A Multiple Velocity Fields Approach to the Detection of Pedestrians Interactions Using HMM and Data Association Filters

Abstract

AbstractThis paper addresses the diagnosis of interactions between pairs of pedestrians in outdoor scenes, using a generative model for the trajectories. It is assumed that pedestrians’ motions are driven by a set of velocity fields, learned from the video signal. This model is extended to account for the interaction among pedestrians, using attractive/repulsive velocity components. An inference algorithm is provided to estimate the attraction/repulsion velocity from the pedestrian trajectory and characterize pedestrians’ interaction. Since we consider multiple motion models switched according to space-varying probabilities, inference is performed by combining a data association filter with a HMM-like forward algorithm. The proposed algorithm is denoted I-PDAF and is tested with synthetic data and pedestrians trajectories.KeywordsHide Markov ModelActivity RecognitionData AssociationAssociation ProbabilityOutdoor SceneThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.