Keywords

1 Introduction

Next-generation wireless networks will facilitate the development of connected automated industrial systems by exploiting disruptive technologies, such as THz frequencies, Reconfigurable Intelligent Surfaces (RIS) as well as Integrated Sensing and Communication (ISAC) systems [15]. Thanks to these new technological developments, distributed computing tools will replace energy-hungry cloud processing functions by pushing the intelligence directly into edge devices or agents [31]. The formation of self-sustained, cooperative networks is beneficial for advanced mobility services as they allow merging (partial) information acquired from spatially-distributed agents and consequently improve the sensing/localization capabilities of the agents themselves. Machine Learning (ML) tools are also paramount in these contexts to extract useful relationships from the data collected by the agents, allowing further positioning/sensing performance enhancements.

Driven by all these key elements, this chapter presents novel cooperative localization and sensing strategies for future mobility systems comprising indoor/outdoor scenarios characterized by complex propagating conditions and/or highly dynamic interactions among the agents. These conditions may arise due to harsh environments in which the agents are deployed (e.g., industrial facilities) or due to the agents’ mobility (e.g., vehicular contexts). To solve such challenges, wireless networks are exploited to enable cooperative schemes where networked devices collaborate in sharing information with the goal of estimating their position, perceiving the surrounding environment, or both. Besides, data-driven approaches are tightly integrated into the proposed algorithms to enable efficient and trustworthy sensing/positioning functionalities.

The chapter initially addresses the problem of high-precision localization and environmental perception as two separate tasks. Individual solutions are proposed for both tasks that aim at augmenting the positioning/sensing performance by exploiting side information from the surrounding environment. Next, a combined approach is proposed where the localization of the mobile agents is integrated with the perception of the environment at each agent by means of a cooperative approach. Finally, careful attention is given so that the proposed methods provide not only accurate positioning/sensing functionalities but also trustworthy responses. This aspect is treated in the last part of the chapter where a trustworthy, yet accurate perception system is developed.

The rest of the chapter presents in more detail the main research activities included in my Ph.D. thesis [3] and in the publications [2, 4,5,6,7,8,9,10,11, 32]. Further results of the research conducted during the Ph.D. can be found in [13, 14, 20, 21, 29, 30], which are omitted here as not fitting the main scope of the Ph.D. thesis. The organization of the chapter is as follows. Sect. 2 addresses the problem of high-accuracy localization in complex propagating environments, while Sect. 3 focuses on providing a solution for accurate environmental perception integrating distributed learning tools. Then, localization and sensing tasks are combined via a unified cooperative approach in Sect. 4, whereas Sect. 5 studies how to enhance not only the accuracy but also the trustworthiness of perception systems. Finally, Sect. 6 draws some conclusions.

2 Localization of Mobile Agents in Complex Environments

Accurate location information has become a fundamental requirement in many of today’s services. Positioning estimates can be typically obtained by harnessing the radio signals exchanged over a wireless network. However, complex environments, such as industrial plants, pose a major problem in this respect, as they heavily affect the quality of the location-dependent information that can be extracted from wireless signals. Therefore, such environments call for augmentation/mitigation strategies able to amend Non Line of Sight (NLOS)-corrupted radio signals and profitably use them to enhance localization accuracy.

Throughout the years, several approaches have been developed to mitigate the impact of complex propagating conditions and improve positioning performance. Statistical characterization of the Channel Impulse Response (CIR) can be used to detect NLOS propagation and consequently correct localization measurements [19]. Another popular approach is to rely on Bayesian tracking filters to integrate information on the propagation environment [27]. More recently, ML methods have been also proposed to provide accurate localization in harsh environments [26]. They mostly rely on supervised techniques to augment the localization accuracy under NLOS propagation.

2.1 A Bayesian Tracking Framework for NLOS Compensation

To address the aforementioned challenges, the Ph.D. thesis proposes a novel NLOS mitigation approach that incorporates multiple (hybrid) localization measurements, namely Time Difference of Arrivals (TDoAs) and Angle of Arrivals (AoAs), as well as an efficient tracking algorithm to estimate the position of the agents accurately. The technique is specifically formulated for Ultra WideBand (UWB) systems. Still, it is general enough to be applied to any wide bandwidth and multi-antenna system, such as the ones foreseen for Beyond 5G networks.

More specifically, the method aims at embedding the propagation information of the environment in which the localization task has to be carried out while mitigating the NLOS and multipath impairments. It does so by jointly tracking the agents’ position and the Line of Sight (LOS)/NLOS conditions, referred to as sight conditions, experienced at the reference stations or Access Points (APs). The sight conditions evolution is modeled as a first-order Markov chain with transition probabilities describing the change of state from LOS to NLOS and from LOS to LOS, and calibrated according to the available layout information of the tracking area. Based on the current estimated value of the APs sight variables, the measurements are statistically described to take into account the actual propagating conditions and compensate for the measurements affected by NLOS. The developed statistical framework is integrated with a Jump Markov System (JMS) that enables the description of the relationship between sight conditions and the position of the agents, allowing the overall problem to be solved via a Bayesian filtering approach. In this respect, a Particle Filter (PF) implementation is considered to track the joint position-sight state efficiently across time. The overall methodology is summarized in Fig. 1.

Fig. 1
An illustration of the Bayesian N L O S mitigation methodology. It includes motion models, measurement models, sight evolution, angle of arrival, and then the Bayesian tracking through particle filtering.

Bayesian NLOS mitigation methodology: the aim is to jointly track the agents’ positions as well as the visibility conditions of the APs so that highly corrupted localization measurements can be compensated and used for improving the positioning performances

Full size image

The proposed NLOS compensation tool has been evaluated considering real raw UWB data collected inside a fully-functional industrial facility and compared against a conventional Bayesian filter that does not compensate for the NLOS as well as state-of-the-art NLOS mitigation methods [5]. Experimental results showed the superiority of the proposal in providing more accurate localization compared to all other methods considered, especially in areas highly affected by non-ideal propagating conditions. Besides, the proper integration of hybrid positioning measurements is beneficial for further improving the position estimate of the agents. We refer the interested reader to [2, 5] for more details about the methodology and additional analyses.

3 Federated Learning for Enhanced Perception

Future mobility systems will require highly-accurate localization and environmental awareness capabilities to detect possible hazardous situations and act accordingly. This section moves in this direction and complements the previous one by proposing a sensing system for accurate environmental perception in high-mobility scenarios, i.e., road vehicles. A wireless network connecting the vehicles is exploited to implement cooperative perception strategies, where a set of networked vehicles equipped with imaging sensors aim at obtaining enhanced perception capabilities.

Conventional cooperative sensing methods rely on data-sharing procedures where raw or partially processed data are exchanged over the network [28]. However, the introduction of regulations restricting the access and distribution of data among multiple parties makes such techniques unfeasible. On the other hand, Federated Learning (FL) procedures can be used to learn a ML model able to provide the same sensing functionalities as standard cooperative perception approaches. FL [32] resorts to the exchange of locally trained instances of a shared ML model without requiring any data exchange. Even though FL represents a promising privacy-preserving solution, communication-efficient designs are required to make FL platforms more sustainable, especially when large models need to be exchanged over the network.

3.1 Communication-Efficient FL Policy

This section discusses how to improve the communication efficiency of decentralized learning policies so as to obtain more sustainable FL-based perception systems without penalizing sensing performances. A communication-efficient design is introduced where the vehicles participating in the FL process are able to intelligently select a subset of the parameters of the ML model to be exchanged via Vehicle-to-Everything (V2X) networking.

As depicted in Fig. 2, the proposed communication-efficient strategy tries to reduce the communication overhead by choosing the layers of the Neural Network (NN) according to the local data quality observed at the vehicles. We develop a layer selection optimizer that dynamically selects the layer parameters according to the normalized squared gradients observed during the local optimization step performed by the vehicles. The gradients are firstly sorted in a descending manner and only the layers associated with the strongest gradient magnitudes are selected and propagated to the neighbors. Intuitively, higher magnitude gradients convey more informative updates; therefore, the corresponding layers should be transmitted more frequently. Additionally, a randomized policy is integrated with the layer selection optimizer that chooses the layers in a independent and identically distributed (i.i.d.) fashion. Besides providing a more fair layer exchange process during the FL process, the combination of gradient-based and randomized selection strategies has been found to provide higher-quality models with improved generalization abilities [9].

Fig. 2
A diagram illustrates the communication-efficient F L policy. It has initialization, local model optimization, model exchange through V Z X network, consensus-based federated averaging, and convergence.

Communication-efficient decentralized FL policy: vehicles exchange fractions of their local models and implement an average consensus policy for fusing the model updates received from their neighbors

Full size image

The performances of the communication-efficient design have been evaluated considering a challenging automotive vertical, where vehicles are required to optimize a large NN for accurately classifying road users/objects present in the driving environment via lidar point clouds [6]. The assessment was focused on characterizing the impact of the layer selection process on the final accuracy of the trained models while also comparing the achieved results against conventional centralized and decentralized learning strategies.

Numerical results showed that the developed design provides substantial communication overhead reduction (up to 80%) while approaching the performances of conventional (uncompressed) FL tools. Additionally, balancing gradient-based and randomized selection policies is beneficial for heavily limiting communication resource consumption without introducing accuracy penalties. Interestingly, layers possessing the least number of trainable parameters should be selected more frequently as they heavily impact the learned models’ quality [9]. The interested reader can find additional layer selection strategies and further numerical evaluations in [6,7,8,9, 32].

4 Cooperative Localization and Sensing in Connected Vehicle Scenarios

The previous sections treated localization and sensing as two separate tasks. However, in next-generation communication systems, such as 6G, these two functionalities are expected to be integrated into the same infrastructure so as to exploit even more their synergy. In line with this trend, this section introduces a more complete system compared to Sects. 2 and 3 integrating cooperative localization and sensing into a unified solution where the goal is to augment the Global Navigation Satellite System (GNSS) performances under complex urban environments.

Perception sensors, particularly Lidar devices, have been increasingly adopted in mobility systems to provide detailed and rich information on the surrounding environment [24]. Cooperative methods have also been studied in such systems [22, 33] to improve environmental awareness by fusing information across multiple interconnected agents. However, considering the sheer amount of data generated by these sensors, conventional signal processing tools may be inadequate as they might introduce large delays. On the other hand, data-driven methods enable the efficient processing of large data volumes while also extracting useful information beneficial for jointly carrying out positioning and sensing tasks [17].

4.1 Data-Driven Joint Cooperative Localization and Perception

Based on the above discussion, this section develops a data-driven cooperative positioning and environmental sensing solution to increase the vehicles’ localization performance compared to conventional GNSS-based systems. The proposal’s main idea is to make the vehicles align their (limited) view of the surrounding environment with other vehicles to improve the detection of the objects along the road and implicitly refine the vehicle positioning as well.

Fig. 3
A diagram illustrates the data-driven localization and sensing. It includes lidar point cloud, D N N model, bounding boxes, G N S S, and measurement aggregation at the road infrastructure, data association, co-operative Bayesian estimation, and vehicle and objects state prediction.

Data-driven cooperative localization and sensing: vehicles employ a ML model to localize static objects in the driving environment via lidar sensors. The individual detections are collected at a centralized infrastructure which is able to refine both the objects and the vehicle positions

Full size image

The developed method extends the Implicit Cooperative Positioning (ICP) framework introduced in [12] by integrating a realistic lidar sensing platform. In particular, a Deep Neural Network (DNN)-assisted sensing framework is designed to recognize and localize road objects (e.g., poles) from lidar sensors available at moving vehicles. The DNN-based detector learns how to recognize static objects as their use has been acknowledged to provide better benefits in ICP [12]. In particular, the detection process focuses on recognizing poles since they are largely present in the driving environment, easily recognizable through the lidar point cloud, and do not require new installations and/or calibrations. Once the vehicles have estimated the position of the poles present in the driving environment, the aggregated information is collected by a centralized road infrastructure which is tasked to cooperatively localize both objects and vehicles employing a Bayesian tracking tool. By doing so, multiple poles estimated at different vehicles can be coherently fused and exploited to improve the vehicles’ positioning accuracy. A block scheme summarizing the main operations required to run the developed approach is shown in Fig. 3.

The evaluation of the proposed approach considers a highly-realistic vehicular scenario simulated using the CARLA software [16], an advanced, high-fidelity autonomous driving simulator that allows defining complex driving conditions as well as generating accurate sensors readings. Numerical results have shown that the developed cooperative localization and sensing approach outperforms a conventional GNSS-based tracking tool while providing similar results to a cooperative oracle system where vehicles always detect all possible poles within the lidar sensing range regardless of actual visibility conditions [4]. The interested reader can look at [4, 11] for the complete description of the methodology.

5 Bayesian Federated Learning for Trustworthy Environmental Perception

Throughout the years, ML tools have been demonstrated to provide excellent performances in solving complex tasks, particularly in big-data regimes where large collections of data are available. However, when data are scarce or limited, NNs trained under the conventional, frequentist, learning paradigm, tend to provide overconfident and often incorrect predictions while also suffering from overfitting. This is further exacerbated when considering FL-based sensing platforms as vehicles may converge to the same unreliable ML model, thereby posing major safety concerns.

Most of the solutions proposed to address the aforementioned challenges rely on Bayesian learning strategies, where the goal is to learn the posterior distribution of the ML model parameters in place of finding a single model parameters’ value that fits well the training data [23]. Some Bayesian FL systems have been recently proposed based on the Partitioned Variational Inference (PVI) framework developed in [1] or on Markov Chain Monte Carlo (MCMC)-based sampling schemes [18]. Still, implementations of Bayesian FL systems over cooperative wireless networks typically assume noiseless communications and, thus, are hardly applicable in real-world scenarios.

5.1 Channel-Driven Bayesian FL Strategy

This section presents a fully decentralized Bayesian FL framework for trustworthy environmental perception in vehicular networks. Compared to the previously-analyzed FL system introduced in Sect. 3, here, the proposal extends the frequentist tools to embrace Bayesian learning strategies. The aim is to obtain ML models that concurrently provide accurate perception capabilities and reliably quantify the uncertainty associated with their predictions. Besides, the proposed method exploits the noise introduced by the propagation in a novel fashion to wirelessly implement the Bayesian FL process.

To obtain an approximation of the global posterior distribution shared by all vehicles, a Bayesian FL system is proposed extending the Decentralized Stochastic Gradient Langevin Dynamics (DSGLD) [18] scheme. The proposal builds on the concept of channel-driven sampling [25], whereby the Bayesian FL strategy is implemented over wireless networks, and the channel noise introduced by the propagation is repurposed for obtaining the final posterior distribution. Indeed, under DSGLD, vehicles update their local posterior samples using Stochastic Gradient Descent (SGD), combine the samples received from their neighbors using a consensus strategy and, finally, add Gaussian noise to obtain a new sample approximating the (global) posterior distribution. Therefore, channel-driven sampling allows each vehicle to directly use the channel noise for the sampling process of DSGLD. An over-the-air computing policy is also proposed to wirelessly aggregate the samples produced at the vehicles in an analog fashion so as to reduce the training latency associated with the cooperative learning process. The block scheme summarizing the proposed strategy is depicted in Fig. 4. For more details on the methodology, the interested reader can refer to [10].

Fig. 4
A block diagram illustrates the decentralized Bayesian F L system. It includes local optimization, sample exchange through the network, and consensus-based sample aggregation.

Decentralized Bayesian FL system: vehicles exchange samples drawn from their local posterior and fuse those received from their neighbors following a consensus-based aggregation strategy

Full size image

The developed Bayesian FL tool is evaluated considering the same cooperative sensing task as in Sect. 3 and is compared against a standard frequentist FL tool. Numerical results show that the proposed strategy provides highly-accurate perception models that reliably quantify the uncertainty of their predictions, while the conventional FL strategy lacks such uncertainty quantification and consequently provides unreliable ML models [10].

6 Concluding Remarks

This chapter presented several methodological advancements aimed at enhancing localization accuracy, environmental awareness, or both in multi-agent networks. Cooperative systems underpin the proposed algorithms in order to provide enhanced environmental awareness or augmented localization performances, thanks to collaborative functions implemented by interconnected agents, devices, or vehicles. ML methods are also instrumental in achieving highly-accurate results especially when conventional methods fail to provide any reasonable outcome. Shifting from standard signal processing tools to ML strategies is also often required, if not mandatory, as model-driven approaches may be too complex to implement or require too much time to be formulated. Moreover, they enable efficient and timely processing of massive amounts of data that may also be required for latency-critical services.

The techniques discussed in this chapter represent fundamental building blocks that can be combined for developing a larger, more refined localization and sensing system, where accuracy is not the only performance metric to be considered. We believe the proposed framework is a starting point that could be extended to embrace novel technologies, ad-hoc implementations, or better integration possibly looking at future technological developments.