised about the escalating computational requirements of training these neural models, primarily in terms of energy consumption and environmental impact. In the situation of optimisation and sustainability, Spiking Neural Networks (SNNs), inspired by the temporal processing of the human brain, have come forth as a third-generation of neural networks, famed for their energy efficiency, potential for handling real-time driving scenarios and processing temporal information efficiently. However, SNNs have not yet achieved the performance levels of their predecessors in critical AD tasks, partly due to the intricate dynamics of neurons, their non-differentiable spike operations, and the lack of specialised benchmark workloads and datasets, among others. This paper examines the principles, models, learning rules, and recent advancements of SNNs in the AD domain. Neuromorphic hardware, hand in hand with SNNs, shows potential but has challenges in accessibility, cost, integration, and scalability. This examination aims to bridge gaps by providing a comprehensive understanding of SNNs in the AD field. It emphasises the role of SNNs in shaping the future of AD while considering optimisation and sustainability.

Autonomous driving and the machine learning (ML) models developed to achieve it have grown rapidly in importance and complexity respectively, and with them has also grown their carbon footprint due to long training times. Given the importance of climate change, it should be necessary to include the CO2 impact of ML models explicitly to encourage competition on more than just model quality. This work presents the implementation of two different Convolutional Neural Network (CNN) training approaches, used for autonomous driving by behavioral cloning in a simulated environment and compares their impact on the CO2 footprint. Using a cloud execution environment and driving data, previously obtained by applying an end-to-end deep learning technique, the first implemented training approach is a classical approach that uses an image generator that carries out the pre-processing and data augmentation during training. The contribution proposed in this paper is a second approach that improves the training by decreasing the training time, carrying out the data augmentation and pre-processing tasks before training the model, storing the result in RAM and then starting the training. The new approach to training presented in this article finishes the training approximately 38 times faster and reduces the carbon footprint impact by approximately 96%. In absolute values, this is a reduction from an average value of approximately 0.1643 CO2eq (kg) to 0.007 CO2eq (kg). To estimate the impact of CO2, the hardware used for the project, the training time and the cloud service provider were all taken into account.

In order to improve road safety, modern vehicles are equipped with smart sensors and Vehicle-to-Everything (V2X) communication technologies that facilitate the exchange of data (e.g., location, speed, road hazards) with other vehicles, the road infrastructure, and pedestrians, thus extending the range of perception beyond the capabilities of on-board sensors. All these data can be processed by a Collision Avoidance service deployed in a mobile edge computing (MEC) platform to guarantee low latency in the detection and localization of road hazards. In this paper, we propose a Collision Avoidance service based on Vanetza, an open-source ETSI ITS protocol stack, and demonstrate its operation using already developed experimental On-Board Units (OBUs) that communicate with the Collision Avoidance service over UDP and MQTT to exchange ETSI ITS messages encoded in ASN.1 and JSON format, respectively. We measure the application-level latency and we observe the benefit of our proposed approach in terms of latency reduction by a factor of 12 with respect to the literature.

The fifth generation (5G) of communication networks aims to accelerate the adoption of incipient vertical industries which will leverage innovative smart applications and services such as Cooperative, Connected and Automated Mobility. One of the objectives globally within that area is reducing to zero the number of fatal vehicle accidents. Unfortunately, human errors are the main cause of them, where vulnerable road users (VRUs) are involved in half of the cases. A possible approach to reduce accidents is estimating the probability of collision between two vehicles based on their estimated trajectories. These trajectories are usually tracked on-board using sophisticated devices such as cameras and LiDAR. However, VRUs are generally not equipped with such equipment and, ideally, VRUs carry smartphones with active geolocation capabilities based on satellite-based positioning systems. In this paper, we propose a novel vehicular service based on a regression algorithm to predict trajectories by uniquely using Cartesian coordinates. We compare different types of regression techniques in terms of prediction time window, position accuracy and processing time using Weka. Results show that the Alternating Model Tree (AMT) technique can predict the next position with an error of less than 3.2 centimeters, increasing up to 1 meter when predicting the next 5 positions with a period of 1 second between consecutive positions. In this case, a prediction time window of 5 s is processed within 1.25 milliseconds. AMT resulted as the lowest complex and most accurate algorithm in a multiple-step prediction position.

The Internet of Things (IoT) is being used to monitor a wide range of physical phenomena. In this paper, we are concerned with the extraction of features from the gathered IoT signals and, specifically, with the online estimation of their rate-distortion relationship. This information is in fact key to the configuration and adaptation of data compression and in-network processing protocols and, in turn, is deemed a prime functionality for IoT networks. The point is that lossy compression can be often applied at the sources to save transmission energy, while meeting application requirements in the reconstruction quality. This task is however signal- and time-dependent as different signals are usually characterized by different relations and the signal statistics may also change as a function of time. Here, we first formulate the rate-distortion estimation task, framing it as a classification problem. Hence, we consider the following clustering algorithms from the literature: multilayer perceptron, support vector machine, random forest and linear discriminant analysis, and use them to automatically assess rate-distortion curves in an online fashion and from a small number of signal samples. These algorithms are compared in terms of classification accuracy, training time and memory footprint. Numerical results reveal that, although the problem is inherently complex, the careful combination of feature extraction and classification tools makes it possible to reach high classification accuracies using only a few signal features (e.g., from one to four). The best algorithm (random forest) also entails a short training time and, if properly tuned, has a modest memory footprint.

Elder adults may have some dependence on performing common activities like zapping on the television through a remote control (i.e. due to possible hand mobility problems). The Internet of Things (IoT), including the Radio Frequency Identification (RFID), interconnects devices to provide a higher variety of services. Together, and by applying intelligence through Machine Learning (ML) techniques, advanced applications can be implemented improving people's life. We present the Smart Surface system, relying on state of the art RFID equipment. It uses the unsupervised machine learning technique K-means clustering to detect and trigger actions by means of simple gestures, in real time and in a non-intrusive way. We implemented and evaluated a prototype of the Smart Surface system achieving an accuracy of 100% gesture recognition.