Video Behavior Analysis Using Topic Models With Structured Information

Video behavior analysis is a meaningful and challenging work in computer vision. The ex-isting approaches to analyzing video behavior can be divided into two categories:one is based on trajectory, and the other is based on motion patches. These first approaches depend on the accuracy of tracking, and the second ones are based on the extracting of motion features. The approaches based on motion features are focused on how to mining the abstract behavior from the motion features.Topic models have the ability to mine the the latent abstract topic of the data, which were mainly used in text classification and retrieval. Topic models have two typical models:Latent Dirichlet Allocation(LDA) and Hierarchical Dirichlet Processes(HDP). The first one has to prefix the number of topics, and the second one is a nonparameterized bayesian model with no need to set the number of topics. It can be viewed as conducting abstraction and dimensionality reduction by applying topics models, which can convert the original data represented by high dimension of words to the data represented by lower dimension of topics. As topic models have such character-istic, many contributions have been applied in video behavior analysis to find out the behaviors of motion features.As the topic models are based on the hypothesis of bag of words, they have the defect of ignoring the structured information. The video data are original spatial-temporal data, without which the analysis will not be accuracy. Most researchers have improved the models by introduc-ing spatial-temporal information. They mainly focused on increasing the levels of the models or modeling the dynamic change. Different from the others, we propose a new method to introduce the structured information. We add the location and direction of the motion patches to the models and reconstruct the features. We improve both LDA and HDP models by introducing structured information. In the LDA model we view the direction as the stamp of the location, we prefix the number of topic and analyze the visualization result to determine the appropriate number of the topics. In the HDP model, we can view the direction as the stamp of the location, and we can also view the direction as the independent variable in which there is no need to prefix the number of the topics. The experimental results on QUML datasets show that our improved LDA model can find the atomic behaviors in the video data, and the improved HDP model can find the appropriate number of the topics. According to the combination of the atomic behaviors we can distinguish the video scene and detect abnormal behaviors in the scene.