Getting robust observation for single object tracking: A statistical kernel-based approach

Mean shift-based algorithms perform well when the tracked object is in the vicinity of the current location. This cause any fast moving object especially when there is no overlapping region between the frames fails to be tracked. The aim of our algorithm is to offer robust kernel-based observation as an input to a single object tracking. We integrate kernel-based method with feature detectors and apply statical decision making. The foundation of the algorithm is patch matching where Epanechnikov kernel-based histogram is used to find the best patch. The patch is built based on Shi and Tomasi [1] corner detector where a vector descriptor is built at each detected corner. The patches are built at every matched points and the similarity between two histograms are modelled by Gaussian distribution. Two set of histograms are built based on RGB and HSV colour space where Neyman-Pearson method decides the best colour model. Diamond search configuration is applied to smooth out the patch position by applying maximum likelihood method. The works by Comaniciu et al. [2] is used as performance comparison. The results show that our algorithm performs better as we have no failure yet lesser average accuracy in tracking fast moving object.