Abstract
Currently, most of urban intersections are being installed surveillance cameras, and several vision-based techniques are emerging to exploit and explore the traffic behavior. Vehicle trajectory classification plays an important role in the traffic flow analysis, such as organizing the structure of urban intersections, detecting traffic events, monitoring the abnormal driving activities … and so on. In this thesis, some methods of vehicle trajectory classification and vehicle location prediction are proposed. For solving the trajectory classification problem within an urban intersection, the training trajectories are roughly categorized using their sources and sinks initially, and for each group of trajectories a statistical model is trained using the Hidden Markov Model (HMM), then iteratively refines all of the models until no trajectory misplaced or accuracy cannot be improved further. Based on these well-trained models, the tracklet of the vehicle can be projected accordingly. Given an identified prefix trajectory, the most likely model is determined and the most probable template (tracklet) with the highest similarity is selected. This template gives the direction to forecast the next few locations. Finally, the real-time tracking of all vehicle trajectories at urban intersections can be possible with the help of this vehicle location prediction solution.