Conventional single-agent localization methods have been demonstrated to provide unsatisfactory performances for mission-critical control applications where strict requirements are imposed, such as ultra-high reliability, ultra-low latency, and ultra-accurate positioning, due to the limited sensing/computing capabilities of ego systems. In contrast, cooperative positioning approaches enable interconnected agents to share information across the network with huge benefits in terms of accuracy, reliability, and safety. This thesis focuses on the development of novel cooperative localization and learning strategies in the context of mission-critical control networks. The goal is to provide competitive solutions for obtaining precise positioning in harsh propagating environments as well as reliable environmental mapping in highly-dynamic scenarios. In the first part of the thesis, the localization problem is initially tackled by proposing novel augmentation strategies allowing to improve the reliability of the sensing at each agent. In particular, we consider a wireless network where each connected agent is tasked to localize itself based on location measurements extracted from radio signals. To cope with complex propagating conditions originating from the environment in which the agents are deployed, the proposed augmentation strategy statistically describes the propagation characteristics of the environment and combines hybrid localization measurements so as to reduce the uncertainty of the agents' position. Once the agents are able to localize themselves, the research is moved to the mapping of the surrounding environment. Specifically, the proposed perception system leverages distributed learning methods for reliable perception at the agents. Fully decentralized, consensus-driven Federated Learning (FL) strategies are developed where in-network processing functions among cooperating agents replace energy-hungry operations carried out at a centralized location to enhance the resilience of the overall training platform. Several communication-efficient designs are proposed to optimize the accuracy, latency, or training time by selecting in an intelligent way the parameters that have to be exchanged over the network during the FL optimization. Then, localization and environmental mapping functionalities are merged into a unified framework. Under this framework, agents are assumed to be equipped with imaging sensors, namely Lidar devices, for collecting information on their surroundings. Data-driven methods are designed to let the agents efficiently process the lidar point clouds and localize passive static targets present in the environment. The cooperation is then exploited to coherently fuse the individual detections made by the agents and consequently improve the localization of the targets. Once the targets have been localized with high accuracy, they are exploited to further refine the agents' position. Finally, the research activities are concluded by proposing a trustworthy environmental perception system. The proposed framework integrates Bayesian inference tools into the aforementioned FL-based perception methods so as to reliably quantify uncertainty arising from limited data availability at the agents. Compared to the previously-studied FL systems that target the learning of a single value for the ML model parameters, the goal of the proposed Bayesian FL approach is to learn the (shared) global posterior distribution across all cooperating agents. Employing such a scheme allows the agents to weigh their decisions according to the posterior and, thus, provide reliable predictions that can be employed under safety-critical conditions. The results achieved during the Ph.D. demonstrate that the proposed approaches can be applied to a wide range of challenging problems that require highly-accurate, low-latency, trustworthy outcomes which are fundamental requirements envisioned for future Industrial Internet of Things (IIoT) and Connected Automated Vehicle (CAV) services.
I metodi convenzionali di localizzazione a singolo agente hanno dimostrato di fornire prestazioni insoddisfacenti nelle applicazioni mission-critical dove requisiti stringenti sono imposti, quali altissima affidabilità, bassissima latenza e posizionamento ultra-accurato, a causa delle limitate capacità di rilevamento/computing dei singoli agenti. Al contrario, gli approcci di localizzazione cooperativa consentono agli agenti di condividere informazioni attraverso la rete con enormi vantaggi in termini di accuratezza, affidabilità e sicurezza. Questa tesi si concentra sullo sviluppo di nuove strategie di localizzazione e apprendimento cooperativo nel contesto delle reti mission-critical. L'obiettivo è quello di fornire soluzioni competitive al fine di ottenere una localizzazione ad alta precisione in ambienti complessi e una mappatura ambientale affidabile in scenari altamente dinamici. Nella prima parte della tesi, il problema della localizzazione viene inizialmente affrontato proponendo nuove strategie che consentono di migliorare l'affidabilità del rilevamento di ciascun agente. In particolare, si considera una rete wireless in cui ogni agente ha il compito di localizzarsi sulla base di misure di posizione estratte da segnali radio. Per far fronte alle complesse condizioni di propagazione derivanti dall'ambiente in cui gli agenti sono collocati, la strategia proposta descrive statisticamente le caratteristiche di propagazione dell'ambiente e combina misure di localizzazione ibride in modo da ridurre l'incertezza della posizione degli agenti. Una volta che gli agenti sono in grado di localizzarsi, la ricerca si concentra sulla mappatura dell'ambiente circostante. Il sistema di percezione proposto sfrutta metodi di apprendimento distribuito per aumentare le capacità di percezione degli agenti. Strategie di Federated Learning (FL) completamente decentralizzate basate sul consenso sono state sviluppate al fine di ottenere una piattaforma di allenamento federato più robusta rispetto a tecniche FL convenzionali basate su architetture centralizzate. Inoltre, diversi metodi per aumentare l’efficienza comunicativa del processo federato sono stati proposti per ottimizzare l’accuratezza, la latenza o il tempo di addestramento andando a selezionare in maniera intelligente i parametri che gli agenti vanno a scambiarsi attraverso la rete. In seguito, le funzionalità di localizzazione e mappatura ambientale vengono unite in unico framework. Il framework proposto considera agenti dotati di sensori di imaging, in questo caso sensori Lidar, che permettono di collezionare informazioni riguardo all’ambiente circostante. Ogni agente utilizza un modello di apprendimento automatico per rilevare e localizzare possibili oggetti statici presenti nell’ambiente. La cooperazione è poi sfruttata per combinare in modo coerente le informazioni estratte dagli agenti e migliorare la localizzazione degli oggetti rilevati. Una volta che gli oggetti vengono localizzati con più alta precisione, la loro posizione viene retroattivamente sfruttata per affinare la posizione degli agenti. Infine, le attività di ricerca si concludono un sistema di percezione ambientale affidabile. Il framework proposto integra strumenti di inferenza Bayesiana nei suddetti metodi di percezione basati sul FL in modo da quantificare in modo affidabile l'incertezza derivante dalla limitata disponibilità di dati presso gli agenti. Rispetto ai sistemi FL precedentemente studiati che mirano all'apprendimento di un singolo valore per i parametri del modello di apprendimento automatico, l'obiettivo dell'approccio FL Bayesiano proposto è quello di apprendere la probabilità a posteriori dei parametri del modello. L'utilizzo di un tale schema consente agli agenti di valutare le loro decisioni in base alla distribuzione a posteriori e fornire decisioni affidabili che possono essere impiegate in condizioni critiche per la sicurezza. I risultati conseguiti durante il dottorato dimostrano che gli approcci proposti possono essere applicati ad un'ampia gamma di problemi complessi che richiedono risultati altamente accurati, a bassa latenza e affidabili, i quali sono requisiti fondamentali previsti per lo sviluppo di servizi Industrial Internet of Things (IIoT) e Connected Automated Vehicle (CAV).
Cooperative Processing and Learning Methods for High-Resolution Environmental Perception
BARBIERI, LUCA
2022/2023
Abstract
Conventional single-agent localization methods have been demonstrated to provide unsatisfactory performances for mission-critical control applications where strict requirements are imposed, such as ultra-high reliability, ultra-low latency, and ultra-accurate positioning, due to the limited sensing/computing capabilities of ego systems. In contrast, cooperative positioning approaches enable interconnected agents to share information across the network with huge benefits in terms of accuracy, reliability, and safety. This thesis focuses on the development of novel cooperative localization and learning strategies in the context of mission-critical control networks. The goal is to provide competitive solutions for obtaining precise positioning in harsh propagating environments as well as reliable environmental mapping in highly-dynamic scenarios. In the first part of the thesis, the localization problem is initially tackled by proposing novel augmentation strategies allowing to improve the reliability of the sensing at each agent. In particular, we consider a wireless network where each connected agent is tasked to localize itself based on location measurements extracted from radio signals. To cope with complex propagating conditions originating from the environment in which the agents are deployed, the proposed augmentation strategy statistically describes the propagation characteristics of the environment and combines hybrid localization measurements so as to reduce the uncertainty of the agents' position. Once the agents are able to localize themselves, the research is moved to the mapping of the surrounding environment. Specifically, the proposed perception system leverages distributed learning methods for reliable perception at the agents. Fully decentralized, consensus-driven Federated Learning (FL) strategies are developed where in-network processing functions among cooperating agents replace energy-hungry operations carried out at a centralized location to enhance the resilience of the overall training platform. Several communication-efficient designs are proposed to optimize the accuracy, latency, or training time by selecting in an intelligent way the parameters that have to be exchanged over the network during the FL optimization. Then, localization and environmental mapping functionalities are merged into a unified framework. Under this framework, agents are assumed to be equipped with imaging sensors, namely Lidar devices, for collecting information on their surroundings. Data-driven methods are designed to let the agents efficiently process the lidar point clouds and localize passive static targets present in the environment. The cooperation is then exploited to coherently fuse the individual detections made by the agents and consequently improve the localization of the targets. Once the targets have been localized with high accuracy, they are exploited to further refine the agents' position. Finally, the research activities are concluded by proposing a trustworthy environmental perception system. The proposed framework integrates Bayesian inference tools into the aforementioned FL-based perception methods so as to reliably quantify uncertainty arising from limited data availability at the agents. Compared to the previously-studied FL systems that target the learning of a single value for the ML model parameters, the goal of the proposed Bayesian FL approach is to learn the (shared) global posterior distribution across all cooperating agents. Employing such a scheme allows the agents to weigh their decisions according to the posterior and, thus, provide reliable predictions that can be employed under safety-critical conditions. The results achieved during the Ph.D. demonstrate that the proposed approaches can be applied to a wide range of challenging problems that require highly-accurate, low-latency, trustworthy outcomes which are fundamental requirements envisioned for future Industrial Internet of Things (IIoT) and Connected Automated Vehicle (CAV) services.| File | Dimensione | Formato | |
|---|---|---|---|
|
Cooperative_Processing_and_Learning_Methods_for_High_Resolution_Environmental_Perception.pdf
non accessibile
Dimensione
23.23 MB
Formato
Adobe PDF
|
23.23 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/202812