Prediction in Dynamic Environments via Identification of Critical Time Points

Abstract

PRIDE (PRediction In Dynamic Environments) is a hierarchical multi-resolutional framework that incorporates multiple prediction algorithms to enable navigation of autonomous vehicles in real-life, on-road traffic situations. At the lower levels, we utilize estimation theoretic short-term (ST) predictions via an extended Kalman filter-based algorithm using sensor data to predict the future location of moving objects with an associated confidence measure. At the higher levels, we utilize a long-term (LT) situation-based probabilistic prediction using spatio-temporal relations and situation recognition. At these levels, moving objects are identified as far as the sensors can detect and these two approaches provide for the prediction of the future location of moving objects at various levels of resolution at the frequency and level of abstraction necessary for planners at different levels within the hierarchy. This paper presents the results from research exploring the integration of these two prediction approaches in a way that the predictions from one can help to enforce the predictions of the other. We propose a methodology at which the short-term prediction algorithm no longer provides results within an acceptable predefined error threshold. We identify the different time periods where the two algorithms provide better estimates and thus demonstrate the ability to use the results of the short-term prediction algorithm to strengthen/weaken the estimates of the long-term prediction algorithm at different time periods. We provide experimental results in an autonomous on-road driving scenario on a closed-track for different cases using AutoSim, a highfidelity simulation tool that models details about road networks, including individual lanes, lane markings, intersections, legal intersection traversability, etc