Bayesian Networks in Lane Change Maneuver Prediction

master's thesis

• Title (croatian): Primjena Bayesovih mreža za predviđanje manevra promjene prometne trake
• Author: Ivan Grabić
• Mentor: Petar Ćurković (mentor)
• Granter: University of Zagreb; Faculty of Mechanical Engineering and Naval Architecture; Zagreb
• Defense date and country: 2020-12-04, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing; Artificial Intelligence

Abstract

Developing autonomous vehicles is a challenging task. One of the reasons for this is that human behavior is unpredictable. Another reason is that this problem is in a high-risk environment. Most traffic accidents are due to human error. Thus a conclusion can be made that autonomous vehicles will make driving safer. One type of accident happens when one participant changes the lane and the other participant doesn't notice it. If a system could predict lane change and alert the driver in time it could prevent an accident. Deep learning methods are state of the art approach to prediction problems. Neural networks are, however, known as black-box models. This is the reason they are not fully suitable for high-risk domains such as traffic environment. This thesis will take an alternative approach to lane-change maneuver prediction. This approach is called Bayesian or probabilistic machine learning. There are two main benefits to this approach. First is interpretability of probabilistic models and second is good uncertainty representation. We will create a Bayesian network for predicting lane change maneuvers of traffic participants. We will look at Highway Drone Dataset (HighD) and show conclusions. From this dataset, we will create a training and test dataset. To create a model we will use a probabilistic programming language called pyro. We will train the model on a training set using an algorithm called Black Box Variational Inference. After the training, the model is evaluated and evaluation metrics are reported. The inference time is appropriate for real-time implementation. Prediction power is comparable with other probabilistic approaches but worse than deep learning models.

Abstract (croatian)

Razvoj autonomnih vozila izazovan je zadatak. Jedan od razloga tome je što je ljudsko ponašanje nepredvidljivo. Drugi je razlog taj što je ovaj problem u visoko rizičnom okruženju. Većina prometnih nesreća dogodila se zbog ljudske pogreške. Stoga se može zaključiti da će autonomna vozila vožnju učiniti sigurnijom. Jedna vrsta nesreće dogodi se kada jedan sudionik promijeni traku, a drugi sudionik to ne primijeti. Ako bi sustav mogao predvidjeti promjenu trake i na vrijeme upozoriti vozača, mogao bi spriječiti nesreću. Metode dubokog učenja su najsuvremeniji pristup problemima predviđanja. Neuronske su mreže, međutim, poznate kao crne kutije. To je razlog što nisu u potpunosti prikladne za rizične domene poput prometnog okruženja. Ovaj rad će zauzeti alternativni pristup predviđanju manevara promjene trake. Taj se pristup naziva Bayesovim ili probabilističkim strojnim učenjem. Dvije su glavne prednosti ovog pristupa. Prva je interpretabilnost probabilističkih modela, a druga dobro prikazivanje nesigurnosti. Stvorit ćemo Bayesovu mrežu za predviđanje manevara promjene trake sudionika u prometu. Razmotrit ćemo "Highway Drone Dataset" (HighD) i pokazati zaključke. Od ovog skupa podataka stvorit ćemo set podataka za učenje i testiranje. Za stvaranje modela poslužit ćemo se vjerojatnosnim programskim jezikom zvanim pyro. Model ćemo trenirati na setu podataka za učenje koristeći algoritam nazvan "Black Box Variational Inference". Nakon učenja, model se evaluira te su prikazane metričke vrijednosti. Vrijeme predviđanja prikladno je za provedbu u stvarnom vremenu. Moć predviđanja usporediva je s drugim probabilističkim pristupima, ali je gora od modela dubokog učenja.

• Keywords: Probabilistic machine learning; probabilistic programming; lane-change maneuver prediction; Bayes networks; autonomous driving.
• Keywords (croatian): Probabilističko strojno učenje; probabilističko programiranje; predviđanje manevra promjene trake; Bayesove mreže; autonomna vožnja.
• Language: english
• URN:NBN: urn:nbn:hr:235:829533
• Study programme: Title: Mechanical Engineering; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka strojarstva (magistar/magistra inženjer/inženjerka strojarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access; Embargo expiration date: 2022-12-04
• Created on: 2020-11-30 17:18:54