Abstract: The Haar-like features used in MIL(multiple instance learning) trackers are not efficient to represent the appearances of the targets， and the noise samples are prone to be involved for classifier training phase， then drift in targets may happen. To solve these problems， an online discriminative feature selection (ODFS) tracking algorithm based on distribution fields (DFs) descriptors at instance level was proposed. Firstly， soft histogram method is manipulated to fastly approximate DFs， and the Haar-like features are replaced with the layers of DFs， which are adopted to represent appearance information. Then， supervised learning with prior information of instance labels is conducted; the ODFS algorithm is used to select the most optimal discrimination layer features， which can handle drift more effectively. The proposed tracking method are tested in benchmark dataset of a large variety of scenarios and under new evaluation indexes. Experimental results show the effectiveness of the algorithm.

Key words: visual tracking, distribution fields descriptors, feature selection, soft histogram, supervised learning