Deep circadian-informed probability refinement network for pedestrian intent classification in urban complex

Abstract

Urban complexes often feature a mix of commercial, entertainment and recreational space serving a wide range of services. Pedestrian intent classification is hence crucial to identify their different destinations and understanding their needs. Moreover, circadian effects generally influence pedestrian behaviour. This paper proposes a deep circadian-informed probability refinement network for pedestrian intent classification (CIPRNet). It incorporates circadian information using a multiplexer network architecture to refine preliminary classification probabilities generated by a preliminary deep learning-based trajectory classifier. A joint loss function is used to co-optimize both the preliminary baseline trajectory classifier and the CIPRNet. Experimental results using real pedestrian trajectories captured from 3D range sensors at the Osaka Asia and Pacific Trade Centre (ATC) on a sunny day and cloudy day show that the CIPRNet can improve the state-of-the-art prediction of pedestrian paths by long short term memory classifier and trajectory unified transformer by approximately 13% and 10%, respectively. The CIPRNet is also extended to trajectory prediction and it outperformed various state-of-the-art algorithms in terms of average and final displacement error reduction. It may serve as an attractive alternative for pedestrian intent classification for urban complexes.