—We present an approach for vehicle classificationin IR video sequences by integrating detection, tracking andrecognition. The method has two steps. First, the movingtarget is automatically detected using a detection algorithm.Next, we perform simultaneous tracking and recognitionusing an appearance-model based particle filter. We presenta probabilistic algorithm for tracking and recognition thatincorporates robust template matching and incrementalsubspace update. There are two template matching methodsused in the tracker: one is robust to small perturbationand the other to background clutter. Each method yieldsa probability of matching. The templates are representedusing mixed probabilities and updated when the appearancemodels cannot adequately represent the variations in objectappearance. We also model the tracking history using anonlinear subspace described by probabilistic kernel principalcomponents analysis, which provides a third probability.The most-recent tracking result is incrementally integratedinto the nonlinear subspace by augmenting the kernel Grammatrix with one row and one column. The product of thethree probabilities is defined as the observation likelihoodused in a particle filter to derive the tracking and recognitionresult. The tracking result is evaluated at each frame. Lowconfidence in tracking performance initiates a new cycle ofdetection, tracking and classification. We demonstrate therobustness of the proposed method using outdoor IR videosequences.
展开▼
