Robust Detection of Traffic Incidents Based on Neural Network

Traffic accidents seriously affect transportation and public safety. To improve the robustness of the model, this paper proposes an LSTM-FNN model combining Long Short-Term Memory Network and Feedforward Neural Network (FNN) using a comprehensive data set of US traffic accidents from 2016 to 2023 and traffic flow data set from PeMS, considering the spatio-temporal heterogeneity. LSTM is based on data sets possessing temporal properties, and FNN uses data sets representing spatial properties. LSTM is very suitable for dealing with highly correlated time series data. FNN has strong ability to capture patterns inherent in the data and contribute to greater efficiency. The experimental results show that the model has satisfactory accuracy and good robustness. The fusion model is also resistant to data defects. The significance of this paper lies in its innovative approach and practical implications, which provide an important scientific basis for accident detection and safety measures in similar transportation systems.