Prediction of driver intended path at intersections

Abstract

The complexity of situations occurring at intersections is demanding on the cognitive abilities of drivers. Advanced Driver Assistance Systems (ADAS) are intended to assist particularly in those situations. However, for adequate system reaction strategies it is essential to develop situation assessment. Especially the driver's intention has to be estimated. So, the criticality can be inferred and efficient intervention strategies can take action. In this paper, we present a prediction framework based on Hidden Markov Models (HMMs) and analyze its performance using a large database of real driving data. Our focus is on the variation of the model parameters and the choice of the dataset for learning. The direction of travel while approaching a 4-way intersection is to be estimated. A solid prediction is accomplished with high prediction rates above 90% and mean prediction times up to 7 seconds before entering the intersection area.

Multimodal Trajectory Predictions for Autonomous Driving using Deep Convolutional Networks

“…Manually engineered models fail to scale to many different traffic scenarios, which motivated the use of machine learning models as alternatives, such as Hidden Markov Model [22], Bayesian networks [23], or Gaussian Processes [24]. In recent Kitani et al [26] used inverse optimal control to predict pedestrian paths by considering scene semantics, however the existing IRL methods are inefficient for real-time applications.…”

Interaction-Aware Probabilistic Behavior Prediction in Urban Environments

“…For highways, this set usually consists of lane change left, lane change right, and keep lane (e.g., [7], [8]). For intersections, the desired route is mostly represented by the turning directions left, right, and straight (e.g., [5], [6]). Besides the intention of a lane change or the desired route, more detailed intentions can be distinguished.…”

“…Interactions between traffic participants are often not considered (e.g., [7], [5], [2], [3]). When the motion of multiple vehicles is interdependent, however, this may result in inaccurate predictions, especially for longer prediction horizons (e.g., if a vehicle approaching an intersection has to decelerate because of a slow vehicle in front, without considering interactions, it might be misleadingly inferred that it intends to turn).…”

Safe Reinforcement Learning on Autonomous Vehicles

“…These methods are limited in that the Gaussian probability models and kinematic transitions assume cars roughly follow a known trajectory. More sophisticated models allow for multiple maneuvers [22] which can be done by including road information (either heuristically or learned for particular intersections [23]) to allow for multiple possible maneuvers. More recent work in vehicle prediction is starting to consider the interactions between multiple vehicles [24], [25].…”