Mehran Zamani Abnili, Nasser L. Azad

Abstract: In this paper a framework for short-term microscopic prediction of traffic participants’ motion is presented and is deployed in a roundabout simulation using SUMO for evaluation. This framework consists of a dynamic Bayesian network where expert knowledge is incorporated and a continuous variable prediction module (CVPM) where continuous variable prediction is handled by a sequential neural network models. The DBN topology was designed to To have a comparison, three CVPM models were experimented with: recurrent neural network (RNN), gated recurrent unit (GRU), and long short-term memory network (LSTM). The results show promising 0.036 RMSE and higher than 0.895 correlation between 10-second predictions and actual data for the worst case.

Keywords: Machine learning, situational awareness, traffic participants behaviour prediction, dynamic Bayesian network, recurrent neural network, gated recurrent unit, long short-term memory

Date Published: May 28, 2021
DOI: TBA

View Article