Bayesian Filtering for Background Change Detection in TV Dramas

With the advancement of image recognition techniques and computing power, video analysis has recently attracted much attention. However, previous approaches have limitations for practical analysis of various unstructured video streams because they focus on very restricted features such as repeating frames. In addition, the previous studies do not ensure a proper model with the faulty assumption on the distribution and data organization. In order to alleviate these difficulties, we introduce a particle filtering-based estimation method requiring minimum prior knowledge and assumptions. The proposed method constructs a latent variable model based on a set of particles and their associated weights. This latent variable model is used to compute the likelihood of a newly observed data given the model. This method is especially effective for domains with irregular changes due to particle filtering's flexibility. We apply the proposed method to background segmenting of TV drama episodes. Each background segment is represented by a set of particles. If the likelihood of a newly observed data given the current segment model is low, then a new model is estimated. We validate the performance by comparing its estimation results with those of human estimation.