L1 tracker with spatially weighted similarity measure based clustering

Abstract

Recently, sparse representation has been successfully applied in visual tracking for its efficiency to varieties of corruptions. It is, however, unqualified for practical applications due to the extremely high computational expense of ℓ1 minimization. This paper proposes a new L1 tracker that resolves the above problem by clustering particles via k-means based on a spatially weighted similarity measure(SWSM) under particle filter framework. The SWSM which incorporates spatial relationships between particles into pixel-wise similarity measure is calculated for each particle pair, and then is fed for k-means clustering. After that, a two-stage selection based on ℓ2 and ℓ1 minimization respectively is applied to jointly determine the target state. Our L1 tracker keeps the diversity of particles from drifting and also largely promotes the tracking efficiency. The good performance of the proposed method is validated by comparison with two other state-of-the-art L1 tracker on four challenging sequences.