Pedestrian Intention Anticipation with Uncertainty Based Decision for Autonomous Driving

Abstract

Being able to anticipate actions is a critical part of many applications nowadays. One of them is autonomous driving, undoubtedly one of the most popular subjects today, where action anticipation can be used to help define how the vehicle should act next. In this work, we present a method for action anticipation in the autonomous driving scenario, specifically to anticipate pedestrian intentions. The method extracts movement features from a video sequence, to which we can add context information from other sensors. These features are used by a deep learning sequential model, which predicts the action being done by a pedestrian. Furthermore, we propose a skeleton completer, which can be used for many other applications. We also explore the concept of decisions under uncertainty, since this is a high risk scenario, and propose an effective method to decide whether or not to anticipate the action. Our methods obtain state of the art results in terms of the anticipation accuracy in two comprehensive datasets.