Gastvortrag von Mauricio Munoz am 09.07.2015

Am Donnerstag, den 09.07.2015, hält Mauricio Munoz, Absolvent des Elitestudiengangs Software Engineering, ab 17:15 Uhr in der Ringvorlesung des Elitestudiengangs einen Vortrag über seine Masterarbeit am MIT. Der Vortrag findet in Raum 1057N statt und hat den Titel 'Driver Behavior Modeling Using Hidden Markov Models'.

Driving is a complex task that encompasses a wide array of challenges. In the general driving scenario, the driver may be intuitively modeled as both a sensor and actuator, constantly both receiving information from the environment (e.g. traffic, vehicle telemetry) and implementing locally optimized strategies for managing and completing subtasks. The mechanism that propels and guides this feedback loop may be often reduced to the processing of cognitive, physical, auditory, visual workload sources that result, for instance, from seemingly latent variables such as the spatial layout of the vehicle's human-machine interface or the modality and structure with which a driver is presented with a new task. Efforts in interface optimization that seek to minimize stress sources are therefore necessarily coupled with understanding how the driver processes and manages this workload. This report aims to analyze the link between these modules by using extensive driving data to model driver behavior patterns and extrapolate the fundamental differences between distinct strategies from both a quantitative as well as qualitative perspective. In particular, this work builds on the observation that the driving task may be decomposed into a series of observations over time, making it a prime candidate for data analysis tools that leverage statistical variable progressions as a key signal for classification. As such, Hidden Markov Models are employed and benchmarked in a series of related yet self-contained predictive analyses. These include inferring task modality based on driver glance data, as well as exploring the correspondence between head rotation and glance allocation. This report significantly extends on previous works by approaching these questions from a predictive machine learning perspective. In addition, these quantitative approaches are supplemented by a qualitative analysis technique designed to highlight the reasons for these differences in visual behavior. Results show that these approaches are not only adequate for the research topics at hand, but raise a vast number of additional questions as well as new potential directions for future work.

Zum Vortrag sind alle herzlich eingeladen.

message of 30.06.2015