Autonomous vehicles have emerged as one of the most promising innovations in the field of transportation and mobility. Predicting the motion of surrounding vehicles is essential for safe autonomous driving, enabling to anticipate potential hazards and make informed decisions. This thesis presents the design and implementation of a map-aware motion prediction algorithm with a focus on intersection scenarios, specifically developed in the context of the AIDA autonomous driving project. The proposed algorithm employs an Extended Kalman Filter (EKF) framework complemented with lanelet-based map information derived from OpenStreetMap data to constrain vehicle trajectories within geometrically feasible paths. For each vehicle in the scene, multiple instances of EKF are initialized, with each filter corresponding to a distinct trajectory hypothesis. The plausibility of each hypothesis is evaluated using a combination of EKF residuals, velocity consistency metrics, and a prior over typical acceleration behaviors, resulting in a multi-modal prediction. Geometric consistency is captured through the EKF framework, while dynamic feasibility is assessed based on velocity patterns. A key contribution of this work is the introduction of speed profiles for each candidate path, which define feasible velocities by accounting for road geometry and regulatory speed limits. The algorithm balances computational efficiency, interpretability, and real-time applicability without requiring extensive training datasets typical of deep learning approaches. To support algorithm development and validation, a set of representative simulated scenarios involving non-signalized intersections was constructed. The system was tested both in simulated scenarios and on real data.
La guida autonoma rappresenta una delle innovazioni più promettenti nel campo della mobilità e dei trasporti. Tra le sue sfide principali, la previsione del movimento dei veicoli circostanti è fondamentale per garantire la sicurezza, permettendo di anticipare situazioni potenzialmente pericolose e di prendere decisioni informate. Questa tesi presenta un algoritmo di previsione del moto sviluppato nell’ambito del progetto di guida autonoma AIDA, con un focus specifico sugli scenari di intersezione non semaforica. L’approccio proposto integra informazioni di mappa in formato lanelet, all’interno di un filtro di Kalman esteso (EKF), al fine di vincolare le traiettorie dei veicoli a percorsi geometricamente plausibili. Per ciascun veicolo presente nella scena, vengono inizializzate più istanze del filtro EKF, ciascuna corrispondente a un’ipotesi di traiettoria. La plausibilità di ogni ipotesi viene valutata combinando i residui di innovazione del filtro, metriche di coerenza della velocità e una distribuzione a priori basata sull’accelerazione, ottenendo così una previsione multi-modale. La coerenza geometrica è gestita tramite i residui dell’EKF, mentre la fattibilità dinamica è valutata in base a informazioni di velocità. Uno dei contributi chiave è l’introduzione di profili di velocità per ciascun percorso candidato, che definiscono le velocità ammissibili tenendo conto della geometria stradale e dei limiti normativi. L’algoritmo è progettato per bilanciare efficienza computazionale, interpretabilità e applicabilità in tempo reale, senza la necessità delle grandi quantità di dati di addestramento tipiche degli approcci basati sulle reti neurali. Lo sviluppo e la validazione dell’algoritmo sono stati supportati da un insieme di scenari rappresentativi in ambiente simulato, con test successivi anche su dati reali.
Map-aware Kalman filter-based motion prediction at intersection scenarios
El Gamrani, Jasmine
2024/2025
Abstract
Autonomous vehicles have emerged as one of the most promising innovations in the field of transportation and mobility. Predicting the motion of surrounding vehicles is essential for safe autonomous driving, enabling to anticipate potential hazards and make informed decisions. This thesis presents the design and implementation of a map-aware motion prediction algorithm with a focus on intersection scenarios, specifically developed in the context of the AIDA autonomous driving project. The proposed algorithm employs an Extended Kalman Filter (EKF) framework complemented with lanelet-based map information derived from OpenStreetMap data to constrain vehicle trajectories within geometrically feasible paths. For each vehicle in the scene, multiple instances of EKF are initialized, with each filter corresponding to a distinct trajectory hypothesis. The plausibility of each hypothesis is evaluated using a combination of EKF residuals, velocity consistency metrics, and a prior over typical acceleration behaviors, resulting in a multi-modal prediction. Geometric consistency is captured through the EKF framework, while dynamic feasibility is assessed based on velocity patterns. A key contribution of this work is the introduction of speed profiles for each candidate path, which define feasible velocities by accounting for road geometry and regulatory speed limits. The algorithm balances computational efficiency, interpretability, and real-time applicability without requiring extensive training datasets typical of deep learning approaches. To support algorithm development and validation, a set of representative simulated scenarios involving non-signalized intersections was constructed. The system was tested both in simulated scenarios and on real data.File | Dimensione | Formato | |
---|---|---|---|
2025_07_ElGamrani_ExecutiveSummary_02.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
5.18 MB
Formato
Adobe PDF
|
5.18 MB | Adobe PDF | Visualizza/Apri |
2025_07_ElGamrani_Tesi_01.pdf
non accessibile
Descrizione: Testo Tesi
Dimensione
73.25 MB
Formato
Adobe PDF
|
73.25 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/10589/240342