Hidden Markov Models for Traffic Observation

An automated method for traffic monitoring based on statistical models of sensor behavior in combination with a Markov model of monitoring points is described. A model of normal traffic flow is automatically constructed. The model´s structure and parameters are optimized using a mini-batch model-merging and parameter-updating algorithm. Incoming velocity vectors are conveyed to the model and the most probable path through the tunnel´s Markov model is computed. An alarm is generated when the sensor values have a low probability under the model. The performance, strengths and weaknesses of the automated traffic flow analysis system are discussed.