Incoherent and online dictionary learning algorithm for motion prediction

Abstract

Accurate model development and efficient representations of multivariate trajectories are crucial to understanding the behavioral patterns of pedestrian motion. Most of the existing algorithms use offline learning approaches to learn such motion behaviors. However, these approaches cannot take advantage of the streams of data that are available after training has concluded, and typically are not generalizable to data that they have not seen before. To solve this problem, this paper proposes two algorithms for learning incoherent dictionaries in an offline and online manner by extending the offline augmented semi-non-negative sparse coding (ASNSC) algorithm. We do this by adding a penalty into the objective function to promote dictionary incoherence. A trajectory-modeling application is studied, where we consider the learned atoms of the dictionary as local motion primitives. We use real-world datasets to show that the dictionaries trained by the proposed algorithms have enhanced representation ability and converge quickly as compared to ASNSC. Moreover, the trained dictionaries are well conditioned. In terms of pedestrian trajectory prediction, the proposed methods are shown to be on par (and often better) with the state-of-the-art algorithms in pedestrian trajectory prediction.