A hybrid generative/discriminative model based object tracking primary exploration

Based on analysis and discussion of object representation, a hybrid model based tracking by detection algorithm is presented as yet a primary exploration. The whole system is made of a learning-detecting two phase loop. Object model is built on a general Haar-like feature space which is automatically generated and extracted by a special random projection. Our proposed algorithm involves two type of methods for object modeling, one is to learn a transformation matrix by Principal Component Analysis (PCA) as the multi-view appearance model of the target object, and the other is to learn a classifier by Fisher Linear Discriminant Analysis (FLD) as the classification between the foreground and the background. We extend the Fisher criterion to a multi-mode background situation, which is used to formulate features´ discriminating power as feature weighting from the online captured positive/negative training data. In additionally, a two-stage detection is involved, in which all input samples firstly are tested by the learned FLD classifier to pick up candidates, then amongst candidates the maximum likelihood to the target template as the final detection result is searched for by PCA code matching. All generative model, discriminative model and target templates should online update due to appearance variation. A number of experiments illustrate that the proposed hybrid model based tracking algorithm does has advantages.