Trajectory Classification For Crowded Scenes Via Topic Models

With the increase of terrorist attacks, the security issues in public places have drawn more attentions worldwide. Effective behavior monitoring, analysis, and prediction for moving objects in public places are of great significance. However, the existing work are mainly related to afterwards event retrieval. With the population of big data of surveillance video, the demand of real time behavior analysis of moving objects in crowded scenes is very urgent.The primary works in this thesis are as follows:1. By taking a full consideration of the challenge of fragmentary and chaotic trajectories, Global-MRF is proposed to preprocess trajectories.2. By taking a full consideration of the accuracy of the trajectory representation and expansibility of framework, BCTM is used to learn middle level features of the trajectories.3. Based on the idea of discriminant extension, the multi-SVM is used to design reasonable classified procedure, which can improve the precision of the classification.4. According to the idea of generative extension, Multi-CTM is proposed to learn middle level features of the trajectories and trajectory classification. In order to reduce the hand-craft calibration, a method of Multi-CTM based weakly supervised trajectory classification is proposed.