The rapid evolution of autonomous vehicles (AVs) is reshaping mobility, promising to reduce traffic accidents and revolutionize road safety. In this context, ensuring sufficient safety and comfort for passengers becomes paramount, particularly in scenarios characterized by uncertainty due to the interaction between agents, which emerges as the most challenging task to face when AV’s situational awareness is too low. In this context, roundabouts represent a complex environment to deal with for an AV, making critical the selection of the proper action. This thesis proposes an innovative approach leveraging deep learning techniques to develop an intelligent controller for an AV, with a focus on assessing the behavior of surrounding agents, notably their aggressiveness, increase situation awareness and safety. Exploiting a convolutional neural network (CNN) classifier, images are retrieved from raw time series and used to capture relevant features in agents’ behavior, providing information about their aggressiveness. Thus the AV makes informed "GO/STOP" decisions at roundabout’s entrance, prioritizing passengers’ safety and reduce collisions. The study emphasizes the importance of naturalistic data-set like the round-D, instrumental in understanding real-world vehicle behaviors in roundabouts. Subsequently, the work delves into the process of labeling agents as either aggressive or non-aggressive through the K-means clustering, an unsupervised algorithm useful to split data into the two classes. The evaluation of clustering results, aided by correlation analysis and silhouette scores, validates the robustness of this approach in distinguishing between aggressive and non-aggressive driving behaviors. Gramian Angular Fields (GAF) and Markov Transition Fields (MTF) are exploited to convert time-series into images, enhancing the granularity and richness of CNN’s input, whose novel architecture exhibits commendable accuracy, exceeding 90% in recognizing aggressive behavior. A testing and validation environment has been developed through Python programming language. The scenario comprehends two agents, one commanded via a PS4 joystick, allowing to emulate human behavior, and the second (the AV) guided by a PID controller acting on throttle and steering. In the latter code, all the pipeline whose aim is to classify the behavior of the former agent is implemented and the integration of these controllers in CARLA simulator permits to assess the real time capabilities of the whole framework.
La rapida evoluzione dei veicoli autonomi (AV) sta ridefinendo la mobilità, promettendo di ridurre gli incidenti e rivoluzionare la sicurezza stradale. In questo contesto, garantire una sicurezza e un comfort sufficienti per i passeggeri diventa di primaria importanza, soprattutto in scenari caratterizzati da incertezza dovuta all’interazione tra agenti, che emerge come un’impegnativa sfida quando la consapevolezza della situazione dell’AV è scarsa. A tal proposito, le rotatorie rappresentano un ambiente complesso da gestire per un AV, rendendo critica la selezione dell’azione corretta. Questa tesi propone un approccio innovativo che sfrutta le tecniche di deep learning per sviluppare un controller intelligente per un AV, con un focus sull’analisi del comportamento degli agenti circostanti, in particolare la loro aggressività, per aumentare la consapevolezza della situazione. Sfruttando un classificatore a rete neurale convoluzionale (CNN), delle immagini sono recuperate dalle time series e usate per catturare caratteristiche rilevanti nel comportamento degli agenti, fornendo informazioni riguardo la loro aggressività. Così l’AV prende decisioni informate "GO/STOP" all’ingresso della rotatoria, dando priorità alla sicurezza dei passeggeri e riducendo le collisioni. Lo studio sottolinea l’importanza di set di dati naturalistici come il round-D, fondamentale per comprendere i comportamenti dei veicoli reali nelle rotatorie. Successivamente, viene approfondito il processo di etichettatura degli agenti come aggressivi o non aggressivi attraverso il clustering K-means, un algoritmo non supervisionato utile per dividere i dati nelle due classi. La valutazione dei risultati del clustering, supportata dall’analisi della correlazione e dai punteggi di silhouette, conferma la robustezza di questo approccio nel distinguere tra comportamenti di guida aggressivi e non aggressivi. Gramian Angular Field (GAF) e Markov Transition Field (MTF) vengono sfruttati per convertire time-series in immagini, migliorando la precisione e la ricchezza dell’input della CNN, la cui nuova architettura mostra un’accuratezza notevole, superando il 90% nel riconoscimento del comportamento aggressivo. Un ambiente di test e validazione viene implementato sfruttando il linguaggio di programmazione Python. Lo scenario comprende due agenti, uno comandato tramite un joystick PS4, che permette di emulare il comportamento umano, e il secondo (l’AV) guidato da un controller PID che agisce su accelerazione e sterzo, in cui è implementato l’intero processo con l’obiettivo di classificare il comportamento del primo agente. L’integrazione di questi controllers nel simulatore CARLA permette di valutare le capacità in tempo reale dell’intero framework.
Driverless strategy accounting for drivers' aggressiveness
Paludo, Andrea
2022/2023
Abstract
The rapid evolution of autonomous vehicles (AVs) is reshaping mobility, promising to reduce traffic accidents and revolutionize road safety. In this context, ensuring sufficient safety and comfort for passengers becomes paramount, particularly in scenarios characterized by uncertainty due to the interaction between agents, which emerges as the most challenging task to face when AV’s situational awareness is too low. In this context, roundabouts represent a complex environment to deal with for an AV, making critical the selection of the proper action. This thesis proposes an innovative approach leveraging deep learning techniques to develop an intelligent controller for an AV, with a focus on assessing the behavior of surrounding agents, notably their aggressiveness, increase situation awareness and safety. Exploiting a convolutional neural network (CNN) classifier, images are retrieved from raw time series and used to capture relevant features in agents’ behavior, providing information about their aggressiveness. Thus the AV makes informed "GO/STOP" decisions at roundabout’s entrance, prioritizing passengers’ safety and reduce collisions. The study emphasizes the importance of naturalistic data-set like the round-D, instrumental in understanding real-world vehicle behaviors in roundabouts. Subsequently, the work delves into the process of labeling agents as either aggressive or non-aggressive through the K-means clustering, an unsupervised algorithm useful to split data into the two classes. The evaluation of clustering results, aided by correlation analysis and silhouette scores, validates the robustness of this approach in distinguishing between aggressive and non-aggressive driving behaviors. Gramian Angular Fields (GAF) and Markov Transition Fields (MTF) are exploited to convert time-series into images, enhancing the granularity and richness of CNN’s input, whose novel architecture exhibits commendable accuracy, exceeding 90% in recognizing aggressive behavior. A testing and validation environment has been developed through Python programming language. The scenario comprehends two agents, one commanded via a PS4 joystick, allowing to emulate human behavior, and the second (the AV) guided by a PID controller acting on throttle and steering. In the latter code, all the pipeline whose aim is to classify the behavior of the former agent is implemented and the integration of these controllers in CARLA simulator permits to assess the real time capabilities of the whole framework.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Paludo_Tesi_01.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
8.23 MB
Formato
Adobe PDF
|
8.23 MB | Adobe PDF | Visualizza/Apri |
2024_04_Paludo_Executive Summary_02.pdf
non accessibile
Descrizione: Executive summary
Dimensione
3.58 MB
Formato
Adobe PDF
|
3.58 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/10589/217818