Perceiving, modeling, and estimating the surrounding environment is essential for autonomous vehicles. In order to achieve this objective, it is important to appropriately process and model measurement sensor data. Nowadays, autonomous cars are equipped with a variety of technology solutions capable of gathering a great quantity of data from their surroundings that needs to be merged. The occupancy grid map is a common method for modeling and representing static environments. The modular structure allows for the representation of objects of any shape and addresses the data association problem between several measurement data and estimated objects intuitively. More advanced solutions capable of managing and representing dynamic objects have been proposed. The Bayesian occupancy filter (BOF) is a solution that implements these advancements while also paving the path for numerous additional increasingly advanced and efficient alternatives such as the Dempster-Shafer probability hypothesis density/multi-instance Bernoulli filter (DS-PHD/MIB). This work presents a solution that is capable to perform environment estimation and obstacle detection deeply merging data from multiple sensors data, proposing a method that covers the process from the collection of raw data from different sensors, up to their fusion within the estimation model. The thesis presents the multi-layer occupancy grid map input framework (MOGMIF), a framework that processes raw sensor data into a multi-layer occupancy grid map data structure. Furthermore, this thesis proposes the multi-sensor DS-PHD/MIB, a solution capable of integrating and merging data from different sensors and estimating more consistently and meaningfully the environment surrounding the vehicle. Additionally, the thesis proposes an efficient implementation of the multi-sensor DS-PHD/MIB. Finally, the thesis includes an evaluation of the promising experiments done using real-world data, specifically employing data from 3D lidar measurements, radar measurements, and camera video images to demonstrate the effectiveness of the proposed approach.
Modellare e stimare in modo fedele l'ambiente circostante è essenziale per i veicoli a guida autonoma. Per raggiungere questo obbiettivo è importante processare e modellare i dati delle misurazioni dei sensori in modo appropriato. Al giorno d'oggi, i veicoli a guida autonoma sono dotati di differenti soluzioni tecnologiche in grado di fornire un grande quantitativo di informazioni dell'ambiente circostante, che hanno però bisogno di essere integrati. La occupancy grid map rappresenta un diffuso stumento per modellare e rappresentare l'ambiente in modo statico. La sua struttura modulare permette la rappresentazione di oggetti con ogni tipo di forma e inoltre risolve in maniera intuitiva il problema di associare i dati di diverse misurazioni con gli oggetti stimati. A seguire, soluzioni più avanzate e in grado di gestire e rappresentare oggetti dinamici sono state proposte. Il Bayesian occupancy filter (BOF) è una soluzione che introduce questi progressi, aprendo la strada a numerosi lavori sempre più avanzati ed efficienti come il filtro Dempster-Shafer probability hypothesis density/multi-instance Bernoulli filter (DS-PHD/MIB). Questo lavoro presenta una soluzione in grado di effettuare una stima dell'ambiente e di individuare ostacoli, fondendo efficacemente i dati provenienti da diversi sensori, proponendo una metodologia che copre il processo a partire a partire dalla raccolta dei dati grezzi, fino alla loro fusione all'interno del modello di previsione. Questa tesi presenta il multi-layer occupancy grid map input framework (MOGMIF), un framework che elabora i dati grezzi dei sensori e li rappresenta in una struttura basata su una multi-layer occupancy grid map. Inoltre, questa tesi propone il multi-sensor DS-PHD/MIB, una soluzione in grado di integrare e fondere i dati provenienti da diversi sensori, e stimare l'ambiente circostante al veicolo, in maniera più affidabile e significativa. In aggiunta, la tesi propone una soluzione efficente del multi-sensor DS-PHD/MIB. In fine, la tesi presenta la valutazione dei promettenti esperimenti eseguiti utilizzando dati del mondo reale. Nello specifico i dati usati sono raccolti da misurazioni di 3D lidar, radar e da immagini video di una videocamera. Questi test dimostrano l'efficacia della soluzione proposta.
Multi-sensor data fusion for grid-based environment estimation
Trivilino, Manuel
2020/2021
Abstract
Perceiving, modeling, and estimating the surrounding environment is essential for autonomous vehicles. In order to achieve this objective, it is important to appropriately process and model measurement sensor data. Nowadays, autonomous cars are equipped with a variety of technology solutions capable of gathering a great quantity of data from their surroundings that needs to be merged. The occupancy grid map is a common method for modeling and representing static environments. The modular structure allows for the representation of objects of any shape and addresses the data association problem between several measurement data and estimated objects intuitively. More advanced solutions capable of managing and representing dynamic objects have been proposed. The Bayesian occupancy filter (BOF) is a solution that implements these advancements while also paving the path for numerous additional increasingly advanced and efficient alternatives such as the Dempster-Shafer probability hypothesis density/multi-instance Bernoulli filter (DS-PHD/MIB). This work presents a solution that is capable to perform environment estimation and obstacle detection deeply merging data from multiple sensors data, proposing a method that covers the process from the collection of raw data from different sensors, up to their fusion within the estimation model. The thesis presents the multi-layer occupancy grid map input framework (MOGMIF), a framework that processes raw sensor data into a multi-layer occupancy grid map data structure. Furthermore, this thesis proposes the multi-sensor DS-PHD/MIB, a solution capable of integrating and merging data from different sensors and estimating more consistently and meaningfully the environment surrounding the vehicle. Additionally, the thesis proposes an efficient implementation of the multi-sensor DS-PHD/MIB. Finally, the thesis includes an evaluation of the promising experiments done using real-world data, specifically employing data from 3D lidar measurements, radar measurements, and camera video images to demonstrate the effectiveness of the proposed approach.File | Dimensione | Formato | |
---|---|---|---|
Manuel_Trivilino_Master_Thesis.pdf
Open Access dal 13/04/2023
Descrizione: Manoscritto Tesi Magistrale
Dimensione
16.3 MB
Formato
Adobe PDF
|
16.3 MB | Adobe PDF | Visualizza/Apri |
Manuel_Trivilino_Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
7.12 MB
Formato
Adobe PDF
|
7.12 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/10589/187083