A deep neural network approach for pedestrian trajectory prediction considering flow heterogeneity. (1st January 2023)

• Record Type: Journal Article
• Authors: Nasr Esfahani, Hossein; Song, Ziqi; Christensen, Keith
• Is Part Of: Transportmetrica. Volume 19:Number 1(2023)
• Publication Date: 2023-01-01
• Subjects: Pedestrian trajectory prediction -- individuals with disabilities -- neural network -- deep neural network -- long-short-term memory; Transportation -- Periodicals; Transportation -- Research -- Periodicals; 388.072
• DOI: 10.1080/23249935.2022.2036262
• Languages: English
• ISSNs: 2324-9935

Abstract

Abstract : Pedestrian trajectory prediction is imperative in specific fields, such as crowd management and collision prevention in automated driving environments. In this study, a novel long-short-term memory (LSTM)-based deep neural network capable of simulating the different walking behaviours of individuals with and without disabilities was designed. This network consists of three modules: the Disability module, the Environmental module, and the Trajectory Prediction module. Data from a large-scale pedestrian walking behaviour experiment involving individuals with disabilities were used to train and test the network. These data correspond to several experiments. Each experiment attempts to capture the essence of individuals' walking behaviour in different situations. By sequencing and normalising the input data and applying regularisation techniques, the network was successfully trained. The results were compared to state-of-the-art models, demonstrating that the network can predict pedestrians' trajectories more accurately, especially when pedestrian heterogeneity is involved.