DSpace Collection:https://hdl.handle.net/1889/6762024-03-28T10:03:21Z2024-03-28T10:03:21ZAn edge-to-cloud framework for privacy-aware management of geospatial dataPenzotti, Gabrielehttps://hdl.handle.net/1889/55732024-03-05T12:58:08Z2024-01-01T00:00:00ZTitle: An edge-to-cloud framework for privacy-aware management of geospatial data
Authors: Penzotti, Gabriele
Abstract: In today's fast-paced technological landscape, characterized by rapid innovation and transformation across sectors, a compelling demand emerges for sophisticated and adaptable systems within Smart Environments. These dynamic settings, marked by an influx of data from diverse sources and intricate distributed systems, offer remarkable opportunities alongside distinct challenges.
This Thesis responds to the need for a flexible, scalable framework that adeptly navigates the complexities of modern distributed data aggregation and processing systems, while upholding the paramount principles of data privacy and security.
At its core, the framework introduces a distributed architecture, housing a service placement algorithm that seamlessly spans from the Edge to the Cloud, meticulously crafted in accordance with the tenets of Fog Computing. This architectural approach indispensably relies on data privacy, significantly influencing applications and prioritizing reliable data management.
A second pivotal contribution is the Seamless Data Acquisition Protocol (SEAMDAP), a standard-based and modern approach meticulously designed to facilitate data collection within distributed systems. Engineered to be both user-friendly and highly customizable, SEAMDAP streamlines the intricate process of gathering data from a multitude of sources, reducing friction, and enhancing flexibility.
Lastly, the Thesis ventures deeply into the critical realms of data integrity and security, acutely acknowledging the inherent importance of georeferenced data and location verification. A robust architecture is proposed within the framework's toolkit, ensuring that data remains secure during transmission, storage, and processing, and culminating in the exploration of advanced processing techniques such as Homomorphic Encryption and Multi-Party Computation.
Crucially, the direction taken with this framework is firmly anchored in the pursuit of standardizing the realm of smart environments while proactively addressing identified issues. Several tools presented herein have undergone rigorous testing in Smart Farming environments, each accompanied by compelling use cases. The framework holds particular promise in settings characterized by high heterogeneity, an abundance of georeferenced data, a critical need for interoperability among systems operated by diverse stakeholders, and an strong commitment to data privacy.2024-01-01T00:00:00ZGeometry and learning for efficient 3D perceptionOrsingher, Marcohttps://hdl.handle.net/1889/55722024-03-05T12:57:06Z2024-01-01T00:00:00ZTitle: Geometry and learning for efficient 3D perception
Authors: Orsingher, Marco
Abstract: Building a 3D representation of the world is a longstanding challenge in computer vision and machine learning, with applications in virtual and augmented reality, autonomous driving, industrial site scanning, cultural heritage preservation, and more. The main goal of this thesis is to develop efficient algorithms for processing 3D data, by combining classical geometry-based methods with modern deep learning approaches. Efficiency is a crucial aspect of 3D perception, since data are typically acquired by low-cost noisy sensors and must be processed on mobile platforms with limited computational budget. Furthermore, the exponential growth of 3D data sources calls for scalable and efficient processing pipelines. Our first contribution is a novel framework for multi-view 3D reconstruction in urban scenarios. We significantly improve a state-of-the-art classical approach for dense reconstruction, by designing a local-to-global optimization strategy that leads to geometrically consistent surfaces. Moreover, we show how to scale it up to arbitrarily large scenes with a divide and conquer procedure that combines view clustering and view selection, thus allowing for a massive parallelization of the 3D reconstruction process. Secondly, we present two algorithmic advances in efficient training of neural representation for novel view synthesis. We propose to speed up the learning process by focusing on informative rays, which are defined in the 2D image space by high-entropy pixels and in the 3D object space by a sparse set of cameras that ensures scene coverage, while keeping optimal relative baseline. Additionally, we leverage multi-view geometry as pseudo-ground truth to guide the neural implicit field towards high-fidelity 3D models. We also tackle the point cloud upsampling task, with the aim of refining noisy and low-resolution data from cheap range sensors into dense and uniform point clouds. To this end, we formulate the first learning-based approach that allows 3D upsampling with arbitrary scaling factors, including non-integer values, with a single trained model. The main idea is to convert the input to a probabilistic representation and to train a Transformer network to map between samples from such domain and points on the underlying object surface. This flexibility is crucial in real-world applications with computational and bandwidth constraints. Finally, we propose two novel methods for neural network compression. We first show that feature-based knowledge distillation can be improved by complementing the direct feature matching baseline with a teacher features-driven regularization loss, thus enabling the student model to learn more robust latent representations. Then, we introduce a neural compression approach that combines network pruning with self-distillation and significantly improves the sparsity-accuracy tradeoff for several perception tasks. This allows to deploy neural architectures on constrained hardware for fast inference with unprecedented performances.2024-01-01T00:00:00ZTrajectory generation strategies for safe autonomous driving in urban scenarioLaneve, Francescohttps://hdl.handle.net/1889/55712024-03-05T12:55:57Z2024-01-01T00:00:00ZTitle: Trajectory generation strategies for safe autonomous driving in urban scenario
Authors: Laneve, Francesco
Abstract: In this dissertation we develop novel strategies based on nonlinear optimal control techniques for trajectory generation of autonomous vehicles. These strategies are designed to enable the development of autonomous vehicles that navigate dynamic environments while enhancing safety and passenger comfort.
In the first part of the work, we introduce a family of reduced-order car models suited for trajectory generation strategies. We derive the equations of motion for both kinematic and dynamic bicycle models, the latter of which includes tire modeling for a more realistic representation of vehicle behavior. We re-write the kinematic model in terms of longitudinal and lateral coordinates, aligning them with the way humans perceive and control vehicle motion.
In the second part, we propose an optimization-based strategy to address merging maneuvers in busy intersections. We describe vehicle dynamics in terms of longitudinal and transverse coordinates and introduce a "virtual target vehicle" constrained to move within the target lane for merging. We formulate an optimal control problem in terms of longitudinal and lateral coordinates, including the kinematic position error between the autonomous vehicle and the virtual target vehicle. We also use obstacle predictions to enforce suitable kinematic constraints for generating collision-free trajectories. We show the efficacy of the proposed strategy through a set of numerical computations and highlight the main features of the generated trajectories.
In the third part, we present a real-time maneuver generation algorithm. Given a planar road geometry with static and moving obstacles along it, we are interested in finding collision-free maneuvers that satisfied the vehicle dynamics and subject to physical and comfort limits. Based on longitudinal and transverse coordinates, we propose a novel collision avoidance constraint and formulate a suitable optimal control problem. The optimization problem is solved by using a nonlinear optimal control technique that generates (local) optimal trajectories. We demonstrate the efficacy of the proposed algorithm by providing numerical computations on a simulated scenario. Experimental results are presented to demonstrate the efficiency
of the proposed algorithm both in terms of computational effort and dynamic features captured.
In the fourth part, we address the lane change maneuver using a parametric model predictive control approach. We recognize that successful lane changes involve both the decision of when to initiate these maneuvers and the generation of collision-free trajectories. Our approach combines decision-making and planning tasks, guiding the low-level policy through upper-level policy search. Additionally, we incorporate self-supervised learning techniques to adapt to dynamic, online scenarios, ensuring the vehicle can handle unexpected changes in its environment. We provide numerical results that highlight the effectiveness of this approach in improving vehicle maneuvering in dynamic environments.2024-01-01T00:00:00ZMotion planning for multiple autonomous vehicles on a graphArdizzoni, Stefanohttps://hdl.handle.net/1889/55702024-03-05T12:53:57Z2024-01-01T00:00:00ZTitle: Motion planning for multiple autonomous vehicles on a graph
Authors: Ardizzoni, Stefano
Abstract: I veicoli a guida automatizzata (AGV) vengono utilizzati per spostare le merci tra diverse posizioni in un magazzino. Il coordinamento di una flotta di AGV è un problema molto complesso, che viene generalmente gestito da un software chiamato Traffic Manager nel modo seguente.
Il Traffic Manager invia a ciascun AGV un task che prevede l'assegnazione di una posizione target nel magazzino, che deve essere raggiunta per effettuare un'operazione di ritiro o consegna. Quando un AGV riceve un task, sceglie il percorso per raggiungere l'obiettivo, senza tenere conto dei percorsi degli altri AGV.
Tuttavia, questa procedura può portare ad una situazione di "deadlock", cioè ad uno scenario in cui un gruppo di veicoli si blocca reciprocamente lungo il percorso loro assegnato, in modo che non possano raggiungere le rispettive posizioni target.
L'obiettivo di questa tesi è studiare un algoritmo che pianifichi i percorsi di tutti gli AGV, al fine di prevenire ed evitare situazioni di "deadlock".
Poiché gli AGV seguono percorsi predefiniti che collegano le posizioni in cui gli articoli vengono immagazzinati o elaborati, associamo il layout dell'insieme di percorsi virtuali a un grafo diretto. In particolare si tratta di digrafi fortemente connessi, cioè grafi orientati in cui è possibile raggiungere qualsiasi nodo partendo da qualsiasi altro nodo.
Il problema principale che deve essere risolto dal pianificatore di percorsi è il problema Multi-Agent Path Finding (MAPF) su grafi, che consiste nel calcolare una sequenza di movimenti che riposiziona tutti gli agenti sui nodi target assegnati, evitando collisioni.
Poiché il nostro interesse primario è evitare situazioni di "deadlock", ci concentriamo sull'importante compito di trovare una soluzione ammissibile per il MAPF, anche in configurazioni molto affollate.
Una volta trovata una soluzione ammissibile, studiamo anche come migliorarla, avvicinandola il più possibile a quella ottimale. Per soluzione ottimale intendiamo la sequenza di movimenti di ciascun agente che consente di raggiungere tutti gli obiettivi nel più breve tempo possibile. Inoltre, studiamo il problema di trovare tali percorsi per gli AGV tenendo conto di alcuni vincoli spaziali, ad esempio dovuti alle dimensioni dei veicoli (questo problema è chiamato C-MAPF), e il problema di trovare il percorso più veloce, considerando la velocità e vincoli di accelerazione, per un singolo agente (chiamato Bounded Acceleration Shortest Path problem).; Automated-Guided Vehicles (AGVs) are used to move items between different locations in a warehouse. The coordination of a fleet of AGVs is a very complex problem,
which is generally handled by a software called Traffic Manager in the following way.
The Traffic Manager send to each AGVs a task that involves the assignment of a target position in the warehouse, which must be reached to carry out a pick up or delivery operation. When an AGV receives a task, it chooses the path to reach the target, without taking into account the routes of others AGVs.
However, this procedure can lead to a "deadlock" situation, i.e., a scenario in which a group of vehicles are mutually blocked along the route assigned to them, so that they are unable to reach their target position.
The objective of this thesis is to study an algorithm that plan the routes of all AGVs, with the aim of preventing and avoiding "deadlock" situations.
Since AGVs follow predefined paths that connect the locations in which items are stored or processed, then we associate the layout of the set of virtual paths to a directed graph. In particular, we deal with strongly connected digraphs, directed graphs in which it is possible to reach any node starting from any other node.
The main problem that must be solved by the path planner is the Multi-Agent Path Finding (MAPF) problem on graphs, which consists in computing a sequence of movements that repositions all agents to assigned target nodes, avoiding collisions.
Since we are primarily interested in avoiding deadlock situations, we focus on the important task of finding a feasible solution to MAPF, even in crowded configurations.
Once we find a feasible solution, we also study how to improve it, bringing it as close as possible to the optimal one. By optimal solution we mean the sequence of movements for each agent, which allows all targets to be reached in the shortest possible time. Moreover, we study the problem of finding such routes for AGVs taking into account spacial constraints, for instance due to the size of the vehicles (this problem is called C-MAPF), and the problem of finding the fastest path, considering velocity and acceleration constraints, for a single agent (called Bounded Acceleration Shortest Path problem).2024-01-01T00:00:00Z