Online multiple support instance tracking

We propose an online tracking algorithm in which the support instances are selected adaptively within the multiple instance learning framework. The support instances are selected from training 1-norm support vector machines in a feature space, thereby learning large margin classifiers for visual tracking. An algorithm is presented to update the support instances by taking image data obtained previously and recently into account. In addition, a forgetting factor is introduced to weigh the contribution of support instances obtained at different time stamps. Experimental results demonstrate that our tracking algorithm is robust in handling occlusion, abrupt motion and illumination.