Destination intention estimation-based convolutional encoder-decoder for pedestrian trajectory multimodality forecast

Ruiping Wang, Siew-Kei Lam, Meiqing Wu, Zhijian Hu, Changshuo Wang, Jing Wang

Abstract: Highlights•Dear Editors, Thank you for handling our revised manuscript. Following the reviewers’ comments, we have reorganized and summarized the innovations of the paper. The main highlights of our paper are listed as follows.•Different from methods that use one-dimensional vectors to represent trajectory information, we design a new trajectory data representation method that transforms pedestrian trajectory data into 2D spatial trajectory heatmap, which not only preserves the spatio-temporal characteristics of pedestrian movements, but also helps the model to perceive the tendency of pedestrians’ movement.•Compared with current approaches that mostly use learning attention weights to construct social interactions, we propose to model group interaction of pedestrians by utilizing physical features, relative speed, and relative distance between pedestrians, which are more intuitive and interpretable for characterizing the properties of the interaction behaviors between pedestrians.•Different from the trajectory multimodal modelling approach that only focuses on a single destination, we propose a novel Conv2D-based encoder-decoder network to address the trajectory multimodality induced by different destinations, which can explicitly model the destination intentions first and then the specific trajectories by efficiently utilizing scenario semantics and historical trajectory heatmaps.

External IDs:doi:10.1016/j.measurement.2024.115470