Pose estimation with respect to a global reference frame is a fundamental requirement for autonomous navigation. In indoor environments, where satellite navigation systems are unavailable, this estimation requires the integration of alternative sources of information. This thesis addresses mobile robot localization in a known map, focusing on perceptual ambiguity in repetitive environments, localization recovery following uncommanded displacements, and the computational constraints of onboard hardware. A two-stage estimation architecture is proposed. The first stage employs a strictly proprioceptive Extended Kalman Filter (EKF), which fuses odometry and inertial measurements to provide a continuous estimate subject to drift. The second stage uses a Particle Filter (PF) to localize the robot within the map by combining two information sources: a dense Bird's Eye View (BEV) visual observation, matched against the map using a likelihood field, and an absolute position sensor with a low update rate, implemented in the case study using an Ultra-Wideband (UWB) system. The separation of information sources between the two stages is a central design principle: the EKF provides a local dead reckoning estimate, while the particle filter incorporates absolute information for global localization. This separation preserves the consistency of the particle filter's motion model by preventing absolute position corrections in the first stage from being interpreted as robot motion. The localization module, designed for integration into ROS 2 architectures for autonomous robots, employs the canonical odometry motion model and systematic resampling with an adaptive particle set size. The absolute position sensor serves three complementary roles: providing asynchronous measurement updates, with contributions accumulated in log space and incorporated into the main observation cycle; supplying a reference for reinitialization; and defining the center of the distribution from which new hypotheses are injected during recovery. Localization error detection and recovery activation rely on three distinct and complementary channels that monitor the mean observation likelihood, the absolute position sensor residual, and the absolute quality of map matching, respectively. The method is designed for integration into different robotic platforms. Development and experimental validation were conducted within the framework of the Bosch Future Mobility Challenge 2026, using a 1:10-scale vehicle that employed the presented localization system during the competition trials. Analysis of the logs recorded during the test sessions on the competition track enables assessment of the overall behavior of the developed system. In the absence of independent pose ground truth, the results support an assessment of system consistency and behavior, but do not allow its absolute accuracy to be quantified.
La stima della posa rispetto a un riferimento globale costituisce un requisito fondamentale per la navigazione autonoma. Negli ambienti interni, dove i sistemi di navigazione satellitare non sono disponibili, tale stima richiede l'integrazione di sorgenti informative alternative. Questa tesi affronta il problema della localizzazione di robot mobili su mappa nota, concentrandosi sull'ambiguità percettiva degli scenari ripetitivi, sul recupero della localizzazione dopo spostamenti non comandati e sui vincoli computazionali del calcolatore di bordo. Si propone un'architettura di stima a due stadi. Il primo impiega un filtro di Kalman esteso (Extended Kalman Filter, EKF), strettamente propriocettivo, che integra odometria e misure inerziali per fornire una stima continua, ma soggetta a deriva. Il secondo utilizza un filtro particellare (Particle Filter, PF) per localizzare il robot sulla mappa mediante due sorgenti informative: un'osservazione visiva densa in Bird's Eye View (BEV), confrontata con la mappa attraverso un likelihood field, e un sensore di posizione assoluta a bassa frequenza di aggiornamento, rappresentato nel caso di studio da un sistema Ultra-Wideband (UWB). La separazione delle sorgenti informative tra i due stadi costituisce un criterio progettuale dell'architettura: l'EKF fornisce una stima locale di dead reckoning, mentre il filtro particellare integra le informazioni assolute per la localizzazione globale. Tale distinzione preserva la coerenza del modello di moto del filtro particellare, evitando che le correzioni di posizione assoluta introdotte nel primo stadio siano interpretate come spostamenti del robot. Il modulo di localizzazione, progettato per integrarsi in architetture ROS 2 per robot autonomi, adotta il modello di moto odometrico canonico e il ricampionamento sistematico, con dimensionamento adattivo dell'insieme di particelle. L'integrazione del sensore assoluto prevede tre ruoli complementari: aggiornamento di misura asincrono, con contributi accumulati in spazio logaritmico e incorporati nel ciclo principale di osservazione; riferimento per le reinizializzazioni e infine centro della distribuzione di iniezione di nuove ipotesi durante il recovery. Il rilevamento degli errori di localizzazione e l'attivazione delle procedure di recovery si basano su tre canali distinti e complementari, che monitorano rispettivamente la verosimiglianza media dell'osservazione, il residuo del sensore assoluto e la qualità assoluta del matching con la mappa. Il metodo è progettato per essere integrato in diverse piattaforme robotiche. Lo sviluppo e la validazione sperimentale sono stati condotti nel contesto della Bosch Future Mobility Challenge 2026, su un veicolo in scala 1:10 che ha impiegato il sistema di localizzazione presentato durante le prove di gara. L'analisi dei log acquisiti nelle sessioni di prova sul tracciato di gara consente di valutare il comportamento generale del sistema sviluppato. In assenza di una ground truth indipendente della posa, i risultati permettono di esaminare la coerenza e il comportamento del sistema, ma non di quantificarne l'accuratezza assoluta.
Architettura EKF–Particle Filter a doppio stadio assistita dalla visione per la localizzazione multisensoriale in ambienti interni di veicoli robotici a guida autonoma
BRUNELLA, FEDERICO
2025/2026
Abstract
Pose estimation with respect to a global reference frame is a fundamental requirement for autonomous navigation. In indoor environments, where satellite navigation systems are unavailable, this estimation requires the integration of alternative sources of information. This thesis addresses mobile robot localization in a known map, focusing on perceptual ambiguity in repetitive environments, localization recovery following uncommanded displacements, and the computational constraints of onboard hardware. A two-stage estimation architecture is proposed. The first stage employs a strictly proprioceptive Extended Kalman Filter (EKF), which fuses odometry and inertial measurements to provide a continuous estimate subject to drift. The second stage uses a Particle Filter (PF) to localize the robot within the map by combining two information sources: a dense Bird's Eye View (BEV) visual observation, matched against the map using a likelihood field, and an absolute position sensor with a low update rate, implemented in the case study using an Ultra-Wideband (UWB) system. The separation of information sources between the two stages is a central design principle: the EKF provides a local dead reckoning estimate, while the particle filter incorporates absolute information for global localization. This separation preserves the consistency of the particle filter's motion model by preventing absolute position corrections in the first stage from being interpreted as robot motion. The localization module, designed for integration into ROS 2 architectures for autonomous robots, employs the canonical odometry motion model and systematic resampling with an adaptive particle set size. The absolute position sensor serves three complementary roles: providing asynchronous measurement updates, with contributions accumulated in log space and incorporated into the main observation cycle; supplying a reference for reinitialization; and defining the center of the distribution from which new hypotheses are injected during recovery. Localization error detection and recovery activation rely on three distinct and complementary channels that monitor the mean observation likelihood, the absolute position sensor residual, and the absolute quality of map matching, respectively. The method is designed for integration into different robotic platforms. Development and experimental validation were conducted within the framework of the Bosch Future Mobility Challenge 2026, using a 1:10-scale vehicle that employed the presented localization system during the competition trials. Analysis of the logs recorded during the test sessions on the competition track enables assessment of the overall behavior of the developed system. In the absence of independent pose ground truth, the results support an assessment of system consistency and behavior, but do not allow its absolute accuracy to be quantified.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Federico_Brunella_MSc_06-10-2026_v2.pdf
embargo fino al 18/04/2028
Dimensione
11.25 MB
Formato
Adobe PDF
|
11.25 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/29203