DIGITAL.CSIC. Repositorio Institucional del CSIC
Publication
Tesis doctoral (DoctoralThesis)
MULTIMODAL INTERACTION-AWARE MOTION PREDICTION IN URBAN ENVIRONMENTS
Autonomous vehicles represent a transformative technological advancement with pro-
found implications for our future mobility and quality of life. These systems promise to sig-
nificantly enhance safety by eliminating human errors on the road, improve traffic flow, and
reduce environmental impact through optimized driving patterns. Moreover, automated
driving allows passengers to use travel time more productively, resulting in less stressful
and more enjoyable commutes. The development of autonomous driving has the potential
to reshape the way we move and profoundly impact our daily lives.
However, ensuring safety while navigating through highly interactive and complex
driving scenarios remains a critical challenge. Motion prediction is a fundamental chal-
lenge in the realm of autonomous vehicles, encapsulating the complex task of foreseeing
how other vehicles in the environment will behave. In the pursuit of safe and efficient
self-driving capabilities, predicting the future movements of the surrounding vehicles is
paramount. This predictive foresight enables autonomous systems to make informed deci-
sions, plan routes, and navigate through dynamic and often unpredictable real-world sce-
narios.
In this thesis, a method to infer the intentions, based on a Dynamic Bayesian Network,
and predict the motion of the nearby vehicles, using Markov Chains, at three main envi-
ronments (highways, roundabouts and intersections) is proposed. This approach considers
all possible corridors of the surrounding traffic participants and takes into account their in-
teractions to infer intentions probabilities, as well as the probability of being in each of the
possible navigable corridors. To achieve this, the Dynamic Bayesian Network is used to
model the relationships between the observed states and the unobserved intentions of the
nearby agents. The Markov Chain model, obtained from a kinematic model, is used to pre-
dict the future motions of the vehicles, taking into account their current state, their inferred
intentions, and the uncertainty associated with the prediction. The resulting multimodal
motion predictions are sent to the ego vehicle to navigate through the scene.
To tackle limitations, the framework is extended to adapt to different driving contexts via
a hybrid approach that combines model and learning-based methods. To handle occlusions
caused by sensor range or obstacles, an extension considers partial perception sensor data in
order to provide the ego vehicle with more reliable predictions in situations where the field-
of-view is obstructed. In order to predict the future motion of the surrounding vehicles in
a timely manner, a collection of Graphics Processing Unit acceleration methods is designed
and implemented to bring down the computational cost of the multimodal probabilistic
interaction-aware motion prediction algorithm, This thesis has been partially supported by the Spanish Ministry of Science and Innova-
tion through the National Project SECREDAS (PCI2018-093144) and by the European Com-
mission and ECSEL Joint Undertaking through the Project SECREDAS (783119). Also, it
has been supported by the Community of Madrid through SEGVAUTO 4.0-CM Programme
(S2018-EMT-4362), Peer reviewed