Fahrerabsichtserkennung und Gefährlichkeitsabschätzung für vorausschauende Fahrerassistenzsysteme
FacultiesFakultät für Ingenieurwissenschaften und Informatik
LicenseCC BY-NC-ND 3.0 Deutschland
The intention of drivers to start discrete manoeuvres (like a lane change or a turn manoeuvre) is identified as a major factor determining the dynamic scene within the next couple of seconds. Hidden Markov Models are successfully applied for an early detection of beginning driving manoeuvres. The models use signals of the driving dynamics of the vehicle, digital maps, and a camera-based lane detection as features. By limiting the Markov chains to linear models, it is possible to train complete manoeuvres, and afterwards extract meaningful submodels. The submodels are chosen in a way to fit on manoeuvre starts, in order to ensure early detection. Manoeuvre recognition can be purely based on the probability of the observation sequence which allows for a computationally efficient classification. Decoding steps are not necessary. The linear structure of the Hidden Markov models also enables the determination of "typical" paths of an observation sequence through the Markov chain. The result is not only a yet more computationally efficient classification possibility, but also a measure that is more robust in the presence of differently scaled variances of observations and difficulties introduced by the geometrically distributed state transitions inherent to Markov Models. The classification performance is evaluated using a manually labelled reference database of field-recorded driving data. Based on the probabilistic information about upcoming manoeuvres, a trajectory prediction can be performed. In order to assess the risk of an evolving traffic situation, collision probabilities are evaluated. As determining collision probabilities is a computationally intense operation, efficient approximation methods are proposed and evaluated.
Subject HeadingsDigitale Karte [GND]
Driver assistance systems [LCSH]
Markov processes [LCSH]