Trajectory Clustering and Behaviour Retrieval from Traffic Surveillance Videos

The system presented here uses a minimum of prior knowledge to retrieve trajectories and abnormal behaviour. This system is based on the premise that most members of society choose to comply with traffic laws. The system input is only two parameters, namely (1) the maximum number of junctions available and (2) the rate of infractions in the city. In an unsupervised manner, the new system is capable of learning all legal behaviour and, on its own, extracting the knowledge required for detection of illegal behaviour. The query model used in this system is based on keywords and user sketches. Keywords are labels applied automatically by the system to express the activity models. The proposed system also optimizes user sketches before implementation. Practical implementations have demonstrated the high efficiency of this system in learning legal behaviour and detecting illegal practices.