Capacity of deep learning algorithms on modelling pedestrian behaviour at crossings with countdown signal timer installations

Abstract

This paper attempts to investigate the capacity of different techniques in the modelling of pedestrian crossing behaviour in an urban signalised intersection with a Countdown Signal Timer (CST). Initially, two (2) models were used for the estimation of the average pedestrian crossing speed. The first model was a Linear Regression (LR) model, while the second one was a Deep Learning (DL) model. The R2 value of 0.56 for the DL model indicated a significant improvement in performance, in contradiction to the R2 of 0.22 for the LR model. Afterwards two (2) more models were exploited for the classification of pedestrians in terms of their crossing behaviour. At first, a Multinomial Logistic Regression (MLR) model was used as baseline, achieving an overall accuracy of 98.2%. A DL was then fit, which reached an accuracy of 99.6% and managed a better classification for the minority classes compared to the base model.