Neural Network Approach to Identifying Anomalies in the Traffic Flow

The paper considers approaches to identifying deviant behavior in traffic flows using the neural networks. It presents a dataset developed in virtual environment using the revised S-NFS model simulating human-driven, autonomous, and deviant human-driven vehicles. Observations include vehicle speed and positions structured as matrices for the neural network embedding. Two goals are set, i.e. identifying the vehicle type and evaluating the classification time. A recurrent neural network with the LSTM layers ensures accuracy of over 99% and far surpasses the 33% random guessing rate. The paper shows that simple traffic models are able to simulate complex behavior, and the longer observation sequences could reach near 100% accuracy. Results indicate that neural networks are able to classify vehicles efficiently and identify deviant behavior in the traffic systems.