Multi-classification of Driver Intentions in Yielding Scenarios

Abstract

Abstract-Predictions of the future motion of other vehicles in the vicinity of an autonomous vehicle is required for safe operation on trafficked roads. An important step in order to use proper behavioral models for trajectory prediction is correctly classifying the intentions of drivers. This paper focuses on recognizing the intention of drivers without priority in yielding scenarios at intersections -where the behavior of the driver depends on interaction with other drivers with priority. In these scenarios the behavior can be divided into multiple classes for which we have compared three common classification algorithms: k-nearest neighbors, random forests and support vector machines. Evaluation on a data set of tracked vehicles recorded at an unsignalized intersection show that multiple intentions can be learned and that the support vector machine algorithm exhibits superior classification performance.

Cooperative Intersection Support System Based on Mirroring Mechanisms Enacted by Bio-Inspired Layered Control Architecture

“…Similar classification approaches may be found in other works; notably [23], which extends the application domain to the classification among four possible intentions (no action, stop, creep and go) in yield scenarios without priority. One important finding was that prediction of the vehicle with priority was necessary and, consequently, observations of the second vehicle were included in the training of the three tested classifiers (SVM, Random Forest and k-nearest neighbors).…”

“…As noted, Fig.7 shows that a warning is issued when a driver is approaching the stop line fast, but is cancelled when the driver corrects the trajectory. The notion of 'false alarm' hence somewhat fades in the interaction between human and artificial agents (the human may think that the agent is too conservative, but anyway capable of 'understanding' his actions); b) human behavior is adaptive: intentions may change at any time (e.g., Fig.7 showed that in a yield situation drivers switch from stopping to crossing) and this variability of intentions does not fit well into the conceptual framework of static classification (at least rolling classification would be necessary, such as in [23]).…”

“…Training was not trivial, using very large number of training samples. Scaling to other scenarios, such as multi-classification in yield scenario is also non-trivial, in practice requiring to re-engineer the system [23]. Conversely, the agent of this paper, because of the implicit driver model, is more general.…”

Biologically Guided Driver Modeling: the Stop Behavior of Human Car Drivers

“…Classification at intersections is also studied in [20], extending the case to yield scenarios and classification among four possible intentions (no action, stop, creep and go). Three machine-learning approaches are studied and observations of the second vehicle are included to account for the mutual influence between vehicles in the yield scenario.…”

Sampling Based Motion Planning for Heavy Duty Autonomous Vehicles

“…The future trajectories of other vehicles are often predicted using the assumption of constant velocity (Ziegler et al, 2014b;Werling et al, 2008;Ferguson et al, 2008). This assumption is often valid in simpler situations but with increasing complexity of traffic situations, intentions such as lane changes, yielding at intersection or turning needs to be predicted (Liebner et al, 2013;Schlechtriemen et al, 2015;Ward and Folkesson, 2015). Further complicating the prediction is the dependency on our own actions.…”