Dynamic Neural Network Algorithm For Traffic Incidents Detection

With the development of the society, Intelligent Transportation System (ITS) becomes more and more important to our everyday life and the whole economy. Traffic Automatic Incident Detection System (TAIDS), which is one of the key parts of ITS, is able to detect and deal with the traffic incidents on the roads so as to decrease the death rate and the loss of wealth. Moreover, TAIDS can help to avoid subsequent incidents, save energy sources, reduce pollution and so on. As the core of TAIDS, therefore, Traffic Incident Detection Algorithm (AID) is worthy of study.According to the characters of traffic flow and the fundamental requirements of traffic incident detection, the techniques and theories relevant to traffic incident detection are systematically investigated in this thesis. After that, a new traffic incident detection model based on the combination of traditional fault detection of states estimation and the dynamic neural network is proposed, and corresponding BP algorithm is developed. The main contribution can be stated as follows:1) The emergence and development of ITS are introduced and analyzed in details. The structures of ITS and TAIDS are systematically discussed. Some difficult problems in the applications and theories of ITS and TAIDS are pointed out.2) After the analyzing the theory of neural network, an incident detection approach based on the dynamic neural network is presented. It uses dynamic neural network to estimate the traffic flow status, and compares the estimated slate values and the instrumented ones to judge the events.3) Aiming at traffic flow dynamic model, a particular structure of dynamic neural network is presented and its rationality is pointed out. Some algorithm based on BP net are designed for this network, whose training algorithm is improved to get faster convergence and more accurate direction.4) Data produced by traffic flow model are used to validate the structure of network presented in this thesis. The simulation results on Matlab platform show that the proposed model can approximate real state and detect traffic incident effectively.