Autonomous vehicles rely on accurate perception of the road environment to operate effectively. Perception systems transform raw sensor data into meaningful representations of road features, enabling vehicles to interpret their surroundings and plan maneuvers. These representations must be accurate enough to ensure safe navigation while remaining computationally tractable in real-time environments. Different operational scenarios therefore call for specifically tailored representations in order to meet these conflicting requirements. This thesis systematically investigates this design challenge, studying the development and maintenance of road representations across key operational contexts: offline mapping with abundant computational resources, online detection under strict real-time constraints, and adaptive strategies for dynamic environments. We focus predominantly on road line markings, distinctive features of road environments and essential for navigation, yet our analysis generalizes to broader road feature challenges. First, we explore offline mapping of road lines using compact geometric models based on clothoid curves. These models achieve high accuracy and provide interpretable trajectory information for real-time control, making them suitable for integration into High-Definition (HD) maps, yet they exhibit limited convenience in complex urban scenarios, such as intersections and roundabouts. Second, we address these limitations and further advance offline mapping by adopting a complementary model-free approach characterized by higher expressiveness. Capturing and aggregating dense line segmentation masks and distilling them into topological graphs, we obtain compact and accurate road line maps especially where geometric models are more limited. Next, recognizing that HD maps may be absent or outdated in certain areas, we investigate online road line detection. By improving frame-by-frame accuracy and incorporating short-term detection aggregation, we enhance robustness against noise and occlusions while satisfying real-time constraints. Finally, we address the dynamic nature of road environments, which requires adaptive perception. We discuss strategies for map maintenance and propose continual deep learning approaches to handle appearance variations. Through this comprehensive analysis, we highlight the fundamental trade-offs faced by road representations and provide concrete solutions for their efficient offline mapping, robust online detection, and adaptive long-term maintenance.
I veicoli a guida autonoma richiedono una percezione accurata dell'ambiente stradale per operare in maniera efficace. I sistemi di percezione trasformano i dati grezzi dei sensori in rappresentazioni significative delle caratteristiche della strada, consentendo ai veicoli di interpretare l'ambiente circostante e pianificare le loro manovre. Tali rappresentazioni devono essere sufficientemente precise da garantire una navigazione sicura, ma al tempo stesso computazionalmente efficienti per l'elaborazione in tempo reale. Diverse condizioni operative richiedono pertanto rappresentazioni specificamente adattate per bilanciare questi requisiti contrastanti. Questa tesi analizza in modo sistematico tale sfida di progettazione, studiando lo sviluppo e la manutenzione delle rappresentazioni di elementi stradali in diversi contesti operativi: mappatura offline con risorse computazionali abbondanti, rilevamento online con vincoli di computazione in tempo reale e strategie adattive per ambienti dinamici. L'attenzione è rivolta principalmente alla segnaletica orizzontale, elemento distintivo dell'ambiente stradale e fondamentale per la navigazione, ma la nostra analisi si estende in modo generale anche ad altre caratteristiche stradali. In primo luogo, esploriamo la mappatura offline della segnaletica utilizzando modelli geometrici compatti basati su curve clotoidali. Tali modelli garantiscono elevata accuratezza e forniscono in tempo reale informazioni di traiettoria interpretabili dai moduli addetti al controllo, rendendoli adatti all'integrazione in mappe ad alta definizione (mappe HD). Tuttavia, questi modelli risultano meno efficaci in scenari urbani complessi che includono incroci e rotonde. In secondo luogo, dunque, affrontiamo queste ultime limitazioni e miglioriamo ulteriormente la mappatura offline adottando un approccio complementare di tipo non parametrico, caratterizzato da una maggiore espressività. Catturando e aggregando maschere di segmentazione dense delle linee stradali e condensandole in grafi topologici, otteniamo mappe compatte e accurate, particolarmente efficaci dove i modelli geometrici risultano più restrittivi. Riconoscendo poi che le mappe HD possono essere assenti o non aggiornate in alcune aree, indaghiamo il rilevamento online della segnaletica orizzontale. Migliorando l'accuratezza fotogramma per fotogramma e introducendo una aggregazione temporale a breve termine, aumentiamo la robustezza rispetto a rumore e occlusioni mantenendo i vincoli di computazione in tempo reale. Infine, affrontiamo la natura dinamica dell'ambiente stradale, che richiede capacità di percezione adattiva. Discutiamo strategie per la manutenzione delle mappe e proponiamo approcci di apprendimento continuo per gestire variazioni di aspetto nel tempo. Attraverso questa approfondita analisi, evidenziamo i principali compromessi alla base delle rappresentazioni stradali e forniamo soluzioni concrete per la loro mappatura offline efficiente, rilevamento online robusto e manutenzione adattiva nel lungo periodo.
Road representations for autonomous navigation: offline mapping, online detection and dynamic adaptation
Cudrano, Paolo
2024/2025
Abstract
Autonomous vehicles rely on accurate perception of the road environment to operate effectively. Perception systems transform raw sensor data into meaningful representations of road features, enabling vehicles to interpret their surroundings and plan maneuvers. These representations must be accurate enough to ensure safe navigation while remaining computationally tractable in real-time environments. Different operational scenarios therefore call for specifically tailored representations in order to meet these conflicting requirements. This thesis systematically investigates this design challenge, studying the development and maintenance of road representations across key operational contexts: offline mapping with abundant computational resources, online detection under strict real-time constraints, and adaptive strategies for dynamic environments. We focus predominantly on road line markings, distinctive features of road environments and essential for navigation, yet our analysis generalizes to broader road feature challenges. First, we explore offline mapping of road lines using compact geometric models based on clothoid curves. These models achieve high accuracy and provide interpretable trajectory information for real-time control, making them suitable for integration into High-Definition (HD) maps, yet they exhibit limited convenience in complex urban scenarios, such as intersections and roundabouts. Second, we address these limitations and further advance offline mapping by adopting a complementary model-free approach characterized by higher expressiveness. Capturing and aggregating dense line segmentation masks and distilling them into topological graphs, we obtain compact and accurate road line maps especially where geometric models are more limited. Next, recognizing that HD maps may be absent or outdated in certain areas, we investigate online road line detection. By improving frame-by-frame accuracy and incorporating short-term detection aggregation, we enhance robustness against noise and occlusions while satisfying real-time constraints. Finally, we address the dynamic nature of road environments, which requires adaptive perception. We discuss strategies for map maintenance and propose continual deep learning approaches to handle appearance variations. Through this comprehensive analysis, we highlight the fundamental trade-offs faced by road representations and provide concrete solutions for their efficient offline mapping, robust online detection, and adaptive long-term maintenance.| File | Dimensione | Formato | |
|---|---|---|---|
|
Cudrano_PhD_Thesis.pdf
accessibile in internet per tutti
Descrizione: PhD Thesis
Dimensione
84.7 MB
Formato
Adobe PDF
|
84.7 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/10589/245017