This paper discusses a modular neural network arterial incident detection model that was developed and evaluated using simulated data. A microscopic traffic simulation model of a commuting corridor in Brisbane was used to generate a total of 36 incidents at different times of the day, with varying severity and duration. The neural network model uses speed, flow and occupancy data, provided in 20-second cycles from both the upstream and downstream stations, in addition to section travel times. The model was trained on 23 incidents and its performance evaluated on the remaining 13 incidents. The initial results reported in this paper demonstrate the feasibility of using modular neural networks for arterial incident detection and provide directions for developing fast and reliable automated arterial incident detection models.
Abstract