A Tree-Structure Classifier Ensemble for Tracked Target Categorization

In this paper, an integrated solution of moving object categorization is proposed, within the context of the visual surveillance system. Tracked targets are classified into four categories: single pedestrian, car, bicycle, and others (also the negative samples). In our framework, a set of strong classifier ensembles is in the form of multi-level particular features, where each feature represents an observable or derivable description associated with a tracked object. Here the classifiers based on heterogeneous features can be divided into two categories: 1) low-level numeric features (with low dimension), being classifier via AdaBoost algorithm, such as blob aspect ratio, blob area, and optical flow based velocity; 2) middle-level vector-descriptor features (with high dimension), being classifier via SVM approach, such as shape context, active template, and histogram of gradient. These classifier ensembles are then organized into a decision tree structure which is based on a proposed feature ranking technique. Compared with the similar AdaBoost and JointBoost framework quantitatively, our approach shows better results using a public LHI benchmark.