Object tracking using random sparse appearance model

Selection of reliable templates is important for robust tracking. To this end a novel appearance model is proposed. This model consists of several representive templates. Each template, whose feature vector is extracted from a group of randomly selected feature points on the object, represents a different view of the object. By means of a sparse representation method, the appearance model is updated according to the sparse coefficients of the best candidate by solving l₁-minimisation. Experiments with both public and the authors own challenging datasets show that the new method outperforms several state-of-the-art methods in accuracy.